Scholar forge

Scholar Forge

One skill, full academic lifecycle — integrity-first, venue-aware, bilingual.

Non-Negotiables

1. No fabricated citations — every reference must be located and verified (title, authors, venue, year, DOI/URL). Mark unverified as [UNVERIFIED].
2. Evidence-first — non-trivial claims need a citation or an explicit result in the document.
3. Label source status — peer-reviewed / preprint / grey literature / web.
4. Venue before prose — fetch target journal or school guidelines before drafting (see venue-intelligence.md).
5. Two-stage writing — outline with claim-evidence map → prose. Never skip outline for full manuscripts.
6. Compile or validate — after LaTeX edits, run latexmk or report the specific error.

Mode Router

Identify the user's task and load the matching task file before writing.

Mode	Triggers	Task file
Literature review	综述, survey, PRISMA, systematic review	tasks/literature-review.md
Research paper	IMRaD, conference, journal article	tasks/imrad-paper.md
Thesis / dissertation	学位论文, PhD, master's thesis	tasks/thesis.md
Grant / proposal	基金, NSF, NSFC, 国自然, R01	tasks/grant-proposal.md
Peer review	review this paper, 审稿意见	tasks/peer-review.md
Revision / rebuttal	rebuttal, response to reviewers, 修回	tasks/revision-rebuttal.md
Material integration	chat log, notes → paper, vibe writing	tasks/material-integration.md

If multiple modes apply, confirm priority with the user. Default order: clarify requirements → venue intelligence → discovery → outline → prose → QA.

Execution Plan (show first)

SCHOLAR FORGE — EXECUTION PLAN
==============================
Mode:           [from router]
Target venue:   [journal / school / funder]
Language:       [EN / 中文 / bilingual abstract]
Citation style: [IEEE / APA / ACM / GB/T 7714 / venue-specific]
Reporting:      [IMRaD / PRISMA / CONSORT / STROBE / none]
Output:         [LaTeX + .bib / Markdown / DOCX via pandoc]
Venue profile:  [PENDING → fetch before writing]

Phase 1 — Requirements (ask if missing)

Essential: document type, topic/contribution, target venue, length limit, citation style, deadline, language.

Use this intake template when helpful:

## Scholar Forge Request
**Mode:** [paper / review / thesis / grant / rebuttal / integrate]
**Topic:** [...]
**Venue:** [Nature Methods / IEEE TNNLS / 清华大学学位论文 / NSFC 青年基金]
**Length:** [words or pages]
**Citation:** [auto from venue or specify]
**Data/files:** [@mentions]
**Language:** [EN / 中文]

Phase 2 — Venue Intelligence

Always run for papers, theses, and grants. Do not rely on stale knowledge.

1. WebSearch official author guidelines + AI authorship policy for the target venue.
2. WebFetch the guidelines page; extract limits, citation style, abstract structure, figure limits, template URL.
3. Save writing_outputs/<project>/venue_profile.yaml using the schema in venue-intelligence.md.
4. For Chinese theses, also load chinese-academic.md.

Phase 3 — Discovery & Verification

Search strategy

Domain	Primary sources
CS / Engineering	Semantic Scholar, arXiv, ACM DL, IEEE Xplore
Biomedical	PubMed, PMC, bioRxiv
General	OpenAlex, Crossref, Google Scholar
Chinese	CNKI 知网, 万方, 维普 (user may need institutional access)

Query patterns: broad survey → recency filter → exact phrase → snowball from 2–3 anchor papers.

Per-source record

## [citation_key]
- Title / Authors / Year / Venue
- Status: peer-reviewed | preprint | grey
- DOI / URL
- Takeaway (1–2 lines) | Use in: [section]

Maintain references.bib as single source of truth. Run validation when scripts are available:

python scripts/validate_bib.py references.bib
python scripts/check_citations.py paper.tex references.bib

Details: integrity-protocol.md, citation-formats.md.

Phase 4 — Outline & Claim-Evidence Map

Before prose, produce:

1. Section outline with word budgets per venue profile.
2. Claim-evidence map (template: templates/claim-evidence-map.md).
3. Figure/table plan with one-sentence takeaway each.

Phase 5 — Writing

Universal prose rules

• Flowing paragraphs in Abstract, Introduction, Discussion, Conclusion — no bullets unless venue allows structured abstract.
• Active voice preferred ("We measured…").
• Hedge appropriately; avoid AI tells: "remarkably", "delve", "landscape", "it's important to note", "in conclusion" opener.
• Match the paper's established voice; add transition sentences between inserted blocks.

Citation styles

Style	Typical use	Guide
IEEE	CS, engineering	citation-formats.md
APA 7	social sciences	same
ACM	computing venues	same
GB/T 7714	中文学位论文、期刊	chinese-academic.md

LaTeX

• Prefer official publisher templates from venue profile.
• Scaffolds: templates/ieee-conference.tex, templates/gb7714-thesis.tex.
• Compile: latexmk -pdf -bibtex paper.tex (or -usebiber for biblatex).

Phase 6 — Quality Assurance

Run before delivery:

Integrity audit

• Every \cite{} key exists in .bib
• No [UNVERIFIED] entries in final bibliography
• Claim-evidence map satisfied for major claims
• Retraction check for critical citations (Crossref / Retraction Watch if available)

Venue compliance

• Word limits (abstract, main text, references)
• Required sections and abstract structure
• AI disclosure statement if required by venue

Writing audit

• AI-voice patterns removed (see writing-quality.md)
• Abstract matches paper content
• Limitations and threats to validity stated

Reproducibility (if empirical)

• Data/code availability statement
• Seeds, splits, baselines, compute budget documented

Output Deliverables

1. Source files (.tex, references.bib, venue_profile.yaml)
2. Brief change log: what was written, integrated, or verified
3. Compilation status or validation script output
4. Open items flagged for user (paywalled papers, unverified claims)

Additional Resources

• integrity-protocol.md — verification, retraction, predatory journals
• venue-intelligence.md — live guideline extraction schema
• chinese-academic.md — GB/T 7714, thesis norms, CNKI workflow
• writing-quality.md — AI-voice audit, hedging, transitions
• reporting-guidelines.md — PRISMA, CONSORT, STROBE, ARRIVE