project-specific-writing-exemplar-workflow

Build and route a manuscript-specific academic-writing exemplar corpus, with a single Core 3 approval gate and anti-copying controls. Use for substantive manuscript drafting, major rewriting, target-j

SKILL.md

GitHub

Per-manuscript writing exemplar workflow

Use this workflow to route a manuscript through a private, isolated exemplar lifecycle and learn reusable scientific argument architecture from verified papers without turning source prose into a copying library. This is an orchestration subskill under the existing agent-auto-sci-scicomm writing owner, not a competing writing owner. Manuscript-specific facts, private corpus records, approval decisions, derived patterns, and research decisions belong in that manuscript's private state, not this public skill.

Trigger and boundary

Trigger for substantive Introduction, Methods, Results, Discussion, Abstract, or major manuscript rewriting. Do not trigger for a typo, one-sentence grammar fix, citation-format change, reference-style change, simple caption correction, or simple translation.

Resolve a stable MANUSCRIPT_ID from explicit input, the active project's private state, canonical title/research question, and source-document identity. A filename, title string, or chat title alone is not a stable ID. If uncertain, mark MANUSCRIPT_IDENTITY_UNCERTAIN, do not inherit another skill, and seek only the minimum clarification. Read the one authoritative private state record; never infer the 20-paper list from chat memory. Codex and ChatGPT Web use this same record.

Before substantive writing, check whether the matching manuscript has a current project-specific exemplar profile/skill. If absent or stale, start that manuscript's initialization. If active, select 4–5 section-specific exemplars from that manuscript's own Master 20 and freshly read their relevant sections before drafting. Micro-edits bypass heavy workflow; an existing skill's lightweight constraints may be read.

Lifecycle

S0_UNINITIALIZED -> S1_MANUSCRIPT_PROFILED -> S2_CANDIDATE_POOL_READY -> S3_MASTER_20_SELECTED -> S4_CORE_3_PENDING_AUTHOR_APPROVAL -> S5_CORE_3_APPROVED -> S6_CORPUS_DEEP_ANALYSIS_COMPLETE -> S7_PROJECT_EXEMPLAR_SKILL_BUILT -> S8_ACTIVE_FOR_WRITING -> S9_REFRESH_REQUIRED

Each manuscript independently builds a recommended 40–80-paper pool, selects exactly 20 for its own Master Corpus, then recommends Core 3 from that set. Stop in S4 until the author explicitly approves, replaces one/two, or rejects the Core 3 proposal. No deep analysis or major rewrite before approval. Approval is the only author gate and automatically unlocks Phase 2; do not ask again. See workflow and state machine for transition, refresh, and storage contracts.

Selection rules

  • Build a broad pool before narrowing; deduplicate by DOI, then authoritative title/author/year checks.
  • Prefer original empirical research and the user's verified Zotero/full-text sources. Use reviews sparingly for unusually strong synthesis architecture.
  • Rank scientific-question, object, outcome, design, evidence-chain, planning/theoretical, target-journal, writing, and full-text proximity above impact factor.
  • Keep CORE, STRONG, SUPPORT, and EXCLUDE as categorical states; never invent a pseudo-precise score.
  • Core 3 should be scientifically closest and, where compatible, functionally complementary: story/question, evidence chain, and discussion/planning contribution.
  • Record missing metadata or unavailable structure explicitly as NOT_VERIFIED, FULLTEXT_NOT_YET_AVAILABLE, or PARTIAL.

Analysis boundary

Before approval, do not perform paragraph, sentence, lexical, phrase-frequency, or NotebookLM batch analysis. Do not create a project-specific writing Skill and do not modify the manuscript. After approval, analyze four levels: article story; paragraph functions (including what is lost if removed); sentence functions (why each sentence exists); functional lexical/phrase roles. Do not report ordinary word frequency as a learning result. After cross-paper quality review, patterns may be corpus-stable or high-quality/low-frequency. Never store long source passages or distinctive phrases. Explicit author constraints outrank frequency.

Writing-time safeguards

Build OUR_EVIDENCE_MAP and OUR_ARGUMENT_MAP before exemplar comparison. For each substantive section, select a fresh set of 4–5 from this manuscript's own Master 20; Core 3 are preferred but not mandatory, and section sets may differ. Reread selected source sections and restore paragraph, sentence-function, and functional-lexicon maps. Follow OUR_EVIDENCE_MAP -> OUR_ARGUMENT_MAP -> SECTION_EXEMPLAR_REFRESH -> STRUCTURAL_COMPARISON -> OUR_DRAFT_OR_REVISION -> ACADEMIC_PROSE_GUARD -> REDUNDANCY_AUDIT -> FIGURE/TABLE_CITATION_AUDIT -> FACT_PLACEMENT_AUDIT -> FINAL_CONSISTENCY_CHECK. Do not transfer source claims, causal interpretation, results, citation chains, or distinctive wording. User constraints override frequency.

Detailed schemas and routing rules are in references/schema.md and references/routing.md.