Imported from biprateep/science-skills (skills/co-writer/SKILL.md). Install upstream with npx skills add biprateep/science-skills --skill co-writer. Copyright stays with the author (MIT).

Co-Writer: Papers in the Author's Voice

Overview

co-writer does two things and only two: preserve the information in whatever it is given, and write it the way the author writes. The input's own style is discarded, not modelled — an agent artefact, a collaborator's draft and the author's own rough notes are all treated identically, as a bag of facts to be re-expressed.

The voice lives in one file, references/voice-profile.md: graded rules, a Never list, the words a de-AI pass must leave alone, and verbatim specimens indexed by section function. Read all of it, every time. Its specimens outrank its rules: where a rule and a specimen disagree, imitate the specimen and say what you departed from. Its provenance — the extraction prompt, the evidence report and the interview — is in references/ for maintenance and is not needed at write time.

When to Apply

Apply when the user asks for text in their voice, or to restyle, rewrite, draft or de-AI a piece of paper prose. Apply to any length: a sentence, a caption, a section, a whole draft.

Do not apply when the request is for the science itself, for structure or argument, or for a register the profile does not cover. co-scientist writes the report; co-writer restyles its prose afterwards. If a request mixes both ("write up these results"), do the writing here only after the results, the claims and the citations already exist in the input.

Process

1. Read the input for information, and write the content list. Before writing anything, list what must survive: every number with its precision, every claim with its hedge level, every citation and where it attaches, every defined symbol, every equation, every figure and table reference. For a paragraph this is a glance; for a section, and always for "de-AI this", write the list down — it is what step 4 composes from.
2. Identify the slot and the register. Slot: abstract, introduction opener, gap-and-contribution, methods, definition, equation, results, interpretation, limitation, close, caption. Register: journal (AAS, MNRAS) or workshop (four-page NeurIPS-style). Infer from the input and the surrounding document; if the user has said, that wins.
3. Load the profile, then read the specimen for the slot immediately before writing. The register is transferred from the specimen, not assembled from the rule list: read it at length and write while it is fresh. Spelling is American whatever the venue.
4. Write, in one sitting, from the content list. When the input reads as model prose — "de-AI this", or check_mannered.py lights up on it — close the input and compose from the list alone, because a writer with the old prose open keeps its sentences and its hooks. Otherwise the input may stay open, for information only. Apply every HARD rule, most STRONG rules, LIGHT rules where they fit. Carry the specimen's construction, never its nouns.
5. Check what is fixed. python <skill>/scripts/check_fixed.py INPUT OUTPUT (write both to the scratchpad in inline mode). It must print "fixed items identical", or only WARN lines after a ruled hard-limit cut. Then the self-check below.
6. The cold read (references/cold-read.md): always for a section or a draft; for a single paragraph, whenever an agent is available.
7. Citations, when the text is file-backed and a citing sentence changed — hand the changed instances to cite-check (see "Citations" below). A section or draft is not done until cite-check's audit is ok.
8. Return the text only. No preamble, no "Here is the rewritten version", no summary of changes, no closing offer. If the user asked for a diff or a rationale, give the text first and the rationale after a rule. Never certify the voice; the passages you are least sure of go into the session log, not into the reply.
9. Log it. Append the entry on the form in references/session-log.md, in the paper's repository. A rewrite that is not logged teaches nothing.

Hard Rules — Preservation

These are correctness rules and outrank every voice rule. A rewrite that reads exactly like the author and breaks one of these is a failure.

• Numbers survive exactly. Value, precision, unit, sign, uncertainty. 0.00898 does not become "about 0.009". A number the input lacks is not invented; write [NEEDS: value] and say so after the text.
• Modality is never strengthened. The profile's claim ladder, weakest to strongest: suggests < is consistent with < indicates < shows < demonstrates < proves; find / observe / see are reporting, not claims; negations and "we cannot rule out" sit below everything. A claim may stay where it is or move down; "suggests" may not become "demonstrates", "we cannot rule out X" may not become "we show X". One exception, from the author: the abstract may sit one rung above the results section for the same claim — one rung, never two. Weakening elsewhere is allowed only if the user asked.
• Flourish is not information. An evaluation with no checkable content — "a substantial leap forward", "proved pivotal", "remarkable" — is style and may be dropped. A comparison with a referent — "outperforms the CNN baseline", "important for this performance" — is a claim and survives.
• No reference is invented. The profile's results pattern opens with "Figure N shows"; if the input has no figure, do not add one. Never introduce a \ref, \cite, equation number or table to satisfy a slot.
• The citation set never grows or moves. keys_out ⊆ keys_in, and every \cite key in the output is attached to the claim it was attached to in the input; a key drops out only under a reported hard-limit cut. This is a set comparison, needs no lookup, and is the only citation check co-writer performs itself. Whether the paper exists, whether the .bib entry is genuine, whether the paper supports the sentence — that is cite-check's, below.
• Markup passes through untouched. \cite, \ref, \label, \eqref, math (inline and display), \begin{figure} … \end{figure} bodies, macros. Rewrite the prose around them. If the input is LaTeX, the output is LaTeX.
• Specimen content does not leak. The profile's quarantine list names the nouns; if any phrase of eight or more words from an specimen appears in the output and was not in the input, remove it.
• Hard limits are met by the trim (references/trim.md), never by writing shorter or by compressing a claim away. What leaves under a ruled cut is listed after the text.

Citations — cite-check does the verifying

co-writer has no citation logic beyond the set comparison above. Existence, provenance and claim support are computed by the cite-check skill (skills/cite-check/, a sibling of this one), and every improvement made there applies here with no change to this file. Before the first citation action in a session read its contract, <skills>/cite-check/references/integration.md: it says how to reach the tools (listed by the harness as mcp__cite-check__*, or run its mcp/setup_mcp.sh once and use the identical CLI for the session, ping once for the ADS token) and it wins wherever this section differs.

Three triggers:

1. A rewrite changed a sentence carrying a citation. extract_cites on the rewritten file, cite-check's support check on every instance whose claim changed, then audit(tex_path, require_support=true). Two rules must both hold: co-writer's, that the sentence is no stronger than the input's; cite-check's, that the paper supports the sentence as written.
2. The user asks for a citation, or the input carries a gap — [cite], \cite{?}, "(REF)". cite-check's Workflow A: search_citation → a high candidate, ask about a medium, never a low → bib_add → write \cite{<key>} → support check. no-official-bibtex → leave it uncited and say so. This is the only way a key enters the output that was not in the input.
3. A draft the user calls finished. audit returns ok: true; relay its failures and warnings verbatim and do not lower require_support.

Inline mode — no file, no .bib — gets the set comparison only, and one sentence saying the support check runs when the paragraph is in its file. Never: write or complete a .bib entry; cite a key, DOI, arXiv id or bibcode from memory; record a support verdict yourself; re-implement a cite-check check with your own reasoning; edit anything under cite-check/mcp/.

Modes by Input

Anti-Patterns

What a model reaching for "good academic prose" will do, and must not:

• Stack signatures. Every paragraph ending "Therefore,", every methods sentence opening "To …", "(also called …)" three times on a page. One or two per paragraph is the author; five is a parody.
• Scrub the author's words as AI-isms. The profile's "words that are his and stay" list exists because a generic de-AI pass removes them.
• Add dashes. The corpus has at most one per page. A rewrite with an em-dash aside in every paragraph is the single loudest tell.
• Add "Notably," / "Importantly," / "Interestingly,". Surprise is marked with remarkably / unsurprisingly / as expected, once.
• Shorten paragraphs into punchy ones, or add a one-sentence paragraph for emphasis. His paragraphs are long and even.
• Convert narrated procedure into a list. "We first … We then …" stays prose.
• Smooth away the concessive sentence. "While X, Y" is how caveats are bound to claims; splitting it into two sentences loses the voice.
• Strengthen a claim to make the paragraph land. See Preservation.
• Move a limitation into its own section. It lives beside the result.
• Import a framing from the specimens. The DESI/photo-z nouns are content.
• Conjure a figure. "Figure N shows …" is how he opens a results paragraph when there is a figure; it is not a licence to reference one the input does not have.
• Add a summary sentence at the end of a section that the input did not have.
• Verify a citation by reasoning. "This DOI looks right", "I recall that paper shows this" — not a check. cite-check's tools are the check.
• Revise against the list. When a stretch misses, re-read the specimen and write it again. Patching sentence by sentence against the Never list yields prose that satisfies every rule and still reads as a model.
• Certify the voice. "This sounds like you" is not a finding. Only the author says that; the writer records what it is least sure of.
• Write a preamble or postamble. Text only.

Self-Check Before Finishing

1. check_fixed.py INPUT OUTPUT prints "fixed items identical", or only WARN lines after a ruled cut. ☐
2. No claim is stronger than it was; the abstract at most one rung above the results. ☐
3. Each \cite key is attached to the claim it was attached to (the script checks the set; you check the attachment). ☐
4. File-backed and a citing sentence changed → cite-check's support pass run; a section or draft is called done only with audit ok. ☐
5. No phrase of eight-plus words from a specimen that was not in the input. ☐
6. check_mannered.py OUTPUT read: every hit looked at, the dash tally at or under one per page. A hit is a place to look, not a fault. ☐
7. Nothing from the Never list. ☐
8. No signature stacked more than twice in a paragraph. ☐
9. American spelling throughout. ☐
10. The cold read has run where the mode requires it, and each surviving item was answered by rewriting from the specimen. ☐

Maintenance

The profile improves from use through the loop in references/maintenance.md: capture the author's edits, collect the transcripts, digest them into candidates he adopts, tests or holds, score each version on the eval set, re-extract when a paper is finished. Nothing there is read at write time.

Interaction With Other Skills

co-writer reads its own references/ and, for citation work, cite-check's references/integration.md; it does not load or defer to any other skill's writing guidance.

• cite-check is the citation engine. co-writer keeps only the set comparison; existence, BibTeX and claim support are cite-check's, called as in "Citations". Improvements to cite-check reach co-writer with no change here.
• co-scientist produces the science and assembles the report; co-writer is run afterwards, section by section, over the prose. co-scientist's manifest says which claims are verified; co-writer preserves claims, it does not vet them.
• plot-style makes the figure; co-writer writes its caption in the author's voice (the profile has a caption slot and specimen).

Files — and when each is read

skills/co-writer/
├── SKILL.md                         # read whole when the skill is invoked
├── references/
│   ├── voice-profile.md             # every rewrite — THE artefact; specimens outrank rules
│   ├── session-log.md               # every rewrite, at the end — the form of the log entry
│   ├── cold-read.md                 # section and draft mode — who reads, what it gets, the brief
│   ├── counterexamples.md           # once after drafting — sentences the author rejected
│   ├── trim.md                      # only with a page or word limit — protocol and proposer's brief
│   ├── maintenance.md               # never at write time — capture, collect, digest, eval, new paper
│   ├── extraction.md                # never at write time — the deep-read prompt
│   ├── extraction-report.md         # never at write time — evidence for every rule
│   └── interview.md                 # never at write time — the questions the text cannot answer
├── scripts/
│   ├── check_fixed.py               # INPUT vs OUTPUT: cites, refs, labels, math, numbers — the preservation gate
│   ├── check_mannered.py            # stock phrases and Never-list hits, with lines — calibration, not a verdict
│   ├── extract_prose.py             # .tex → readable prose; --counts gives words per section for the trim
│   ├── capture_edits.py             # session log + author's edits → corrections.md
│   └── collect_transcripts.py       # Claude Code / Antigravity transcripts → .co-writer/transcripts/
└── eval/
    ├── README.md                    # the per-version protocol
    └── 01-abstract.tex … 06-caption.tex   # fixed inputs, one per slot

skills/cite-check/references/integration.md   # before any citation action — the contract this skill follows

<paper repo>/.co-writer/             # gitignored, written at use time
├── log/                             # one file per rewrite + one session note
├── corrections.md                   # from capture_edits.py
└── transcripts/                     # from collect_transcripts.py