Skip to content

mikeparcewski/wicked-garden / skills/learn

wicked-garden-learn

Verified

Learn from outside sources, every claim cited. Learn mode: research a topic or event, synthesize a claims ledger (each claim with its source URL and observed_at), walk the operator through it, then author .wicked/knowledge/<slug>/ plus text-only skills that cite the claims; a refresh proposes only changes (confirm, supersede, contradict, retire), never a silent delete. Report mode (the `sourced-report` workflow): the same research and synthesis, without the write-up, from staged documents, the job's picked sources and the web, then a fidelity check by a seat that did not gather. It changes nothing. Use when: "learn about X", "research X and save what we learn", "what happened with <event>", "refresh the knowledge on X", "check the learned knowledge", a brief with a `Lens:` line, or a phase of the `learn` or `sourced-report` workflow. NOT for this repository's own code (wicked-garden-repo-learn) or for what the operator told you (a memory: wicked-garden-mem).

FileKindSize
SKILL.mdskill md10.7 KB
refs/claims.mdmarkdown2.7 KB
refs/refresh.mdmarkdown2.8 KB
refs/report-analyze.mdmarkdown1.2 KB
refs/report-check.mdmarkdown4.4 KB
refs/report-claims.mdmarkdown7.1 KB
refs/report-competitive.mdmarkdown1.2 KB
refs/report-critique.mdmarkdown1.3 KB
refs/report-due-diligence.mdmarkdown1.3 KB
refs/report-estimate.mdmarkdown1.4 KB
refs/report-research.mdmarkdown1.1 KB

refs/report-check.md

View on GitHub
# Report mode: the fidelity check

The last step of a report run. You are the evaluator: you did not gather or synthesize this report, and you
re-check every claim at its source yourself. If you gathered or synthesized any of it, say so and give no
verdict line (wicked-crew already keeps those seats off this step; this is the second guard).

You change nothing: write no file in any repository and commit nothing. Scratch goes in your notes folder,
outside every repository.

## 1. Run the floor

Copy the packet (the json block in `synthesize`'s output) to `packet.json` in your notes folder, then run,
with the documents folder's absolute path from `gather`'s preflight:

```
wicked-garden run scripts/learn/report_check.py check --packet <notes folder>/packet.json --records <documents folder> [--repo <repository>]
```

Paste its JSON and its exit code. Leave out `--records` only when the job staged no documents. Every finding
fails the run, with one exception: a `quote-not-found-in-pdf-text` where you read the pdf page and the quote
is there (the text extraction lost it). Say so for that source id. With no launcher or no Python, walk
refs/report-claims.md's manual alternative and say you did.

## 2. Re-check every claim at its source

| Source kind | How you re-check it |
|---|---|
| web | First match the record to a fetch receipt `gather` pasted in its output: the same URL, and `observed_at` equal to that receipt's `fetched_at`. A web record with no matching receipt is a finding (a page the job may never have fetched). Then fetch the URL again through the same web source (in a crew run, the `fetch` tool of the job's `web` source; outside one, your own web tools) and read it. Does it say what the claim says? |
| file | Open the file at the page and find the quote. For `brief`, first check that the packet's `brief` is the text the brief really carries. Does the claim follow from the quote, with the same numbers, dates, scope and hedges? A file the floor listed as `unchecked` is read in the original. |
| code | Open the repository file at the line and find the quote. Does the claim follow from it? |
| tool | Call the same tool of the same source with the record's `args`, read-only. Does the result still support the claim? A tool record you cannot call again (no `args`, the source refuses) leaves its claim unverified. |

Each claim ends in one of three states:
- **holds**: the source says it, and the claim says no more than the source does;
- **does not hold**: the source does not say it, says something else, or the claim stretches it (a number,
  a date, "all" for "some", a cause where the source gives a correlation);
- **unverified**: you could not reach the source again (the fetch was refused or failed, the pdf cannot be
  read, the call cannot be repeated). Name what failed. An unverified claim is not a held claim.

## 3. Judge the rest

- Every factual sentence in the report cites a claim. An uncited factual sentence is a finding.
- No source is invented: every URL and file is in the log, every web record matches a fetch receipt, and
  the log matches what the job could read. Read the report as it renders: every link a reader could open
  (a reference link, an email or `www.` address, an HTML link included) is a logged web source, or it is a
  finding. A web source when the packet says the web was not available is a
  fabrication.
- When the web was not used, the report says why, in plain words.
- The lens's shape is there (refs/report-<lens>.md): for example, an estimate states every assumption,
  and its price is effort × rate. Recompute every derived number.
- A stopped packet states a true reason: check that the documents or sources it says are missing are
  really missing.

## 4. Your output

1. **The checked report.** The report with every claim that does not hold, and every unverified claim,
   removed. They are listed under `## Did not hold at the check`, each with the claim id and why. For a
   stopped packet, its reason.
2. **Findings**, most severe first, each citing the claim id and the source id.
3. **Counts:** claims checked N, held H, did not hold D, unverified U, and the floor's exit code.
4. The last line is `VERDICT: PASS` only when the floor is clean (apart from extraction losses you
   confirmed in the pdf), every claim holds, and nothing in step 3 failed. Otherwise `VERDICT: FAIL`. A FAIL
   holds for the person, who sees the findings and the checked report.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall