alirezarezvani/claude-skills / finance/skills/stock-analysis
stock-analysis
UnverifiedAnalyze, research or value a listed stock or company (US, global, NSE or BSE) with verified data, ratios, scores and valuation. Use whenever asked if a stock is a buy, what it is worth, or how a company looks.
references/17-process-and-epistemics.md
View on GitHub# Research Process, Epistemics and the Price Layer
Use this when: you are deciding *how* to run the analysis rather than what number to compute — at the start (scoping, competence, budget), at the point of judgement (Stage 8-10, forming a view), and before you write the verdict.
Everything else in this skill tells you what to measure. This file tells you how to think while measuring, and how to know when you are fooling yourself. It matters because the dominant failure mode in equity research is not arithmetic error — it is a well-executed analysis of the wrong question, or a correct analysis whose conclusion was fixed before the work began. The governing rule of this skill (a metric is meaningless until you know its sector and the company's own history; never rank on a single number; for banks, insurers, REITs and miners the standard ratios are undefined or inverted) is itself an epistemic rule, not a formatting rule — it exists because context-free numbers produce confident, invertible conclusions. Part B adds the price layer, which is a genuine input to risk control and timing and a rationalisation engine if you let it lead.
## Contents
**Part A — Process, epistemics and decision hygiene**
- [1. Circle of competence and the "too hard" pile](#1-circle-of-competence-and-the-too-hard-pile)
- [2. Research budget and diminishing returns](#2-research-budget-and-diminishing-returns)
- [3. Falsification-first research design](#3-falsification-first-research-design)
- [4. Steelmanning the bear case; reading short reports](#4-steelmanning-the-bear-case-reading-short-reports)
- [5. Base rates and the outside view](#5-base-rates-and-the-outside-view)
- [6. Checklists that work; scores that must not decide](#6-checklists-that-work-scores-that-must-not-decide)
- [7. False precision: ranges, reverse DCF and decimals](#7-false-precision-ranges-reverse-dcf-and-decimals)
- [8. Conviction is not certainty: confidence tiers and sizing](#8-conviction-is-not-certainty-confidence-tiers-and-sizing)
- [9. The research file and the evidence trail](#9-the-research-file-and-the-evidence-trail)
- [10. The decision journal and calibration](#10-the-decision-journal-and-calibration)
- [11. Re-underwriting on a schedule](#11-re-underwriting-on-a-schedule)
- [12. Update discipline: signal vs noise](#12-update-discipline-signal-vs-noise)
- [13. Source quality, incentives and management assertions](#13-source-quality-incentives-and-management-assertions)
- [14. Knowing when to say no](#14-knowing-when-to-say-no)
- [15. Pre-mortems and post-mortems](#15-pre-mortems-and-post-mortems)
- [16. Anchoring, framing and order effects](#16-anchoring-framing-and-order-effects)
- [17. External challenge and echo chambers](#17-external-challenge-and-echo-chambers)
- [18. Auditing the process itself](#18-auditing-the-process-itself)
- [19. Epistemic rules specific to an AI analyst](#19-epistemic-rules-specific-to-an-ai-analyst)
**Part B — Price action, technical and timing inputs**
- [20. What the price layer is for](#20-what-the-price-layer-is-for)
- [21. Chart data hygiene](#21-chart-data-hygiene)
- [22. Trend and moving-average structure](#22-trend-and-moving-average-structure)
- [23. Relative strength vs index and vs sector](#23-relative-strength-vs-index-and-vs-sector)
- [24. 52-week positioning and momentum](#24-52-week-positioning-and-momentum)
- [25. Volume, delivery and the institutional footprint](#25-volume-delivery-and-the-institutional-footprint)
- [26. Volatility, beta and capture](#26-volatility-beta-and-capture)
- [27. Drawdown history and regime behaviour](#27-drawdown-history-and-regime-behaviour)
- [28. Liquidity, float and tradability](#28-liquidity-float-and-tradability)
- [29. Short interest, crowding and positioning mechanics](#29-short-interest-crowding-and-positioning-mechanics)
- [30. Event behaviour and non-fundamental flow](#30-event-behaviour-and-non-fundamental-flow)
- [31. Staged accumulation and invalidation levels](#31-staged-accumulation-and-invalidation-levels)
- [32. Price-fundamental divergence: the tape as evidence](#32-price-fundamental-divergence-the-tape-as-evidence)
- [33. Where technicals add value and where they mislead](#33-where-technicals-add-value-and-where-they-mislead)
- [Checklist](#checklist)
---
# Part A — Process, epistemics and decision hygiene
## 1. Circle of competence and the "too hard" pile
Before opening a spreadsheet, write one page in plain language: how does this company make money, who pays and why, what would have to be true for it to earn materially more in five years, and which three variables drive the outcome. No jargon, no sell-side deck, no "platform". If you cannot write that page from primary sources, the correct output is not a weak analysis — it is a documented pass.
Keep a literal, dated **too-hard list** with the reason recorded, so the same name is not re-litigated every quarter. Recurring members: pre-revenue biotech; opaque cross-border holding structures (VIEs, layered offshore SPVs); banks and insurers where you cannot read the loan book, restructured-asset disclosure or reserve triangle; commodity trading houses; serial acquirers whose growth is unauditable roll-up accounting; crypto-adjacent balance sheets; companies where the majority of profit sits in unconsolidated or related-party entities.
**India-specific too-hard triggers:** promoter groups with a history of related-party fund diversion; companies with high or rising promoter pledge and opaque end-use; frequent auditor changes with no CARO explanation; a listed holdco whose value is entirely unlisted subsidiaries with no published accounts; names under SEBI's GSM/ASM surveillance framework where price formation itself is impaired.
Watch for **competence drift** — the slide from a real competence area (consumer staples) to an adjacent-sounding one (specialty pharma) whose value drivers are entirely different. The tell is that you start reasoning by analogy rather than from unit economics.
*Why:* most permanent loss comes from owning what you could not assess, not from missing winners. Passing has zero cost; being wrong with capital committed does not.
## 2. Research budget and diminishing returns
Set the depth budget before you start and tie it to the stakes and reversibility of the decision, not to how interesting the company is. Track *which source actually changed your estimate* of the three key variables. When the last block of work produced no revision to those three, stop — that is the diminishing-returns signal, and the honest thing to do is write up.
Prefer **breadth of source type** (filings, competitor filings, customer disclosures, regulator dockets, industry data, trade press) over depth in one type (the tenth broker note). Guard both errors: the over-researched name that must be bought because forty hours were spent on it, and the "obvious" idea presented with the same confidence as a researched one.
*Why:* the payoff curve on research is strongly concave — the annual report, the competitor's annual report and the cash flow statement capture most of the edge. Later hours add confidence without accuracy, and sunk effort biases the conclusion toward action.
## 3. Falsification-first research design
Write the thesis as a **falsifiable statement with numbers and dates** before gathering supporting evidence: "segment revenue compounds above 12% through FY29 while gross margin holds above 45%". Then list the specific observations that would prove it false, and assign each disconfirmer a data source and a check frequency.
Then spend the *first* research block hunting disconfirmation: the bear case, the losing competitor's commentary, the regulator's docket, customer complaint channels, the changed risk factors. No verdict is written until at least one serious attempt to kill the thesis has been documented and survived.
*Why:* confirmation bias is the default state of research — once you like an idea, every subsequent fact reads as support. Reordering the process is one of the very few debiasing techniques that works, because it changes *what you look at* rather than asking you to feel less biased. It also converts narrative into a testable claim, which is a precondition for ever knowing you were wrong.
## 4. Steelmanning the bear case; reading short reports
Write the bear case yourself, in its strongest form, *before* reading anyone else's — then compare, and treat the gap as a measure of your blind spots. Then actively source the other side: short-seller reports, the most negative covering analyst, bearish threads that contain actual numbers, borrow cost and short-interest trends, and the year-over-year *diff* of the company's own risk factors (a newly added risk factor is a disclosure event).
When reading a short report, separate three layers and treat them differently:
| Layer | How to treat it |
|---|---|
| **(a) Verifiable facts** — filings, court records, customs/import data, permits, registry entries | Check each one yourself against the primary source. These are the only part that can change a thesis. |
| **(b) Interpretation** of those facts | Argue with it. Reasonable people read the same filing differently. |
| **(c) Rhetoric, price target, framing** | Discount entirely. It is marketing for a position. |
Note the author's incentive and horizon, but never dismiss a report on incentive alone — everyone publishing has a position. Ask the decisive question: *if every checkable fact in this report is true, does my thesis survive?* Then watch the company's response — a specific, itemised, numbers-based rebuttal is informative; a defamation notice plus ad hominem is also informative.
**India note:** dedicated short reports are rare and the borrow market (SLB) is thin, so adversarial work is under-supplied. Substitute: rating agency rationales and downgrade notes, CARO qualifications, auditor resignation letters filed with the exchanges, SEBI orders and adjudication notices, NCLT filings, and the MCA/ROC accounts of unlisted group entities.
*Why:* shorts are the only participants paid to do adversarial forensic work on a company, and they have repeatedly surfaced accounting and related-party issues years ahead of the market. The failure is symmetric: dismissing them costs you fraud blow-ups, following them blindly costs you the many short reports that are simply wrong about well-run companies.
## 5. Base rates and the outside view
Before accepting any company-specific forecast, find the reference class and state its base rate. What share of companies sustain 20%+ revenue growth for a decade? What share of large acquisitions create value? How often do turnarounds actually turn? What is the historical distribution of margin expansion for firms already at this margin level? How many entrants in this category survived a full cycle?
Then compare the company's own five-to-ten-year record and management's prior guidance-versus-delivery against the current promise, and state **explicitly why this company should beat the base rate** — a specific, durable mechanism. "Great management" and "large TAM" are not mechanisms.
Always sanity-check the terminal implication: what share of the addressable market must the company hold in year 10 for the model to work, and has any company ever held that share in this industry structure? A model that implicitly requires 40% share of a fragmented, low-switching-cost market has already told you it is wrong.
*Why:* the inside view — a bottom-up story built from company detail — is systematically overoptimistic because it ignores how rarely the story class succeeds. High growth mean-reverts, high margins attract entry, most M&A destroys value. Base rates are the cheapest available correction and they are usually devastating to the aggressive case.
## 6. Checklists that work; scores that must not decide
Tier the checklist. Run a short **kill-switch list** on every name — roughly ten to fifteen disqualifiers: auditor resignation or qualified opinion, going-concern language, unexplained related-party flows, chronic negative operating cash flow with rising debt, promoter pledge above a threshold with falling price, restatement, undisclosed encumbrances, a covenant cliff inside twelve months. Only names that survive get the long list.
On scoring (see `references/11-scoring-rubric.md`), enforce four rules:
1. **Normalise every factor within sector or against the company's own history.** Raw cross-sector numbers in a composite reproduce exactly the single-metric error this skill exists to prevent.
2. **Veto factors override any score.** Averaging converts a fatal binary flaw — fraud, a covenant cliff, an unfixable governance problem — into a small point deduction.
3. **Check for compensating errors.** A very strong score on one factor masking a fatal weakness on another is the composite's characteristic failure.
4. **Test weight sensitivity.** If a plausible reweighting reorders your top names, the score contains no information and must not be reported as if it does.
Audit checklist use periodically: which items have never once changed a conclusion (delete them), and which post-mortems traced to an item that was skipped (promote them).
*Why:* a 200-item list run without attention becomes a tick-box ritual that manufactures false confidence, and weights chosen without evidence are priors wearing a spreadsheet costume. Scoring exists to structure judgement, not to replace it.
## 7. False precision: ranges, reverse DCF and decimals
Force every valuation output into a **range with stated assumptions**, and check that the low case is genuinely bad — recession *plus* share loss *plus* margin reversion — not merely "slightly less good". Count decimals: if a DCF prints a fair value to the cent while terminal value is 75% of the total, delete the decimals and demote the model to a scenario-comparison tool.
Make **reverse DCF the primary tool** (`references/06-valuation.md`): what growth, margin and reinvestment rate does today's price imply, and does that sit inside or outside the company's own historical distribution and the sector's? Judging whether an implied assumption is plausible is a far easier task than forecasting the future outright.
Stress only the two or three assumptions that drive most of the variance — sensitising forty inputs is theatre. Reconcile the model's implied unit economics to something physical: stores, seats, tonnes, MW, beds, subscribers, capex per unit of capacity. And never inherit precision from a consensus number or third-party model you have not reproduced.
*Why:* a model's precision is capped by its least reliable input. A five-year revenue estimate accurate to ±30% cannot produce a fair value accurate to the rupee — but the false precision is exactly what generates the confidence to overweight and to ignore contradicting evidence.
## 8. Conviction is not certainty: confidence tiers and sizing
Conviction is a feeling produced by familiarity and effort. Certainty is a property of the evidence. They diverge most dangerously after long research. Keep two things separate in the write-up:
- **How likely the thesis is to be right** — and on what evidence.
- **The payoff spread** — how bad the downside is if wrong, how good the upside if right.
A high-probability / low-upside case and a low-probability / high-upside case are not the same recommendation even if both are "positive". State both.
Cap stated confidence by the **quality of the information**, not the strength of the narrative: what share of the thesis rests on verified primary data, versus management assertion, versus your own extrapolation? Any fraud or governance question, or any single-point-of-failure risk (one customer, one product, one regulator, one plant, one country), caps confidence regardless of how good the numbers look.
Where the report discusses sizing or risk control, keep it **generic and principle-based** — volatility- and drawdown-aware sizing, hard caps for governance-questionable names, checking whether five holdings are really one factor bet. Do not write personalised allocation instructions; you are producing analysis, not advice, and the report should say so.
*Why:* sizing is where epistemics become financial. An excellent process with reckless sizing still ends in ruin; modest sizing lets you be wrong often enough to keep compounding. If you cannot articulate why confidence is high rather than moderate, the analysis is not finished.
## 9. The research file and the evidence trail
Maintain one file per name containing: five to ten years of annual reports and proxies (India: annual report + notice of AGM + shareholding pattern; US: 10-K, 10-Q, DEF 14A via EDGAR), transcripts, the **page or paragraph citation behind every key claim**, the peer set with the reason for each inclusion and exclusion (`references/10-peer-set.md`), the model, and links to primary data — regulatory filings, trial and patent registries, customs/import-export data, court and NCLT dockets, environmental permits, job postings.
Tag every input by epistemic status, and keep the tags visible in your working notes:
| Tag | Meaning | Weight it carries |
|---|---|---|
| **F** | Audited or regulator-filed fact | Highest; can be relied on with a citation |
| **A** | Management assertion (concall, press release, investor deck) | Claim about the future or about unaudited detail — never restate as fact |
| **E** | Third-party estimate (broker, industry report, data vendor) | Usable as context; must be attributed and dated |
| **I** | Your own inference | Must be labelled; it is where most errors enter |
Retain superseded versions rather than overwriting, so thesis drift is visible. Never let a screener or data vendor figure drive a conclusion without reconciling it to the filing — vendors routinely mis-map segments, mishandle leases (IFRS 16 / Ind-AS 116), treat preference shares and perpetual instruments inconsistently, mis-state net debt, use unadjusted share counts, and silently restate history. Adjusted earnings, net debt, share count and segment data are the four that must always be traced to source.
*Why:* without a trail you cannot re-underwrite honestly, because you will not remember which numbers were verified and which were absorbed from a summary. The file is also what makes a post-mortem possible — you can only learn from an error if you can reconstruct what you actually believed and why.
## 10. The decision journal and calibration
At the moment of every conclusion — buy, add, trim, exit, or **pass** — record, before the outcome is known: the thesis in one paragraph; the three key variables and the forecast for each; the price and valuation at the time; what would trigger a reversal; a numeric probability; and what you expect to happen and by when. Include the state of the world (market regime, what else was happening), because that is what lets you spot mood-driven decisions later.
Review on a fixed cadence and score **calibration**: of the calls made at 80%, how many happened? Then tag each closed decision on the two-by-two — good process/good outcome, good process/bad outcome, bad process/good outcome, bad process/bad outcome — and take the lesson only from the *process* column.
Journal the passes too. A portfolio-only record cannot show whether the too-hard pile is protecting you or quietly costing you.
*Why:* memory reconstructs the past to fit the outcome. Hindsight bias makes every loss look foreseeable and every win look deliberate, which destroys the feedback loop that learning depends on. A contemporaneous, pre-outcome, written record is the only defence.
## 11. Re-underwriting on a schedule
At least annually, and after any material event, re-underwrite from scratch. The test is: **would you initiate this position today, at today's price, with today's facts, at this size?** Write the fresh thesis *before* rereading the old one, then compare — the gap is where thesis drift lives.
Classify explicitly:
- **Thesis intact, price fell** → potentially add.
- **Thesis broke** → exit; the loss is already taken, the only question is whether more capital should stay.
- **Thesis was quietly replaced with a new one to justify continuing to hold** → exit. This is the growth story becoming a value story becoming a dividend story becoming a "cheap on book" story. It is a permanent loss wearing the costume of long-term conviction.
Check the original key variables against actual delivery, and ask whether the original reason to own has already played out. Counter the endowment effect with a periodic clean-slate exercise: list holdings anonymously, with only their metrics and theses, and rank them against new candidates.
*Why:* portfolios decay silently. The world moves, the reason for owning expires, and the position persists on inertia and familiarity. Re-underwriting converts every holding into an active decision, which is the only fair standard.
## 12. Update discipline: signal vs noise
Pre-specify, **at the time of the conclusion**, what counts as thesis-relevant. Write the list: "gross margin below 42% for two consecutive quarters matters; a single quarter's revenue miss on FX does not; loss of the top customer matters; a broker downgrade does not."
When news arrives, ask one question: does this change one of the three key variables, or only the sentiment? A material adverse fact on a key variable triggers a full re-underwrite within a defined window — not a reflex trade, and not a silent explaining-away.
Guard both errors. **Anchoring** — refusing to update after a genuine break, which is most acute when already losing money on the name. **Over-updating** — rewriting the thesis every earnings call, which produces turnover, costs and worse decisions. And check whether you are updating on the *price move* rather than on evidence: price is information about other participants' views, not about the business, though a persistent unexplained divergence deserves an explanation (Section 32).
*Why:* most information flow is noise, but the rare genuinely thesis-breaking fact must be acted on immediately — and it is precisely the one you will most want to rationalise, because you are already down on it. Pre-specifying removes the judgement call from the moment when judgement is most compromised.
## 13. Source quality, incentives and management assertions
Grade every source by proximity to the fact and by incentive:
| Tier | Sources | Caveat |
|---|---|---|
| 1 | Audited financials, regulator filings and orders (SEBI, RBI, IRDAI, SEC), court/NCLT records, customs data | Still read the notes and the auditor's opinion — Tier 1 is not "unread and trusted" |
| 2 | Concall transcripts, investor presentations, management commentary | **Assertion, not fact.** Tag as A (Section 9) |
| 3 | Rating agency rationales, exchange filings by peers, industry association data | Rating rationales are unusually high-value in India — they contain covenant, pledge and liquidity detail found nowhere else |
| 4 | Paid industry reports, sell-side research | Note who commissioned it; note the coverage incentive |
| 5 | Media, newsletters, social, forums, AI-generated summaries | Zero weight as evidence; useful only as a pointer to a primary source |
For every critical claim, trace it to the original document rather than to a summary of it. Ask of every source: who paid for this, what is their horizon, what do they gain if I act on it?
**Score guidance versus delivery.** Pull three to five years of transcripts and compare what management said would happen with what happened — capacity commissioning dates, margin targets, debt reduction promises, capex budgets, subsidiary turnaround timelines. Note whether the *language* changes when results deteriorate (rising abstraction, new adjusted metrics, more time on strategy and less on numbers). This is one of the highest-yield and least-performed checks in equity research, because it is a direct quantified measure of whether this management's forecasts are worth anything.
**India note:** many Indian companies give no formal numeric guidance, so score qualitative commitments instead — commissioning schedules, capex plans, deleveraging targets, stated dividend policy, promises about monetising a subsidiary. Also weigh: the CARO annexure, the auditor's Key Audit Matters, and whether the concall Q&A allows non-scripted questions from institutions.
**US/global note:** guidance is explicit and quantitative, so scoring is easier — but so is managing to it. Check whether delivery came from operations or from buybacks, one-time gains, acquisitions and definitional changes to "adjusted" measures.
Treat expert-network calls and channel checks as **small, biased samples** and record the sample size. Ask whether investor-relations access is shaping the view, and whether the same conclusion survives without it.
## 14. Knowing when to say no
Adopt an explicit default of **no**. The idea must clear a written bar — understandable, verifiable from primary sources, adequate margin of safety, better than the weakest existing holding, executable at a sensible size — rather than needing a reason to be rejected.
Watch for manufactured pressure to act: cash drag, benchmark envy, someone else's winner, a fresh model that "needs" to justify itself, a research budget already spent. Track the pass list and its subsequent performance, but judge each pass on whether it was correct *given what was knowable at the time*, not on the outcome.
*Why:* there are no called strikes in investing. Inaction is a free option, and the largest source of avoidable underperformance for most investors is doing too much — too many names, too much turnover, too many positions bought because the work was already done. Missing a winner is not a loss; losing capital is. A well-defended "no" pile is a portfolio-level asset.
For this skill specifically: **"insufficient basis for a verdict" is a legitimate, useful output.** Say it plainly, say exactly what is missing, and say what would change the answer.
## 15. Pre-mortems and post-mortems
**Pre-mortem, before concluding.** Assume it is three years later and the position has lost 60%. Write the most likely *story* of how that happened — a concrete causal chain, not a list of risk words. Then ask of each path: is it cheap to monitor? Is it hedgeable? Is it likely enough to change the size or kill the idea outright?
**Post-mortem, after any position closes — winner or loser.** What was right, what was wrong, and was the error in the *facts*, the *interpretation*, the *sizing*, or the *timing*? Was it a recurring error type?
Maintain a running tally of your recurring error types and convert each repeat offender into a checklist item. Common ones: overpaying for quality, catching falling knives in structurally declining industries, trusting promotional management, ignoring dilution and stock comp, mis-set peer groups, mistaking cyclical peak earnings for a trend, believing a turnaround before cash flow confirms it.
*Why:* prospective hindsight — imagining the failure as already having happened — measurably improves risk identification versus asking "what could go wrong", because it forces a causal story instead of a vague list. Post-mortems on *winners* matter as much as on losers: a profitable outcome from a broken process is the most dangerous lesson you can teach yourself, and it is the one that gets repeated at larger size.
## 16. Anchoring, framing and order effects
Control the **sequence** of the work. Form a view of the business and its economics before looking at the price, the chart, the analyst target, or any entry cost. When the price is known first — which is often unavoidable — note it explicitly as an anchor in the working file.
Rules that follow:
- Purchase price is irrelevant to any hold/exit decision. The only question is whether you would buy at today's price.
- A 60% decline says nothing about cheapness. Refresh the anchor deliberately: re-derive value from current fundamentals, not from the old high.
- Examine absolute cash flows and per-share amounts, not only percentages and multiples; look at nominal debt in crore or millions, not only at ratios. Framing in ratios hides scale; framing in absolutes hides trend. Use both.
- Where practical, review the numbers with the company name and your prior view stripped out, and check whether the same conclusion arrives.
*Why:* anchors operate below awareness and are not neutralised by knowing about them, so the only real defence is procedural — controlling what you see and in what order. Entry price is the most destructive anchor in practice: it produces both the refusal to sell losers and the reluctance to add to winners.
## 17. External challenge and echo chambers
Before any high-conviction conclusion, have the thesis attacked by someone whose job in that conversation is to kill it — and give them **the file, not the pitch**. Require the challenge to be specific and evidence-based, and write down the objections that survive.
Cultivate at least one genuinely bearish source on the name and engage with them rather than around them. Avoid public commitment before the analysis is complete: once a view is stated to an audience, consistency pressure makes updating feel like a status loss, and investors routinely defend a broken thesis long past the evidence because they defended it in public first.
Audit your information diet. If everyone you read owns the same names, treat that as a warning, not a confirmation.
*Why:* self-generated criticism is systematically weaker than adversarial criticism, because you cannot see the assumptions you did not know you were making.
## 18. Auditing the process itself
Track process-level statistics, not just returns:
| Process metric | How to compute | Indicative reading | Why it matters |
|---|---|---|---|
| Hit rate | Share of closed decisions that were profitable | 40-60% is normal and compatible with excellent results | On its own it means nothing; only meaningful alongside win/loss size |
| Win/loss ratio | Average gain on winners ÷ average loss on losers | >1.5x for a concentrated long-only process | A 40% hit rate with 3x asymmetry beats a 70% hit rate with 0.5x |
| Calibration | Predicted probability vs realised frequency, bucketed | 80% calls should happen ~80% of the time | The only direct measure of whether your confidence means anything |
| Contribution by idea source | Return attributed to screen / competitor filing / spin-off / referral | — | Tells you where to spend research time next year |
| Contribution by thesis type | Compounder / cyclical / turnaround / special situation / deep value | — | "I lose money on turnarounds" is far more actionable than an aggregate return |
| Holding period, actual vs intended | Median realised holding period vs stated horizon | Large gap = process not matching stated style | Style drift shows up here before it shows up in returns |
| Turnover cost | Commissions + spread + impact + tax, as % of average capital | India: add STT and short-term capital gains treatment | Activity has a measurable price; most investors never compute it |
*Ranges are indicative only — they vary by market, strategy, cycle and sample size, and a short sample tells you almost nothing. Compare a metric to your own history before comparing it to any external norm.*
Audit annually: which idea sources actually produced returns; which thesis types you are demonstrably bad at (candidates for the too-hard pile); whether losses cluster in a sector, a market-cap band, or a behavioural pattern. Then change exactly **one or two things** and record the change with a date, so a future audit can attribute the effect.
*Why:* returns over any short horizon are dominated by luck and market beta, so judging a process by its P&L is a slow, low-signal and often misleading feedback loop. Process metrics are higher-frequency and attributable. And without a dated record of process changes, you cannot tell improvement from a regime change.
## 19. Epistemic rules specific to an AI analyst
These are non-negotiable and specific to how you fail, not how a human fails:
1. **Never state a company-specific number you have not retrieved.** No estimated revenue, no "approximately" margins, no plausible-looking ratio reconstructed from memory. A fabricated figure that happens to be near-correct is worse than a gap, because it will be trusted and propagated.
2. **Date everything.** Every figure carries an as-of date and a period label (FY25 vs CY25 vs TTM to a specific quarter). Indian fiscal years end 31 March; US filers vary. Mixing periods silently is one of the most common quiet errors.
3. **Distinguish "not disclosed" from "zero" from "not retrieved".** These have entirely different implications. Non-disclosure of a segment, a related-party balance, or a contingent liability is itself evidence.
4. **Cite to the document, not to a summary.** If a number came from a screener, say so and mark it unverified until reconciled to the filing.
5. **Do not let fluency substitute for evidence.** A well-written paragraph and a well-evidenced one are indistinguishable in tone. Tag each claim F / A / E / I (Section 9) while working, and make sure the final verdict rests mostly on F.
6. **Do not resolve conflicts between sources by averaging.** Find out which one is right, or report the conflict as a finding — a discrepancy between the annual report and the vendor is often the most interesting thing on the page.
7. **State units and currency explicitly.** Crore vs lakh vs million vs billion; INR vs USD; and never mix reported and converted figures in one table without labelling the rate and date.
8. **Refuse gracefully.** Where the data is insufficient or the business is outside what can be assessed, say so and stop. Declining is an output, not a failure.
9. **Analysis, not advice.** Present findings, evidence and risks. Do not issue personalised buy/sell/allocation instructions; state clearly that the output is research, not investment advice.
---
# Part B — Price action, technical and timing inputs
## 20. What the price layer is for
This layer is **supplementary**. It does not tell you whether a business is good, what it is worth, or whether the accounts are honest — those are settled by Parts 01-16 of this skill. What it legitimately provides is four things: (i) a risk-control frame (volatility, drawdown history, liquidity) that determines whether a correct thesis is survivable; (ii) a timing/staging frame that removes single-point entry guesses; (iii) a *challenge* signal — persistent relative weakness against peers is evidence that someone knows something; and (iv) pre-committed invalidation levels that stop a losing thesis from being rewritten indefinitely.
The rule for the whole of Part B: **a technical signal never overrides a fundamental conclusion, and a fundamental conclusion never entitles you to ignore a persistent price divergence.** Price gets to force a re-examination. It does not get to make the decision.
Report these items in a clearly labelled section, separate from the fundamental verdict, so a reader can discard the whole layer without disturbing the analysis.
## 21. Chart data hygiene
Before reading anything from a price series, verify it. Almost every wrong technical conclusion is downstream of a corrupted series, and the analyst never discovers the error.
- Confirm the series is **split- and bonus-adjusted**, and know whether it is price-only or total-return — for a 6% yielder the two diverge enormously over a decade.
- Confirm you are on the **primary listing**, not a thin ADR/GDR or a secondary line, and that currency is consistent throughout.
- Use **log scale** for multi-year charts; linear only for short windows. A linear twenty-year chart makes early percentage moves invisible and late ones look parabolic.
- Check the vendor's handling of spin-offs, reverse splits, ticker changes, scheme-of-arrangement demergers and rights issues. Cross-check two historical points against the company's own filings or the exchange's own bhavcopy/historical data.
- **India:** NSE and BSE prices differ slightly; pick one and stay on it. Adjust for bonus issues, stock splits and rights entitlements; verify the adjustment around any demerger, where vendors frequently leave an artificial gap.
## 22. Trend and moving-average structure
Identify the dominant trend on both weekly and daily charts: is price above or below the 50-day and 200-day moving averages, and — more informative than the level — is the 200-day sloping up, flat, or down? Note the sequence of higher highs/higher lows versus lower highs/lower lows. Check whether short-, medium- and long-term timeframes agree; conflict between them is itself information (usually that a change of regime is in progress).
Treat crossover signals (golden/death cross) as descriptive context, never as triggers. They lag by construction and whipsaw in range-bound markets.
*Why:* trend is the single most robust thing a chart provides, and the slope of the long-term average is a rough proxy for whether the market's assessment of the business is improving or deteriorating. **The specific value to a fundamental analyst:** a statistically cheap stock in a confirmed downtrend is the classic value-trap setup — cheapness plus a falling 200-day usually means estimates have further to fall, and the multiple is low because the E is wrong.
## 23. Relative strength vs index and vs sector
Absolute price movement conflates the company with the market. Plot the **ratio line** — the stock divided by the relevant index — over one, three and five years, and note whether it is making new highs, new lows, or basing.
Then decompose it in two steps, which answers the question "what do I actually own?":
1. **Stock vs sector index** — is this a stock-specific story or a sector story?
2. **Sector vs broad index** — is the sector leading or lagging the market?
Also compare the stock against three to six *named direct competitors* on one normalised chart, using the peer set you built in `references/10-peer-set.md` — not a GICS bucket. A stock breaking down while its peers hold up is a company-specific problem and demands an explanation before you conclude mispricing. The whole group breaking is a sector or macro problem, and the fundamental work should shift accordingly.
