draft

name	draft
description	Draft a single section by loading outline slice, querying section's evidence cards, drafting in main session, running style-auditor, and applying safe fixes. Gate: outline must be approved.
disable-model-invocation	true
argument-hint	<section>

Draft a Section

This skill drafts a single section (Introduction, Methods, Results, Discussion, Conclusion, etc.). The draft binds every empirical claim to evidence cards, and runs style audit before finalizing.

Language: every user-facing string below (gates, aborts, summaries) is written in Turkish, the kit's conversational default (CLAUDE.md). If the active project's PROJECT.md sets Conversation language: en, present the equivalent content in natural English — translate meaning and structure (headings, bullets, ✅/❌ markers), don't reproduce the Turkish text verbatim. This is independent of Output language, which governs the drafted manuscript prose itself, not these gate/status messages.

GATE: Outline Must Be Approved

Check: Read outline/outline.md header.

if "**Status:** approved" not in header:   # /outline writes this exact bullet on approval
    ABORT with:
      "❌ İskelet henüz onaylanmadı.
       Lütfen /outline komutunu çalıştırarak ve 'evet' yazarak onaylayın."

If gate passes: continue.

---

PROCEDURE

1. Parse Section Argument

User provides section name (case-insensitive):

• introduction / intro / background
• methods / methodology
• results
• discussion
• conclusion / concluding-remarks
• abstract / özet
• literature-review / lit-review
• related-work

Normalize to canonical heading name from outline/outline.md.

If section not found: print options:

❌ Bölüm "[input]" bulunamadı.

Kullanılabilir bölümler:
  - introduction
  - methods
  - results
  - discussion
  - conclusion

Lütfen bölüm adını girin: /draft methods

2. Load Outline Slice

Query outline/outline.md for the chosen section. Extract:

• Full heading hierarchy (e.g., "I. Introduction", "I.1. Background", "I.2. Existing Work")
• RQ tags from HTML comments (e.g., <!-- RQ: rq1, rq2 | Cards: EV-0001... -->)
• Card IDs bound to this section

Example extraction:

Section: Introduction
Headings: I, I.1, I.2, I.3, I.4, I.5
RQ Tags: background, rq1, rq2, rq3
Card IDs: EV-0001, EV-0002, EV-0003, EV-0010, EV-0011, ...

3. Query Evidence Cards

For each RQ tag in the section, query cards:

python scripts/evidence_db.py query --tags "rq1" --min-strength 3

Load all returned cards into context (typically 5–15 per section). Extract:

• Claim (1 sentence)
• Detail (2–4 sentences, justification)
• Page (for quotes)
• Verbatim (if quote)
• Metrics (if quantitative finding)
• Source key (bib_key for citation)

Print progress:

✅ Bölüm kartları yüklendi: [N] kanıt kartı
   Örneğin: EV-0001 (strength: 4), EV-0002 (strength: 3), ...

4. Draft in Main Session

Now draft the section using the outline as skeleton and cards as evidence. Each sub-heading gets:

Sub-heading text: typically 3–6 paragraphs.

Paragraph structure (from academic-writing skill):

• Sentence 1: Claim / topic sentence
• Sentence 2–3: Evidence (cite card[s] via [@bib_key]) + detail from card
• Sentence 4: Interpretation / connection to RQ
• Sentence 5: Bridge to next paragraph or idea

Citation discipline:

• For every empirical claim: bind to a card ID via HTML comment after the sentence.

  Example: "CNNs have shown 95% accuracy on bearing fault detection [@Li2022DeepLearning]."
           <!-- EV-0045 -->
  

• For uncardable claims (general background, your own reasoning): tag [CITATION NEEDED].

  Example: "Bearing fault detection is critical for industrial safety. [CITATION NEEDED]"
  

Knowledge isolation — the three tags are not interchangeable:

Write factual content from the session's materials (evidence cards, analysis/ outputs, PROJECT.md, user-supplied text), not from model memory. This restricts FACTUAL CONTENT only — claims, numbers, citations, procedures. Academic conventions, argument structure, register and phrasing are yours to supply freely; that is the writing, not the evidence.

Tag	When	Blocks finalization
[CITATION NEEDED]	the claim is sound, a source for it isn't attached yet	yes
[MATERIAL GAP]	the outline needs content no session material covers	yes
[LLM-SUPPLEMENTED]	the user explicitly authorized filling a gap from model knowledge	no — reported

When the outline calls for something the materials do not cover, tag [MATERIAL GAP] and move on. Do not fill it from memory, and do not quietly soften the outline to avoid the hole — a plausible invented sentence is indistinguishable from a sourced one once written, which is precisely why the gap has to be marked at writing time rather than discovered later.

Surface every gap at the end of the section (Step 6 already reports counts). The user then either supplies material, or authorizes a fill — in which case replace the tag with [LLM-SUPPLEMENTED] so the decision leaves a trace. An authorized fill that leaves no trace is indistinguishable from an unauthorized one.

cite_audit.py counts all three; [LLM-SUPPLEMENTED] is reported and never blocks. Turkish drafts may use [KANIT YOK] and [MODEL BİLGİSİ] — both are recognized.

Tense discipline (per section):

• Introduction: present tense (literature facts), past tense (prior work)
• Methods: past tense (what you did)
• Results: past tense (what was found)
• Discussion: present/past mixed (interpret findings, position vs literature)
• Conclusion: present tense (implications, future)

Hedging examples (from academic-writing skill):

Strong: "X demonstrates..." / "X shows..." (experimental findings, high confidence)
Moderate: "X indicates..." / "X suggests..." (multiple studies, convergent evidence)
Weak: "X may suggest..." / "X may imply..." (single study, preliminary)

4a. Methods-Specific Completeness Check (Methods/Methodology Only)

Every other section drafts adequately from outline-skeleton + bound cards alone — Methods does not. A thin card set produces a thin Methods ("we trained four models" and nothing else), and the generic GATES SUMMARY warning ("no cards → synthesis-only, continue if user agrees") is not strong enough for a section whose whole job is enabling replication. Before drafting Methods prose, work through this checklist and draft to fill every applicable row — most rows are answerable from your own analysis scripts/RUNLOG.md (own-data studies) or from the bound method-type cards (literature-derived methodology), not from new research:

Checklist item	Source to check	If missing
Design — experimental / observational / quasi-experimental / simulation / secondary-analysis, stated explicitly	Outline + PROJECT.md Type	Ask the user; do not infer silently
Data/participants — N, source, inclusion/exclusion criteria, missingness handling	analysis/RUNLOG.md, own-result cards, PROJECT.md "Own data"	If a data-scientist ran the analysis, this MUST already be recorded — pull it, don't re-derive
Measures/variables — what each variable is, how it was operationalized/encoded	analysis/*.py scripts, method-type cards	Flag as a gap; do not guess an encoding scheme
Procedure — exact steps in the order performed (preprocessing → split → train → evaluate)	analysis/*.py filenames/order, RUNLOG.md	Reconstruct from script order; note if scripts are missing
Analysis plan / statistical tests — which test, why it fits the data/hypothesis, assumptions checked	data-scientist's numeric-results digest (Statistics discipline step)	If the data-scientist agent did not report this, ask it to before drafting, rather than writing the test name from memory
Reproducibility — seeds, package versions, hyperparameters, code/data availability	analysis/RUNLOG.md (data-scientist writes this every run)	Missing RUNLOG.md for an own-data study is a real gap — flag it, do not draft "results are reproducible" without the seed/version record backing it

For deeper guidance on any row, consult (read only the rows you need, not the whole file): .claude/skills/ml-research/reference/experiment-design.md (design + analysis-plan guidance) and reference/reproducibility.md (reproducibility checklist detail). For a second opinion on whether the Methods design fits the stated RQs/hypotheses, dispatch methodology-advisor (read-only, opus) — it is not normally in the /draft dispatch chain, so do this explicitly when the design is non-trivial or you're unsure a test/split choice is defensible, not by default for a routine comparison.

If a checklist row is genuinely inapplicable to this study (e.g. "participants" for a synthetic-data simulation), say so explicitly in a code comment in the draft rather than silently omitting the row — a reviewer should see it was considered, not guess whether it was missed.

Example draft section (Introduction):

## I. Introduction
<!-- RQ: background, rq1, rq2, rq3 | Cards: EV-0001, EV-0002, EV-0010, EV-0011 -->

### I.1. Background

Bearing components are among the most critical elements in rotating machinery, and their failure can cascade into catastrophic system shutdowns. Condition-based maintenance strategies rely on early fault detection to prevent unplanned downtime. <!-- EV-0001 -->
<!-- EV-0001: "Bearing failure costs industry $X billion annually..." -->

Modern approaches to bearing fault diagnosis employ signal processing and machine learning techniques. Traditional methods such as Fourier transform analysis and envelope analysis have been widely adopted. <!-- EV-0002 -->
<!-- EV-0002: Method card citing classical papers -->

### I.2. Existing Work

Deep learning has emerged as a powerful paradigm for time-series anomaly detection. [@Li2022DeepLearning] demonstrates that convolutional neural networks (CNNs) can outperform classical baselines on bearing fault datasets, achieving 94–97% accuracy across multiple benchmarks.
<!-- EV-0010 -->

However, questions remain about generalization to real industrial settings, where data distributions shift and labeled examples are scarce. [@Chen2023TransferLearning] suggests that transfer learning can address this gap, though validation remains limited.
<!-- EV-0011 -->

### I.3. Research Gap

Existing work has not comprehensively compared CNN architectures under realistic conditions: noisy, imbalanced, multi-bearing datasets typical of industrial deployments. [CITATION NEEDED] Furthermore, reproducibility concerns persist: most published models lack code availability statements or environment specifications.

### I.4. Research Questions

This work addresses three interconnected questions:
1. Which CNN architectures are most robust to noise and class imbalance in bearing fault diagnosis?
2. What datasets are available for benchmarking, and do they represent real-world scenarios?
3. How does CNN performance compare to signal-processing baselines under fair comparison protocols?

### I.5. Contribution

This article provides: (1) a systematic comparison of 5 CNN architectures on 3 public bearing datasets; (2) ablation studies isolating the impact of architecture components; (3) code and trained models for reproducibility.

The rest of this paper is organized as follows...

5. Run Style-Auditor (Sub-agent)

After drafting, dispatch style-auditor agent:

Input contract:

• Draft text (drafts/<section>.md)
• Output language (from PROJECT.md: TR or EN)
• Hedging audit: are claims appropriately hedged?
• Tense discipline: are tenses consistent per section?
• Terminology consistency: is "X" always called "X" (not sometimes "Y")?
• Register (academic vs colloquial): is tone appropriate?

Example style-auditor return (plain text, NOT JSON — one line-referenced item
per line: file:line · severity · rule-id · current → suggested, plus a 3-line
summary; style-auditor never returns structured JSON, it returns this exact
format so a cheap model can't accidentally produce malformed edits):
drafts/methods.md:12 · fix · article-usage · "the CNN models" → "CNN models" (drop "the")
drafts/methods.md:25 · review · hedging-overstrong · "CNNs are the best..." → "CNNs show promise..." or "CNNs outperform..."
drafts/methods.md:38 · fix · term-inconsistency · "bearing fault" vs "anomaly detection" → standardize to one term
Summary: 3 items (2 fix, 1 review); dominant problem: article usage; terminology drift pairs: bearing fault/anomaly detection.

6. Apply Safe Fixes

Main session (you, not the model) apply style-auditor's suggestions to drafts/<section>.md:

• ✅ Apply: grammar fixes, article fixes, hedging clarifications, term standardization
• ❌ Do NOT apply: rewrites that change meaning, subjective style opinions, contradictions of your intent

Print:

✅ Stil denetimi tamamlandı. [N] öneri uygulandı:
  - Makale kullanımı: 3 değişiklik
  - Terim tutarlılığı: 2 değişiklik
  - Eğilim: 1 öneri (gözden geçirin)

Dosya: drafts/introduction.md

7. Word Count Report

word_budget.py scans every file in drafts/ against PROJECT.md's budget in one pass — there is no per-section flag. Run it and read the entry matching this section's filename from the returned sections array:

python scripts/word_budget.py --json

Output:

Bölüm: introduction
Sözcük sayısı: 1,240 / 1,500 bütçe (82.7%)
[Breakdown by sub-section if available]

Kalan bütçe: 260 sözcük

Warn if over budget:

⚠️ introduction bütçeden 340 sözcük fazla. Lütfen gözden geçirin veya
/word_budget --trim introduction komutunu çalıştırın.

8. Final Output

Print:

✅ Bölüm taslağı tamamlandı!

Dosya: drafts/introduction.md
Sözcükler: 1,240 / 1,500
Kanıt kartları: 10 bağlantı
[CITATION NEEDED] sayısı: 2 (iddia sağlam, kaynağı henüz bağlanmadı)
[MATERIAL GAP] sayısı: 1 (oturumda bunu karşılayan hiçbir malzeme yok)
[LLM-SUPPLEMENTED] sayısı: 0

Sonraki adımlar:
1. /draft methods — sonraki bölümü taslak yap
2. /verify-citations — alıntı denetimi çalıştır (sonra tüm taslak bitmeli)
3. /status — ilerleme raporu

Not: [CITATION NEEDED] kaynak eksikliğidir — kaynak bulun veya iddiayı kaldırın.
[MATERIAL GAP] ise malzeme eksikliğidir: hafızadan doldurulmadı. Malzeme verin,
ya da doldurmaya açıkça izin verin — o durumda etiket [LLM-SUPPLEMENTED] olur ve
karar kayda geçer. İkisi de bitmiş bir metinde delik sayılır ve kapıyı bloklar.

Her [MATERIAL GAP] için tek satırla NE eksik olduğunu söyleyin — "hangi malzeme gerekiyordu" bilgisi olmadan kullanıcı boşluğu kapatamaz:

[MATERIAL GAP] §2.3 — örneklem büyüklüğü gerekçesi: güç analizi ya da
benzer çalışmaların N değerleri hiçbir kartta yok.

---

GATES SUMMARY

Gate	Trigger	Action
Outline not approved	Yes	ABORT, ask user to approve outline first
Section not found in outline	Yes	ABORT, list available sections
No cards for section	Yes	WARN: this section will be synthesis-only; continue if user agrees
Methods drafted with most checklist rows (4a) unfilled/unanswerable	Yes	WARN: list which rows are missing and why before finalizing; do not silently ship a thin Methods

---

SCRIPT USAGE

python scripts/evidence_db.py query --tags "rq1,rq2" --min-strength 3
python scripts/word_budget.py --json   # scans all of drafts/; read the matching section entry

---

STABILITY CHECK (optional, diagnostic — not a gate)

Drafting is LLM-mediated, so its accuracy is unmeasured and stays that way: no gold set exists and the kit refuses to invent one (eval/MEASUREMENT.md). Its stability can be measured with no labels at all — draft the same section 2-3× independently, into separate files, then:

python scripts/draft_stability.py runs/methods-1.md runs/methods-2.md runs/methods-3.md

Agreement is reported on the three things that have no honest reason to vary between renderings of one section from one evidence set: which sources are cited, which quantities are reported, and how strongly each source is claimed. Prose is never compared — two honest renderings differ in every sentence.

When to spend the tokens: a section carrying the paper's central numeric claims, or one where claim_check.py already flagged an overclaim. Read the result as variance only — a stage that is consistently wrong scores a perfect 1.00, and that sentence belongs in any report of the number.

---

EVIDENCE CARD BINDING FORMAT

Every empirical claim must have a card comment:

Claim sentence [@citation_key].
<!-- EV-0045 -->

Or if no source:

General claim [CITATION NEEDED]

Why: The comment establishes traceability. Later, /verify-citations will cross-check [@citation_key] against sources/registry.jsonl and evidence cards.

---

ERROR MESSAGES (Turkish)

❌ İskelet henüz onaylanmadı. Lütfen /outline çalıştırın.
❌ Bölüm "[section]" iskelet dosyasında bulunamadı.
⚠️ Bu bölüme hiçbir kanıt kartı bağlanmadı. Lütfen /lit-review veya /analyze-data çalıştırın.
⚠️ Bölüm bütçenizi aştı: [N] sözcük fazla.

---

FOLLOWING STEPS

After /draft <section>:

• User runs /draft <next-section> to draft another section.
• After all drafts: user runs /verify-citations to check bibliography.
• After verify passes: user runs /critique for peer review.