Red-teaming a grant proposal before reviewers see it
Grant writing is one of the most collaborative faculty AI uses — 70% augmentation rather than automation in Anthropic's data — and a Wiley survey of nearly 5,000 researchers found most expect AI in grant writing to become widely accepted. You cannot use AI to review other people's proposals, but you can make it the harshest reader your own draft meets before submission.
You are a skeptical, experienced grant reviewer for {{funder}} — the kind who reads fifteen proposals in a weekend and rewards clarity over ambition. Review the draft section below against the funder's published criteria. Draft section: {{draft_text}} Review criteria, pasted from the funder's official guidance: {{review_criteria}} My biggest worry about this proposal: {{main_concern}} Output: 1. A reviewer-style summary: three genuine strengths, then weaknesses ranked by how much each would hurt the score 2. Line-level flags: quote every sentence where a claim outruns the evidence presented, where aims appear dependent on each other's success, or where jargon would slow a panelist outside my subfield 3. The five questions a reviewer would most likely raise in panel discussion 4. Suggested rewrites for the three weakest passages only, keeping my voice and technical content Constraints: do not invent citations, preliminary data, methods, or results — where support is missing, flag the gap rather than filling it. Judge only against the criteria I pasted, not criteria you recall from memory. Do not soften criticism to be encouraging; I need the panel's objections now, not in the summary statement.
Fill in your details and the prompt updates live — then copy.
**Weakness 1 (highest impact on score):** The Overview claims the framework will "generalize to any coastal aquifer," but the validation plan covers two sites with similar hydrogeology. A panelist will ask what happens in karst systems — either narrow the claim or justify the extrapolation. **Line flag:** "Preliminary results confirm the approach is feasible" — no preliminary data appear in this section. State what exists or cut the sentence. **Panel question 3:** If the Aim 1 model underperforms, is Aim 3's forecasting work still fundable on its own?
The full workflow
- Paste the actual solicitation criteria from the funder's site — never let the model work from remembered criteria.
- Run the red-team on one section at a time; whole proposals blur the feedback.
- Fix the flagged gaps with real data and citations from your own library, not the model's suggestions.
- Check your funder's AI policy before submission — NSF encourages disclosing AI use in the project description and holds you responsible for accuracy.
Watch out for
Never paste a proposal or manuscript you are REVIEWING into any AI tool: NIH prohibits generative AI in peer review outright, and NSF treats uploading proposal content to non-approved tools as a confidentiality violation.
Your own unfunded ideas are competitive intellectual property. Prefer an institution-licensed AI tool with a no-training agreement over a free consumer account for proposal text.
Any citation or preliminary claim the model adds is a fabrication risk — and fabrication in a federal proposal is research misconduct, not a typo.
Where this comes from
Every use case on this site is grounded in real reports from working professors — not invented by us.