paper-writer
VerifiedMedical/scientific paper writing workflow skill. Manages the full pipeline from literature search to submission-ready manuscript. Creates and manages a project directory with IMRAD-format section files, literature matrix, reference management, and quality checklists. Supports both English and Japanese papers. Triggers: 'write paper', 'paper-write', 'start manuscript', '論文を書く', '論文執筆', '論文プロジェクト', 'manuscript', 'research paper', '原稿作成'.
templates/research-question.md
View on GitHub# Research Question Forge
> **Phase −1 — Discovery.** Runs *before* Phase 0 (Project Init).
> Input: a raw clinical observation **or** an existing question/protocol. Output: **one selected, sharpened research question**
> ready to hand to `references/novelty-check.md` and `templates/study-design.md`.
>
> This template implements the Google AI co-scientist pattern (generate → debate →
> rank → evolve) with the human as the irreplaceable final selector.
> See `references/ai-for-science-model.md §2` for why the 💡 IDEA gate is never automated.
---
## 0. Two entry modes
**Decision header:** Did the question come from a clinical spark (→ Mode A) or does a question/protocol already exist (→ Mode B)?
| Mode | Entry condition | Steps to run | Output |
|------|----------------|-------------|--------|
| **A — Forge** | Raw clinical observation; no prior question | §1 → §2 → §3 → §4 → §5 💡 → §6 | Full §6 record |
| **B — Resume/Refine** | A sharpened question or draft protocol already exists; study not yet locked | (a) PECO/PICOTS back-fill → (b) Attack pass → (c) FINER score (N = ?) → (d) §5 💡 re-confirmation → (e) §6 record | Same §6 record |
### Mode B — step-by-step
Run these four sub-steps in order. Skip Steps 1–4 of the forge entirely.
**(a) PECO/PICOTS back-fill.** Fill the §2 framing table for the existing question. Any element left blank is a precision gap — resolve it before continuing.
**(b) Attack pass (single).** The steelman is implicit — the question already survived to protocol stage. Run only the Attack half of §4 Step 2:
- Is the outcome measurable with data you can actually obtain?
- Is there an obvious confounder that would sink the interpretation?
- Would a positive result change anything, or is the effect size too small to matter clinically?
- Is the design implied by the question achievable in your setting?
- Does it risk HARKing — i.e., were you drawn to this question because you already suspect the answer from your data?
Any "yes" to the last point, or a hard "no" on measurability, is a design-change trigger, not a reason to abandon the question.
**(c) FINER scoring.** Score F / I / E / R as in §3. Leave N as `?` — run `references/novelty-check.md` next. A Σ below 15 (counting N = 3) is a signal to revisit the framing, not necessarily to kill the question.
**(d) 💡 Human re-confirmation (§5 sovereign gate).** Explicitly answer: "Is this still the question worth running?" This check is mandatory even for advanced studies — context, feasibility, and clinical relevance can shift between conception and data collection.
> Mode B still produces the identical §6 Output record (P/E/C/O/T/S, FINER scores, rationale, next steps). Downstream files (`references/novelty-check.md`, `templates/study-design.md`) consume §6 and are unaffected by which mode produced it.
---
## 1. Input: the clinical observation
Fill this in before running the forge. The puzzle must come from you — it cannot come from a literature gap AI found on its own. That distinction is the 💡 gate.
```
OBSERVATION RECORD
──────────────────────────────────────────────────────────────
What I saw / heard / measured in the clinic:
[e.g., "Parents arriving at my outpatient clinic increasingly mention
they consulted generative-AI tools before coming in. Several had delayed
presenting a febrile child after the AI told them to 'watch and wait.'"]
Why it surprised or unsettled me:
[What expectation was violated? What gap does it suggest?]
What I do NOT yet know that would change my practice if I knew it:
[The clinical knowledge gap — not a literature gap.]
Population I care about:
[e.g., infants <12 mo with fever, parents in urban tertiary-care setting]
Rough outcome I have in mind:
[e.g., delayed presentation, inappropriate antibiotic request, diagnostic accuracy]
──────────────────────────────────────────────────────────────
```
> 💡 **Human input required.** AI may prompt or scaffold this form, but cannot
> supply the observation. If the observation originates from an AI-generated
> literature gap rather than a clinical experience, label it clearly — the
> question will need extra novelty justification.
