메뉴
SUPERSEDED by /paper-draft (Pass B). Cold-read fidelity audit comparing each outline draft against integrated manuscript.tex. Retained for historical reference (referenced by past plans/*-revision.md); no longer auto-invoked.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
SUPERSEDED by /paper-draft (Pass A). Delta-based integration of revised outline prose into manuscript.tex; carried a hard-coded heading map now replaced by the PAPER.md anchors block. Retained for historical reference (it is referenced by past plans/*-revision.md); no longer auto-invoked.
Render the outline source-of-truth into manuscript.tex in two passes — the integrating phase of the paper factory. PASS A delta-integrates each `review`-bucket section's outline `## Draft` into manuscript.tex using the PAPER.md ```anchors``` heading map (never a hard-coded map), preserving every \cite and applying the invariants.md LaTeX prose conventions; `make build` must stay green. PASS B delegates to a FRESH subagent for two honest regimes — a fidelity cold-read (outline vs tex) and a genuinely BLIND claim + citation + invariant audit — which REPORTS and STOPS (no auto-fix). On human approval, fixes are applied and cleared sections advance review -> integrated. Absorbs and generalizes the legacy latex-integration-{first,second}-pass skills.
Human-gated applier of the paper factory's self-improvement findings. Reads the repo-root META.md via meta_status.py, reads the harness-log.md ledger (anti-thrash), shapes with the human which harness improvements to make, previews a concrete re-derived diff per fix, applies each as an atomic [harness] commit directly on draft, flips finding status open->applied/rejected/deferred in the same commit, and records every decision in harness-log.md. Meta/maintenance — NOT a lifecycle phase. Never touches PAPER.md's FSM, manuscript.tex, or the paper spine; only edits .agents/ (skills, templates, scripts, factory docs) + AGENTS.md/rules. Never weakens an invariant, never writes META findings, never recurses.
Fold research into section prose — the drafting phase of the paper factory. Picks one outline section (`outline/0N-*.md`), reads its `## Key Points`, the REFNOTEs for the C-IDs it `satisfies` and its topic (`outline/notes/refs/*.md`), and any `outline/snippets/`, then writes the section's `## Draft` prose: every Key Point covered, each non-obvious claim cited to a source, one voice, target_words as a soft cap. Advances the section `draft → review` via set_status.py, DRAFT-commits, and writes a session log. Strictly serial and single-voice, RE-RUNNABLE per section, one section per invocation by default. NO LaTeX here — that is /paper-draft; the `## Draft` block is the source of truth (see .agents/factory/methodology.md).
Ship the paper's `draft` branch to `main` — the release phase of the paper factory. Strips `DRAFT:` prefixes, fast-forward merges into `main`, and (opt-in) bumps the version, tags, and publishes a GitHub release that triggers the PDF build via CI. A RELEASE GUARD runs `make build`, `make check`, and `paper_status.py` first: check_paper.py errors (dangling cite, over page limit) and open must-fix items or `release_ready: false` are hard blockers requiring explicit human override. For a camera-ready release, packages the ACM TAPS zip via `make upload`. On success advances macro_phase to `released`. Default invocation is conservative (merge-and-push only); squash, bump, tag, release, and camera-ready are enabled by free-form instructions after the slash command.
Literature search and structured summarization for the paper, ONE source at a time and re-runnable (the additive analog of hs-plan's research fan-out). Given a bibkey, a URL, "find sources on <topic>", "next", or "status": fetches/reads the source, derives a bibkey, writes outline/notes/refs/<bibkey>.md from the REFNOTE template (key_findings + sample_sentences + supports: [C-IDs]), adds/verifies a references.bib entry, marks the source complete in the refs README checklist, tags the C-IDs it supports via set_status.py --add-evidence, and DRAFT-commits. Advances macro_phase scoped -> researching. Does NOT write section prose (that is /paper-outline). Second step of the paper-factory lifecycle (see .agents/factory/methodology.md).
| name | latex-integration-second-pass |
| disable-model-invocation | true |
| description | SUPERSEDED by /paper-draft (Pass B). Cold-read fidelity audit comparing each outline draft against integrated manuscript.tex. Retained for historical reference (referenced by past plans/*-revision.md); no longer auto-invoked. |
Superseded — retained for history. This cold-read fidelity audit is now
/paper-draft(Pass B), which runs it in a fresh subagent alongside a blind claim/citation/invariant audit. Kept because pastplans/*-revision.mdreference it by name.disable-model-invocation: truestops auto-firing; invoke/paper-draftfor new work.
Cold-read fidelity audit comparing each outline draft against the integrated
manuscript.tex. This is intentionally separated from the first-pass
integration to avoid self-confirming errors — the agent re-reads both sources
fresh and compares sentence-by-sentence without relying on memory of what it
just wrote.
Use this skill for task 5.2 of plans/second-revision.md (or any future
integration phase). Run after the first pass
(latex-integration-first-pass skill) has been committed.
Execute Subphase B of Phase 5: cold-read fidelity audit.
CONTEXT: Read `plans/second-revision.md` to confirm task 5.1 is complete.
The first-pass integration has been committed. This is a fresh audit pass.
PROCEDURE — for each section in order (00 through 06):
B1. RE-READ BOTH SOURCES. Read the outline file (`outline/0N-*.md`) and
the corresponding section in `manuscript.tex` from scratch. Do NOT
rely on memory from any previous pass.
B2. SENTENCE-LEVEL COMPARISON. Walk through the outline `## Draft` prose
sentence by sentence. For each sentence, confirm:
- The prose appears in the LaTeX (with appropriate LaTeX formatting)
- No words were dropped, substituted, or reordered
- No stale text from previous drafts remains in the LaTeX
- Paragraph breaks match the outline structure
B3. CITATION CHECK. For each `\cite{...}` in the LaTeX section:
- Confirm the citation key exists in `references.bib`
- Confirm it is attached to the correct claim or sentence
- Cross-reference the outline frontmatter `citations:` field:
every key listed there should appear in the LaTeX section
B4. FORMATTING CHECK. Verify LaTeX conventions are correct:
- ``...'' for double quotes (not "...")
- \emph{...} for italics where the outline uses *...*
- \texttt{...} for inline code where the outline uses `...`
- Non-breaking spaces (~) where appropriate
- No raw --- em-dashes (project convention: commas or reworded)
- Heading commands match the A4 heading map from the first pass
B5. REPORT DISCREPANCIES. Compile all findings into a summary list:
- Section name
- Line number in `manuscript.tex`
- Issue description (dropped word, stale text, misplaced cite, etc.)
Present this report and STOP. Wait for confirmation before applying
any fixes. Do NOT auto-correct — the user may want to review or
make judgment calls on specific items.
After fixes are approved and applied:
- Verify `make build` succeeds.
- Commit: `WIP: Third integration fidelity audit (second pass)`
with `Co-Authored-By: Oz <oz-agent@warp.dev>`
- Mark tasks 5.1 and 5.2 as `[x]` in `plans/second-revision.md`.