grant-review

# Grant Proposal Review Routes a grant proposal to an **independent, fresh-context reviewer** that scores it against the official NIH or NSF criteria and returns a structured review. This skill is a thin dispatcher: it decides how to run the reviewer and in which mode. The review procedure, criteria, scoring rubrics, and output templates all live in `references/` and are loaded by the reviewer, not duplicated here. ## When to use Activate when the user wants feedback on a grant proposal (specific aims, research strategy, project description) evaluated against NIH or NSF review criteria. ## Why a fresh-context reviewer Review validity depends on independence: a reviewer that shares the conversation that produced the proposal is biased toward it. Run the reviewer in a separate context and pass only **framing** (proposal path, mechanism/agency, resubmission status, target program), never the authoring rationale. This is also why the reviewer is a subagent on tools that support one, and an inline procedure where they do not. ## Modes (user decides each run) - **Single (default):** one independent reviewer applies the full procedure end to end. - **Panel (opt-in):** spawn N independent reviewers in parallel, then a synthesis pass. Trigger on "mock study section", "panel review", or an explicit request for multiple reviewers. NIH: 3 reviewers; NSF: 2-3 panelists. Each reviewer scores independently from `references/`; a final chair pass reconciles them into one output. Surface genuine disagreement rather than blindly averaging scores. - **Readability (run BEFORE the panel):** one fresh-context reviewer restricted to clarity, with no merit scoring. Follow `references/readability-procedure.md`; a panel can run clean while the first page still contains prose the PI cannot follow on first read. The default pass reports line-anchored, length-neutral rewrites for long chained sentences, ambiguous pronouns, premature terms or abbreviations, rhetorical inversions, and gratuitous self-undermining while preserving substantive limitations. If requested, the same reviewer can add the optional conciseness hunt as a second pass; do not fan out a second agent unless the caller asks for it. Conciseness rewrites must be strictly shorter while preserving every number, citation, qualification, and required element. ## Dispatch Pick the branch for the current tool. Single and panel reviews follow `references/review-procedure.md`; readability reviews follow `references/readability-procedure.md` and never score merit. - **Claude Code:** `Agent(subagent_type: "grant:grant-review", ...)` passing the proposal path, mechanism, and mode. The agent selects the readability procedure when `mode=readability`; for panel mode, issue one `Agent` call per reviewer role in a single message so they run in parallel, then a final synthesis `Agent` call. - **Codex CLI:** plugin installation exposes this skill, not a Codex subagent. To use a fresh-context Codex reviewer, first copy `agents/templates/grant-review.toml` to `~/.codex/agents/` or `.codex/agents/`, then invoke that configured agent if the current Codex surface supports `/agent`. Pass `mode=readability` for the non-scoring pass. For panel mode, ensure `max_threads` covers the reviewer count. If no Codex subagent is configured or available, use the fallback branch. - **Copilot CLI:** plugin installation exposes this skill and, through `.github/plugin/plugin.json`, the `.agent.md` reviewer in `agents/templates/`. Invoke that configured agent when the current Copilot surface supports custom agents; pass `mode=readability` for the non-scoring pass and use `/fleet` for panel mode when available. If running outside a plugin install, copy `agents/templates/grant-review.agent.md` to `.github/agents/` or `~/.copilot/agents/`. If no custom agent is available, use the fallback branch. - **Fallback** (no subagent support, or the user wants an interactive in-thread review): first locate the rubric (`$CLAUDE_PLUGIN_ROOT/skills/grant-review/references`, else `find . -type d -path '*/skills/grant-review/references' | head -1`); if it cannot be found, stop and tell the user to install the grant plugin rather than reviewing from memory. Then follow `references/readability-procedure.md` for `mode=readability`, or `references/review-procedure.md` for other modes. ## The brain (do not duplicate into dispatch or agent shells) - `references/review-procedure.md` -- step-by-step procedure: mechanism ID, ingest, score, synthesize, output. - `references/readability-procedure.md` -- non-scoring first-read clarity pass with anchored, page-neutral rewrites and an optional concise-writing hunt. - `references/nih-review-criteria.md`, `references/nih-career-training-criteria.md`, `references/nsf-review-criteria.md` -- criteria and scoring. - `references/sbir-sttr-review-criteria.md` -- SBIR/STTR (R41, R42, R43, R44) criteria. These are **not** reviewed under the RPG Simplified Review Framework; they keep the five classic scored criteria plus commercial-potential questions, Commercialization Plan assessment, Phase I milestones, Phase I progress, and Fast-Track acceptability. - `references/review-best-practices.md` -- calibration and common reviewer comments. - `references/review-output-templates.md` -- NIH and NSF output format. - `examples/sample-nih-r01-review.md` -- worked review; `examples/sample-r01-aims.md` -- sample proposal input for testing. - Sister skill `manuscript:humanizer` (invoke through the host's skill-invocation mechanism) -- AI-writing patterns to flag in grant prose.