---
## 2. PICO / PECO / PICOTS framing
Choose the frame that matches your likely design before generating candidates.
| Element | Interventional (PICOTS) | Observational (PECO) | Meaning |
|---------|------------------------|----------------------|---------|
| **P** | Population | Population | Who exactly (age, diagnosis, setting) |
| **I** | Intervention | Exposure | What was done or what the patient was exposed to |
| **C** | Comparator | Comparator | The alternative; may be "no exposure" or standard care |
| **O** | Outcome | Outcome | Measurable; specify direction and time horizon |
| **T** | Time / follow-up | — | When outcome is measured |
| **S** | Study design | — | RCT, cohort, cross-sectional, SR… |
**Pediatric example (fills the form above):**
| Element | This example |
|---------|-------------|
| P | Parents/caregivers of febrile children aged 3 mo–5 yr presenting to a general pediatric outpatient clinic |
| E | Pre-visit use of a generative-AI chatbot for symptom assessment |
| C | No pre-visit AI use (sought advice from family, searched a conventional search engine, or no prior advice) |
| O | Time-to-presentation (hours from symptom onset to clinic arrival); secondary: antibiotic request rate at visit |
| — | Cross-sectional survey + chart review |
---
## 3. FINER scoring rubric
Score each candidate question 1–5 on each criterion. Sum → priority rank.
Novelty score is a placeholder (marked `?`) until `references/novelty-check.md` runs.
| Score | Feasible | Interesting | Novel | Ethical | Relevant |
|-------|----------|-------------|-------|---------|----------|
| **5** | Data exist, N achievable in your setting with current resources | Likely to change practice or policy if answered | No prior study found; fills a clear gap | No consent/safety concerns; IRB straightforward | Directly addresses a daily clinical decision |
| **4** | Minor resource gap (one extra funding step or collaboration) | Results would inform guideline update | 1–2 prior studies, conflicting or methodologically weak | Minor procedural requirements (waiver likely) | Addresses a decision you face weekly |
| **3** | Achievable but requires pilot or phased approach | Interesting to subspecialty audience; publication likely | 1–2 prior studies with moderate quality, your angle is incremental | Standard consent required; no special sensitivity | Relevant to your field but not your immediate practice |
| **2** | Requires external data, extended timeline, or significant funding | Primarily fills a textbook gap; limited clinical impact | Prior RCT or SR with consistent results; gap is minor | Sensitive population requires enhanced protections | Relevant to adjacent field; distant from your patients |
| **1** | Not achievable at your institution within a reasonable timeframe | Academic curiosity only; unlikely to change anything | Already definitively answered | Significant ethical barrier; IRB approval unlikely | Addresses a problem that does not exist in your setting |
**Maximum possible score: 25.** Questions below 15 total are usually not worth pursuing unless one criterion (typically Relevance or Feasibility) is a hard blocker that can be addressed.
---
## 4. The forge procedure
### Step 1 — Generate (diverge widely)
Ask AI to generate 5–15 candidate questions from your observation record. Instruct it to vary:
- **Outcome angle**: primary outcome, secondary outcome, process measure, patient-reported
- **Population angle**: narrow (infants only) vs. broad (all pediatric age groups); high-risk subgroup
- **Temporal angle**: acute event vs. long-term behavior change
- **Design angle**: descriptive → etiologic → interventional
- **Comparator angle**: active comparator vs. unexposed vs. historical
> 🤖 AI runs this step. Cast wide; odd angles often reveal the best question.
> Do not prune yet — the goal is breadth.
### Step 2 — Debate (steelman + attack each)
For each candidate, AI runs two passes in sequence:
**Steelman:** Why is this the *best* question in the batch? What would the finding enable?
**Attack:** What makes it weak?
- Is the outcome measurable with data you can actually obtain?
- Is there an obvious confounder that would sink the interpretation?
- Would a positive result actually change anything, or is the effect size too small to matter clinically?
- Is the design implied by the question achievable in your setting?
- Does it risk HARKing — i.e., is there temptation to pick this question because you already suspect the answer from your data?
> 🤖 AI runs steelman + attack. Human reads; begins informal ranking.
### Step 3 — Rank (FINER table)
Fill the FINER table for each surviving candidate. Leave Novelty as `?` — it requires the full `references/novelty-check.md` sweep to score honestly.
| # | Candidate question (one sentence) | F | I | N | E | R | Σ |
|---|----------------------------------|---|---|---|---|---|---|
| 1 | … | | | ? | | | — |
| 2 | … | | | ? | | | — |
| … | | | | | | | |
Sort by Σ (counting N as 3 until confirmed). Flag any question where:
- Any single criterion = 1 → likely a kill
- Feasibility < 3 → defer unless funding is imminent
### Step 4 — Evolve (sharpen top 2–3)
Take the top 2–3 by Σ. For each, apply:
1. **Narrow the population** — remove ambiguity about who is in and who is out.
2. **Operationalize the outcome** — name the instrument, the threshold, the time window.
3. **Name the comparator** explicitly.
4. **State the direction** — are you testing whether exposure *increases*, *decreases*, or *is associated with*?
5. **One-sentence PECO/PICO rewrite**: "Among [P], does [I/E] compared to [C] affect [O] measured at [T]?"
> 🤖 AI drafts evolved versions. Human reviews wording and clinical accuracy.
### 💡 STEP 5 — HUMAN SELECTS (sovereign gate — never automated)
**Stop. The AI does not choose.**
Present the evolved candidates to the research team (or to yourself, if solo). Decide:
| Option | Action |
|--------|--------|
| **Select** | Proceed to novelty check. Record rationale in Output Template (§7). |
| **Kill all** | Return to Step 1 with a revised observation or a different angle. |
| **Reframe** | Combine elements from two candidates; re-enter at Step 4. |
| **Defer** | The observation needs more clinical data before the question is ripe. |
> From `ai-for-science-model.md §2`: "An AI-generated hypothesis that no human chose
> to pursue is noise, not science." A selected question carries the researcher's
> judgment about what is *worth knowing*. Protect that.
---
## 5. Worked pediatric example
**Observation (filled form):**
Pediatric outpatient clinic, 愛育病院. Over several months, parents increasingly mention consulting a generative-AI chatbot before the visit. Three cases this quarter where a febrile infant's presentation was delayed; parents report the AI reassured them. Unsettled because delayed presentation in young infants risks missing serious bacterial infection. Population: infants and young children. Rough outcome: presentation delay.
**Six candidates generated (Step 1):**
| # | Candidate question |
|---|-------------------|
| Q1 | Among parents of febrile children <5 yr, does pre-visit generative-AI use increase time-to-presentation compared to no pre-visit AI use? |
| Q2 | Among parents who used a generative-AI chatbot before a pediatric visit, what proportion received advice inconsistent with published fever guidelines? |
| Q3 | Does pre-visit AI use predict inappropriate antibiotic request at the pediatric visit? |
| Q4 | What factors predict whether a parent uses AI vs. #8000 (Japan's pediatric telephone triage line) as a first contact for pediatric fever? |
| Q5 | Among parents who used generative-AI for fever triage, does the specific platform used (ChatGPT vs. other) affect advice accuracy? |
| Q6 | Can a structured AI chatbot intervention reduce unnecessary nighttime ED visits for non-urgent pediatric fever compared to standard parental education? |
**Steelman / attack highlights (Step 2):**
| # | Steelman | Attack |
|---|----------|--------|
| Q1 | Directly tests patient-safety outcome; time-to-presentation is chart-extractable | Causal interpretation blocked: parents who use AI may differ systematically from those who don't; observational confounding is severe |
| Q2 | Fully AI-auditable; could drive platform accountability | Defining "inconsistent" requires a guideline operationalization; no patient outcome attached |
| Q3 | Antibiotic stewardship angle; outcome is measurable at visit | Antibiotic request is a distal outcome of AI advice; effect size probably small |
| Q4 | Novel behavior-science angle; understanding the choice architecture is upstream of all other questions | Descriptive only; does not answer a clinical question; limited policy traction |
| Q5 | Platform comparison would interest regulators | Requires head-to-head access to multiple platforms at time-of-use; not feasible with chart data |
| Q6 | RCT design, highest evidence level | Requires intervention development, consent infrastructure, and long follow-up — not Phase −1 feasible |
**FINER table (Step 3):**
| # | F | I | N | E | R | Σ (N=3) |
|---|---|---|---|---|---|---------|
| Q1 | 4 | 5 | ? | 4 | 5 | 21 |
| Q2 | 4 | 3 | ? | 5 | 3 | 18 |
| Q3 | 4 | 4 | ? | 5 | 4 | 20 |
| Q4 | 5 | 3 | ? | 5 | 3 | 19 |
| Q5 | 1 | 3 | ? | 4 | 2 | 13 → kill |
| Q6 | 2 | 5 | ? | 3 | 5 | 18 → defer |
*Q5 killed (Feasibility = 1). Q6 deferred (long-horizon RCT). Top 2: Q1 and Q3.*
**Evolved candidates (Step 4):**
> **Q1-evolved:** Among caregivers of febrile children aged 1–60 months presenting
> to a general pediatric outpatient clinic, does self-reported pre-visit use of a
> generative-AI chatbot for symptom assessment — compared to no AI use — associate
> with longer time from symptom onset to clinic arrival (hours), after adjusting
> for fever height, child age, and parental education?