**India:** use the appropriate Nifty sector or thematic index (Bank, IT, Auto, FMCG, Pharma, Metal, PSE, Realty) rather than Nifty 50 alone, and compare a mid- or small-cap against Nifty Midcap 150 / Smallcap 250, because the broad index is dominated by large caps and will misstate relative performance badly.
*Why:* a stock up 10% in a market up 25% is quietly losing; a stock flat in a market down 20% is quietly winning. Persistent relative weakness ahead of bad news is one of the more reliable warning patterns in equity markets, and it is exactly the signal a fundamental analyst is temperamentally inclined to dismiss.
## 24. 52-week positioning and momentum
| Metric | Definition / how to compute | Indicative reading | Why it matters |
|---|---|---|---|
| Distance from 52-week high | (52wk high − price) ÷ 52wk high; note the date the high was set | Within ~10% in a rising trend = leadership; >50% off = repair job | Proximity to highs is a well-documented momentum anchor; investors under-react to good news |
| Distance above 52-week low | (price − 52wk low) ÷ 52wk low | Context only | Distinguishes "basing after a fall" from "still falling" |
| Distance from all-time high | Same, vs the all-time high and its date | A high set 8 years ago changes the story entirely | A multi-year lower high is a business-quality question, not a chart question |
| 1 / 3 / 6 / 12-month total return | Total return including dividends, absolute and relative to index and sector | — | The raw momentum inputs |
| 12-1 momentum | Trailing 12-month return excluding the most recent month | Top/bottom decile of the universe is the meaningful read | The standard cross-sectional momentum measure; skipping the last month avoids short-term reversal |
| Momentum quality | Is momentum backed by rising earnings estimates and delivered results, or by multiple expansion alone? | Earnings-backed is durable; multiple-only is fragile | Tells you what regime will hurt you |
*All indicative readings vary by market, cycle and period, and mean nothing without comparison to the stock's own history and its sector cohort.*
Two traps. First, **"down 60% so it must be cheap"** is pure anchoring: percentage decline says nothing about value, and a stock down 60% can fall another 60% (a further 60% decline from there leaves 16% of the original price). Second, momentum is prone to violent reversals at market inflection points — momentum crashes cluster at bear-market bottoms when beaten-down names rip hardest. Know whether your thesis is implicitly a momentum bet.
## 25. Volume, delivery and the institutional footprint
Compare volume on up days versus down days over the last four to thirteen weeks. Look for advances on volume above the 50-day average and pullbacks on declining volume; flag the opposite pattern — rallies on thinning volume, declines on expanding volume. Examine volume specifically around earnings, guidance and investor-day dates.
Technical accumulation/distribution indicators (OBV, A/D line) are noisy proxies. **Ownership filings are the real evidence**, so cross-reference:
- **US:** 13F quarter-over-quarter changes, number of institutional holders, active vs passive split, Form 4 insider transactions — separating genuine open-market purchases from option exercises and pre-scheduled 10b5-1 sales — and buyback execution disclosed in the 10-Q.
- **India:** the quarterly **shareholding pattern** (promoter, FII/FPI, DII, mutual fund, public — and critically, changes in promoter holding and pledge), **bulk and block deal** disclosures on NSE/BSE, SEBI insider-trading (PIT) disclosures by designated persons, and buyback/open-offer filings. Promoter pledge invocation is a distinctive Indian source of sustained mechanical selling.
- **India-specific volume quality:** the **delivery percentage** (deliverable quantity ÷ traded quantity), published daily by the exchanges. High volume with low delivery is intraday churn and says little; a rising delivery percentage on rising price is a genuinely different signal. There is no direct US equivalent.
*Why:* volume is the closest thing a chart has to a conviction measure — it shows how much capital was willing to transact at those prices. A breakout on no volume is often a liquidity artifact that fails. Heavy-volume down days indicate a motivated seller, often an institution unwinding a full position, which can cap a stock for months regardless of fundamentals — and knowing that changes entry pacing entirely.
## 26. Volatility, beta and capture
| Metric | Definition / how to compute | Indicative range | Why it matters |
|---|---|---|---|
| Realised volatility | Annualised stdev of daily log returns over 30 / 90 / 250 days | Large-cap staples ~15-25%; broad market ~15-20%; small/mid caps and cyclicals 35-60%+ | The primary input to any sizing decision |
| ATR % | Average true range ÷ price | 1-3% daily for liquid large caps; much higher for small caps | Sets how wide any stop must be to avoid noise-triggered exits |
| Implied vol / IV percentile | Option-implied vol vs its own 1-year history | IV percentile >80 = options expensive | Tells you whether the market expects an event, and whether hedging is cheap |
| Beta | Regression of stock returns on index returns, 1 / 3 / 5-year windows | Stability across windows matters more than the level | An unstable beta means the single number is meaningless |
| Up-capture / down-capture | Average stock return in up-index months ÷ index; same for down months | Up-capture > down-capture is the desirable asymmetry | Many stocks capture 80% of downside and 60% of upside — structurally a bad deal that a single beta hides |
| Factor exposures | Sensitivity to market, size, value, quality, momentum; plus rates, oil, FX where relevant | — | Exposes hidden concentration: ten "different" long-duration rate-sensitive growth names are one position |
*Indicative ranges vary enormously by market, market-cap band, cycle and measurement window; Indian small caps are structurally more volatile than developed-market large caps. Always compare against the stock's own history and its sector cohort first.*
Sizing a 60%-volatility name like a 15%-volatility name is how a correct thesis still produces an unacceptable outcome.
## 27. Drawdown history and regime behaviour
Chart the full drawdown series: maximum peak-to-trough decline, the depth of a typical annual drawdown, and the time to recover each one. Look specifically at behaviour in 2008, 2011, 2013 (India: taper tantrum and the rupee), 2015-16, Q4 2018 (India: the NBFC/IL&FS credit event), March 2020, 2022, and any sector-specific shock. For each decline, ask **what fundamentally caused it and whether that vulnerability still exists** — that question is what converts a chart into analysis.
Then test regime sensitivity: rising versus falling rates, steepening versus flattening curves, inflation up versus down, risk-on versus risk-off, dollar strength versus weakness, expansion versus recession. For India, add crude oil direction, USD/INR, monsoon and rural demand, and the government capex cycle. Note whether the sensitivity comes from the *business model* or only from the *multiple* — the distinction determines whether it is a permanent or a temporary problem.
*Why:* the relevant question is not what the stock returns but what you must survive to collect it. A name with a history of 45% drawdowns will produce another one; if that would force a sale for psychological, mandate or leverage reasons, the expected return is not actually available. Correlations also converge in stress, so diversification measured in calm markets is largely illusory.
## 28. Liquidity, float and tradability
| Metric | Definition / how to compute | Indicative reading | Why it matters |
|---|---|---|---|
| Average daily traded value | Mean of (price × volume) over 20 and 60 days, in ₹ crore or $m | Judge against intended position size, not an absolute band | Liquidity is the constraint that turns a good idea into an unexecutable one |
| Bid-ask spread | Typical quoted spread as % of price | <0.1% liquid large cap; >1% is a warning | A wide spread is a permanent tax on every entry and exit |
| Free float | Shares outstanding less promoter/insider, government and strategic stakes | India: promoter holding is disclosed quarterly, so float is precisely knowable | Low float amplifies both squeezes and collapses and makes the chart less informative |
| Position as multiple of ADV | Intended position ÷ average daily value | Days-to-exit matters far more than days-to-build | If exiting takes 15 days at 20% of volume, the stop-loss does not exist |
| Passive / index ownership | Index membership and share held by index funds | — | Determines rebalance-driven flow and how the name behaves in an index event |
**India-specific tradability constraints** — these have no US equivalent and can dominate everything above:
- **Circuit filters / price bands** (2%, 5%, 10%, 20% depending on the scrip): a locked upper or lower circuit means no exit at any price that day. Serial lower circuits are a genuine liquidity failure, not a technical pattern.
- **Trade-to-trade (T2T) segment**: no intraday netting; every trade must be settled by delivery. Sharply reduces liquidity.
- **ASM / GSM surveillance** (Additional / Graded Surveillance Measure): additional margins, periodic call auctions and trading restrictions. A name under GSM has impaired price formation — treat its chart as uninformative and consider the name too hard.
- **F&O eligibility and ban periods**: when market-wide position limit utilisation exceeds the threshold, the stock enters an F&O ban where only position reduction is allowed, distorting cash-market price.
## 29. Short interest, crowding and positioning mechanics
**US/global:** track short interest as a percentage of float and of shares outstanding, its trend, days-to-cover (short interest ÷ ADV), borrow cost and availability, and hedge-fund crowding measures. Distinguish genuine directional shorts from convertible-arb, merger-arb and index-hedge shorting — the latter carry no view on the business. Cross-check against unusual option open interest and skew.
**India:** there is no equivalent retail-visible short-interest disclosure and the SLB market is thin, so use substitutes — stock futures open interest and its direction relative to price (rising OI with falling price suggests short build-up; falling OI with rising price suggests short covering), the cost of carry / futures basis, options open interest concentration at strikes, FII derivative statistics, and the F&O ban list. Promoter pledge levels are the closest Indian analogue to a forced-selling overhang.
*Why:* heavy short interest cuts both ways. It can be informed scepticism worth investigating — shorts do real forensic work and are often early — or it can set up violent squeezes that make price completely uninformative about fundamentals for weeks. Either way, crowding tells you the price on the screen may reflect positioning mechanics rather than any view of the business, which is exactly when a fundamental analyst should stop reading the tape as evidence.
## 30. Event behaviour and non-fundamental flow
Catalogue the stock's reaction to the last eight to twelve results: average absolute move, direction, whether gaps filled or held, and whether the stock continued to drift in the same direction for the following weeks (post-earnings-announcement drift). Note reactions to guidance changes, investor days and regulatory rulings, and how far ahead of results the stock typically moves. This is the empirical distribution for the single largest recurring risk event you will face while holding — and gaps that *hold* indicate genuine repricing, while gaps that consistently *fill* indicate an overreactive shareholder base.
Then identify **non-fundamental flow drivers**, because a meaningful share of short-horizon price movement has nothing to do with value:
- Index additions/deletions and rebalance dates (S&P/MSCI/FTSE; India: Nifty and Sensex reconstitution, MSCI free-float factor changes, which move Indian mid-caps hard).
- Lock-up expiries and offer supply. **India:** IPO anchor-investor lock-ins (staggered 30/90 days) and promoter lock-in, plus minimum-public-shareholding-driven promoter sell-downs and OFS.
- Follow-on offerings, shelf registrations, QIPs, convertible and FCCB issuance, preferential allotments to promoters.
- Buyback blackout windows; December tax-loss selling (US) and March fiscal-year-end effects (India); dividend and bonus record dates.
- Business seasonality — festive season and monsoon in India, holiday quarter in US retail — which must be compared year-on-year, never sequentially.
Recognising mechanical flow prevents two errors: reading forced buying as validation, and reading forced selling as deterioration. Occasionally it creates the opportunity, when mechanical selling temporarily overwhelms fundamentals.
## 31. Staged accumulation and invalidation levels
Before any entry is contemplated, write down: the full intended position size, the number of tranches, and **what specifically triggers each subsequent tranche** — a price level, a time interval, a valuation threshold, or a fundamental milestone such as a confirmed inflection in the metric the thesis depends on. Also write down the conditions under which you will **not** add, and cap the maximum size regardless of how attractive the name looks after a decline.
Then define **invalidation in fundamental terms first**: "gross margin fails to recover above X by FY27", "net debt/EBITDA exceeds Y", "the top customer is not renewed", "receivable days do not normalise within two quarters". Only after that, translate it into a price or drawdown level if you use one. Size so that hitting the invalidation level costs a pre-defined, acceptable share of capital, using volatility (ATR or realised vol) rather than a fixed percentage — an identical percentage stop is far tighter on a 15%-vol name than on a 55%-vol one.
Decide in advance whether stops are hard, mental, or **time-based** (a thesis that has not progressed by a stated date is a failed thesis, even if the price has not moved). Be explicit that stops in gapping or illiquid names may fill far away, and in Indian names locked at a lower circuit may not fill at all.
*Why:* staging converts entry timing from one high-stakes guess into a process, because nobody times bottoms. And a *pre-written* rule is the only thing that distinguishes disciplined averaging from averaging down into a broken thesis — the latter is the most common way a survivable loss becomes a portfolio-damaging one. Without a pre-committed invalidation level, sizing is arbitrary and losing positions get rationalised indefinitely, because the mind reliably rewrites the thesis to fit the price.
## 32. Price-fundamental divergence: the tape as evidence
When price disagrees sharply and persistently with the fundamental view — the stock falls steadily while your model and consensus estimates hold — do not resolve it by asserting that the market is wrong. Hunt for what the market might be discounting:
- credit signals: bond prices, CDS, rating agency outlook changes, commercial paper rollover, and in India the rating rationale and any covenant breach disclosure;
- supplier, customer and competitor results and commentary — the read-across usually arrives before the company's own numbers;
- insider and promoter behaviour, including pledge changes and creeping sell-downs;
- channel checks, hiring/attrition data, app and web traffic, customs and import-export data;
- short reports and litigation dockets;
- whether the divergence is stock-specific or shared by the whole peer group (Section 23).
Set an explicit rule for how much unexplained relative weakness triggers a formal thesis re-review, and record the outcome of that review either way.
*Why:* price aggregates many participants' information, some of it better than yours. Dismissing persistent divergence as "the market being wrong" is comfortable and occasionally correct, but it is also the standard prelude to a large permanent loss. The discipline is not to obey the tape — it is to let it force a genuine re-examination with a documented conclusion.
## 33. Where technicals add value and where they mislead
Be explicit about the boundary, and state it in the report so no reader over-weights this layer.
| Use | Verdict | Why |
|---|---|---|
| Trend direction and 200-day slope | **Useful** | Robust, simple, and the single best guard against value traps |
| Cross-sectional momentum (12-1) and relative strength | **Useful** | Among the most persistent documented empirical effects, though prone to crashes at inflections |
| Volatility and drawdown history for sizing | **Useful** | The honest input to how much risk a position actually carries |
| Liquidity, float, days-to-exit, circuit/surveillance status | **Useful** | Hard constraints; they determine whether the idea is executable at all |
| Pre-committed invalidation levels | **Useful** | Enforces a bounded loss and imposes a decision date |
| Volume confirmation and delivery percentage | **Moderately useful** | Directional evidence about conviction; noisy on any single day |
| Support/resistance and base structure | **Moderately useful, partly self-fulfilling** | Levels matter mainly because many holders sit near break-even there, and because many participants act on them |
| Named chart patterns, Fibonacci retracements, Elliott wave counts | **Not useful** | Pattern-matching on noise; the human eye finds structure in random walks |
| Multi-indicator "confluence", short-horizon oscillator signals | **Actively harmful** | Add enough indicators and one always confirms the view you already held |
Two enforcement rules. **Pre-commit to a small, fixed indicator set before looking at the chart** — adding indicators until one agrees with you is the characteristic failure of this entire discipline. And **score your own technical calls** the same way you score fundamental ones (Section 10); if they show no calibration, delete the layer rather than keep it as decoration.
---
## Checklist
**Process and epistemics**
- [ ] Wrote the plain-language one-pager (how it makes money, who pays, what must be true, three key variables) before modelling.
- [ ] Confirmed the business is inside the circle of competence; if not, added it to the dated too-hard list with a reason and stopped.
- [ ] Set a depth budget up front and stopped when new work stopped revising the three key variables.
- [ ] Stated the thesis as a falsifiable claim with numbers and dates, and listed the specific disconfirmers with sources.
- [ ] Ran the disconfirming search *first*, and documented one serious attempt to kill the thesis.
- [ ] Wrote the bear case before reading anyone else's; separated verifiable facts from interpretation from rhetoric in any short report.
- [ ] Stated the reference-class base rate and the specific mechanism by which this company beats it.
- [ ] Sanity-checked the terminal implication (required market share in year 10) against industry history.
- [ ] Ran the kill-switch checklist on the name before the long checklist.
- [ ] Normalised every scored factor within sector or against own history; applied vetoes; tested weight sensitivity.
- [ ] Output a valuation range, not a point; used reverse DCF; stressed only the two or three dominant assumptions.
- [ ] Reconciled implied unit economics to something physical.
- [ ] Separated probability from payoff, and capped stated confidence by evidence quality, not narrative strength.
- [ ] Tagged every input F / A / E / I; traced adjusted earnings, net debt, share count and segments to the filing.
- [ ] Scored management guidance-versus-delivery over three to five years (India: qualitative commitments and commissioning dates).
- [ ] Graded every source by proximity and incentive; traced critical claims to the original document.
- [ ] Recorded the decision — including a pass — with thesis, key variables, probability and reversal triggers, before the outcome.
- [ ] Pre-specified what future news counts as thesis-relevant and what is noise.
- [ ] Ran a pre-mortem: the concrete story of a 60% loss three years out.
- [ ] Formed the fundamental view before looking at price, target prices or entry cost; noted any unavoidable anchor.
- [ ] Sought at least one genuinely adversarial reading of the file, not the pitch.
- [ ] Confirmed that "insufficient basis for a verdict" was considered and rejected on evidence, not on effort already spent.
- [ ] Invented no company-specific number; dated and labelled every figure; stated units and currency.
**Price layer**
- [ ] Verified the price series: split/bonus-adjusted, primary listing, consistent currency, log scale for multi-year.
- [ ] Recorded trend state: price vs 50-day and 200-day, and the 200-day slope.
- [ ] Plotted relative strength vs the broad index, the correct sector index, and named peers.
- [ ] Recorded distance from 52-week and all-time highs, with the dates they were set.
- [ ] Computed 1/3/6/12-month and 12-1 momentum, and judged whether it is earnings-backed or multiple-driven.
- [ ] Checked up/down volume; India: checked delivery percentage; cross-referenced 13F / shareholding pattern, insider and promoter activity, and pledge changes.
- [ ] Recorded realised volatility, ATR%, beta stability and up/down capture.
- [ ] Charted drawdown history and identified the fundamental cause of each, and whether it still applies.
- [ ] Assessed liquidity: ADV, spread, free float, days-to-exit; India: circuit band, T2T, ASM/GSM, F&O ban status.
- [ ] Checked short interest and crowding (India: futures OI, basis, pledge overhang).
- [ ] Reviewed the last eight to twelve earnings reactions and identified pending non-fundamental flow events.
- [ ] Defined invalidation in fundamental terms first, then in price terms; wrote the staging plan and the maximum size cap.
- [ ] Investigated any persistent price-fundamental divergence and documented the conclusion.
- [ ] Used only the pre-committed indicator set; reported the price layer separately from the fundamental verdict and labelled it supplementary.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |