--- name: grant-section-drafter description: > Draft one section of a grant application, manuscript, cover letter, addendum, or any extended prose that must read in the researcher's own voice. Use when Heath asks for a first-pass draft of a Significance section, a Specific Aim, an Approach narrative, a rebuttal, a cover letter paragraph, or any piece of scientific prose longer than ~200 words that will be reviewed line-by-line or submitted to a funding agency, journal, or search committee. Do NOT use for short chat replies, tool-output summaries, or bullet-point notes. --- # Grant Section Drafter ## When to Use Trigger this skill whenever the request involves drafting extended prose that carries institutional weight — grant sections, manuscript sections, cover letters, NIH-style Specific Aims, addenda, progress reports, or response-to-reviewer text. The test: would a program officer, journal editor, or search-committee member eventually read this? If yes, use this skill. If the output is a chat reply or an internal summary, do not use it. The skill is also activated automatically by the `nightly_grant_drafter` job (`agent/jobs/nightly_grant_drafter.py`), which reads the researcher's linked artifacts overnight, identifies the next unfinished section via a gap-finder call, and invokes this skill's principles to draft that section. --- ## Voice Principles All prose produced under this skill must match the researcher's published writing register. The governing rules are drawn directly from `_DRAFTER_SYSTEM` in `agent/jobs/nightly_grant_drafter.py` and the `SCIENTIST_MODE` constant in `agent/jobs/__init__.py`: **Direct, quantitative, concrete.** Claims need numbers, not adjectives. Instead of "we have extensive experience with karyotype analysis," write "we have assembled and curated eight open karyotype databases totaling 85,000+ records across six major taxonomic groups." Vague abstractions are rejected at draft time, not during review. **No corporate register.** The prose must not sound like a consulting deck or a software company's marketing copy. The reader is a scientist. Write like one. **Calibrated assertion.** The researcher's published writing distinguishes hypothesis from finding, correlation from causation, single-study result from replicated pattern. Drafts must mirror that calibration. Never overstate what the preliminary data show. If the effect size is uncertain, say so plainly rather than smoothing it away. **Terse.** Skip validation-forward openers ("In this proposal we will demonstrate that..."), meta-commentary, and padding. Get to the claim. --- ## Banned Phrases The following words and phrases are prohibited in any output produced by this skill. Their presence signals AI-assistant cadence or corporate register — both fatal to grant prose: - comprehensive (as a modifier: "comprehensive framework", "comprehensive analysis") - robust (as a vague quality claim: "a robust approach", "robust methods") - holistic - leveraging (as a verb applied to anything non-mechanical) - cutting-edge - queryable (outside a database-schema context) - paradigm-shifting - novel — permitted only after explicit comparison to existing work that makes the novelty specific ("novel relative to Smith et al. 2019, which did not account for...") - revolutionary, groundbreaking, transformative (any superlative that is not a direct quote from a reviewer) - "In this proposal we will..." - "a comprehensive framework" - "state-of-the-art" When reviewing a draft for these phrases: flag every instance with a concrete alternative, as required by the `` stance. Do not silently delete them — name the phrase and the replacement. --- ## Preliminary-Data Discipline Every claim about preliminary results must be grounded in three components: 1. **n** — the sample size or number of observations behind the claim 2. **test** — the statistical or analytical method applied 3. **effect** — the direction and magnitude of the result Format: "We find [direction + effect] (n = [N]; [test])." Example: "Karyotype instability correlates positively with genome size across Coleoptera (n = 847 species; Pearson r = 0.41, p < 0.001, PGLS controlling for phylogeny)." When one of these three components is absent from the notes or context available: do not invent it. Write the placeholder exactly as: ``` [Heath: confirm n / test / effect] ``` This placeholder is intentional: it flags real gaps for the researcher's morning review rather than propagating confident-sounding fabrications into a grant draft. The `nightly_grant_drafter` job uses this convention and the researcher's review workflow (`review_overnight_draft`) is designed around it. --- ## Limitations-Acknowledgment as Default Strong grant prose acknowledges the scope conditions on every major claim. This is not weakness — reviewers at NIH, NSF, and top journals reward intellectual honesty more than bravado. Per the drafter system prompt: > "Acknowledge limitations explicitly when you make a strong claim — what would > falsify it, what scope is conditional." Operationally: after every bold positive claim, add one qualifying sentence that names the boundary condition or the observation that would falsify the claim. Omit this only when the claim is already narrow enough to be self-limiting. --- ## Section Matching The `nightly_grant_drafter` job passes a `section_label` and a `context_already_present` block from the gap-finder call. This skill must match both: - **Section label** — the draft's heading and scope must match the section being filled (e.g., "Significance", "Approach Aim 2", "Broader Impacts"). Do not expand or contract scope. - **Context already present** — the surrounding sections already contain claims, citations, and structure. The draft must not repeat or contradict what is already written. Read `context_already_present` before drafting a single sentence. In chat mode (not the nightly job), the researcher will usually supply the section context verbally. If he does not, ask for one sentence: "What section is this and what's already written around it?" --- ## Voice Exemplars Before drafting, retrieve voice exemplars from the TF-IDF index: ```python from agent.voice_index import retrieve_exemplars exemplars = retrieve_exemplars(query=section_label + " " + what_is_missing, k=4) ``` Or call the convenience wrapper used by the drafter job: ```python from agent.voice_index import voice_system_prompt_addendum addendum = voice_system_prompt_addendum(query, k=3) ``` The addendum is prepended to the system prompt as a style anchor. It surfaces passages from the researcher's published prose (Discussion sections, grant narratives, lab-website text) ranked by TF-IDF cosine similarity to the draft topic. Match their: - **Density** — how many claims per sentence - **Hedging level** — how explicitly uncertainty is flagged - **Quantitative specificity** — how often numbers appear in support of claims If `voice_system_prompt_addendum` returns an empty string (index not yet populated), proceed with the voice principles above and flag that the index was unavailable. --- ## Output Format - Output is **markdown only** with no preamble. The first line of the output is the section content, not a meta-comment about it. - No "Here is the draft:" opener. No "I've drafted the following section:" closer. - Headings, if appropriate to the section type, use `##` or `###`. - Citation placeholders use `(Author et al., YEAR)` format until real DOIs are confirmed. - The `citation_suggester` module in `agent/citation_suggester.py` is invoked after drafting by the nightly job; in chat mode, suggest 1-3 specific citations inline using your knowledge, flagged as `[suggest citing: ...]`. --- ## Integration with the Nightly Job The full automated flow in `agent/jobs/nightly_grant_drafter.py`: 1. For the active project with the soonest deadline, read its linked Google Doc artifact via `read_drive_file`. 2. Send the artifact to a gap-finder Sonnet call that returns `section_label`, `what_is_missing`, and `context_already_present`. 3. Retrieve voice exemplars for the topic. 4. Draft the section using this skill's rules (the `_DRAFTER_SYSTEM` prompt encodes them verbatim). 5. Run a critic pass (`agent/critic.py`, `rubric_name="grant_draft"`). 6. Create a new Google Doc tagged `[draft]` — never overwrite the source artifact. 7. Insert a row in `overnight_drafts` and a briefing of urgency `warn` or `critical` (based on critic score) so the draft surfaces in the morning. The `_DRAFTER_SYSTEM` constant (lines 62-78 of `nightly_grant_drafter.py`) is the canonical machine-readable form of this skill; this SKILL.md is the human-readable companion. --- ## Anti-Patterns Checklist Before finalizing any draft, check each item: - [ ] Contains no banned phrases (see list above) - [ ] Every preliminary-data claim has n / test / effect, or a `[Heath: confirm ...]` placeholder - [ ] At least one limitation or boundary condition named per major claim - [ ] No validation-forward opener ("In this proposal we will...") - [ ] Register matches voice exemplars (density, hedging, quantitative specificity) - [ ] Section scope matches `section_label` — not broader, not narrower - [ ] No fabricated citations, file paths, or numerical results - [ ] Output is raw markdown with no preamble