>
> **Q3-evolved:** Among caregivers presenting with a febrile child aged 1–60 months,
> does pre-visit generative-AI use associate with higher odds of explicitly
> requesting antibiotic treatment at first contact, compared to caregivers who
> used no AI or used the #8000 telephone triage line?
**💡 HUMAN SELECTS: Q1-evolved chosen.** Rationale: time-to-presentation is an objective chart-extractable outcome; Q3's antibiotic-request outcome requires visit-level documentation that is less reliable in current workflows. Q1 addresses the safety concern directly.
---
## 6. Output template — Selected Research Question record
Fill this after the human selects. This record is the handoff to novelty check and study design.
```
SELECTED RESEARCH QUESTION
──────────────────────────────────────────────────────────────
Full question (one sentence, PECO/PICO format):
[Q1-evolved text]
P: [Caregivers of febrile children aged 1–60 months, general peds outpatient]
E: [Pre-visit generative-AI chatbot use for symptom assessment]
C: [No AI use before visit]
O: [Time from symptom onset to clinic arrival, hours]
T: [Measured at visit (cross-sectional + chart review)]
S: [Cross-sectional observational, single-center]
FINER scores (N score confirmed after novelty check):
F [4] I [5] N [?→TBD] E [4] R [5] Σ = 21+N
Why it beat the alternatives:
[One paragraph: what Q1 can do that Q3/Q4 cannot]
Why I chose this question (clinical judgment, not AI judgment):
[Your sentence: what this would change in your practice]
Killed / deferred alternatives and reason:
Q3 — [outcome reliability concern]
Q5 — [not feasible]
Q6 — [defer to future funded study]
Next steps:
□ Novelty check → references/novelty-check.md
□ Confirm N score; re-rank if literature settles the question
□ Study design + power → templates/study-design.md
□ Pre-registration before data access → templates/preregistration.md
──────────────────────────────────────────────────────────────
```
---
## 7. Common mistakes
| Mistake | Why it fails | Fix |
|---------|-------------|-----|
| Question too broad ("What is the role of AI in pediatrics?") | No study design can answer it; no operationalizable outcome | Apply PECO; force a single measurable outcome |
| Outcome not measurable with available data | Data collection will fail or be biased at collection | Name the specific instrument or variable before scoring Feasibility |
| No real gap ("We wanted to confirm X") | Confirmation of known findings is not a research question | Run novelty check first; if question is settled, kill it |
| Design-led question ("We have a dataset so we asked…") | Question was shaped by available data, not clinical need | Return to the observation; redesign the question before touching the dataset |
| HARKing risk: choosing question to match suspected results | Pre-registered analysis becomes a sham; destroys confirmatory value | Pre-register before data access; any post-hoc question is labeled exploratory only. See `templates/preregistration.md` and `ai-for-science-model.md §5.1` |
| AI selects the question autonomously | Violates 💡 IDEA sovereignty; embeds AI's prior biases about what is "interesting" | Hard stop at Step 5; document who selected and why |
| Five-word question with no comparator ("Effect of AI on outcomes") | Uninterpretable; no direction, no comparison group | Apply PECO test: if any element is missing, the question is not ready |
---
## 8. Team-mode note
In a multi-agent setup, Steps 1–2 (generate + debate) map naturally to parallel
generator agents: each agent receives the observation record and independently
produces a candidate list, then a ranker agent synthesizes them into the FINER
table. This reduces anchoring — agents that do not see each other's candidates
produce a more diverse slate.
However: **Step 5 (human selection) is never delegated to an agent, a ranker, or
a consensus vote among models.** From `ai-for-science-model.md §6` (the autonomy
dial): the 💡 IDEA gate has no autopilot setting, regardless of how the rest of
the loop is configured. Document who selected and their explicit reasoning in the
Output Template — this is the audit trail that `templates/human-loop-ledger.md`
will record.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |