AI grant proposal writing: what helps and what hurts

What does AI-generated text look like to a grant reviewer?

To a reviewer working through a queue, generic AI text has a recognizable shape. These are the tells:

• Openers that say nothing. Restating the question, or praising the ecosystem before saying what you will build.
• Adjectives instead of numbers. "Significant adoption" where "14 integrators since June" should be.
• Uniform structure. Every section the same length, every list three bullets long.
• Stock phrases. "Innovative solution", "empower users", "foster collaboration", "comprehensive approach".
• Claims nobody can check. "Thousands of users" with no link, address or dashboard.
• Wrong program facts. Old grant sizes, closed programs described as open, requirements the program never had.
• Text that fits any chain. The Solana Foundation asks you to explain why your project needs Solana rather than another chain; a generic answer fails that test on sight.

Volume makes this worse for you. Arbitrum's D.A.O. Grant Program received 1,262 applications in Season 3, which closed on March 17, 2026 with no Season 4 proposed. The Stacks Endowment drew 325 applications in its first cycle; its Q4 2026 dates had not been announced as of October 2026. A reviewer facing that queue skims the proposals that sound alike and stops at the first concrete number.

What is AI good for in a grant proposal?

AI is good at the mechanical half of proposal writing: shaping, trimming and checking text you supply.

• Turning criteria into an outline. Paste the program's review criteria and get one heading per criterion. Ink Spark, for example, reviews on Ink alignment, live utility, traction, priority fit, use of funds and builder credibility; Cardano's 2026 Catalyst pilot, which closed in August with no next round announced, scored business questions at 70% and technical questions at 30%.
• Cutting. Shortening a section to a word limit without losing facts.
• Finding gaps. Asking which criterion a draft does not answer, or which claims have no source.
• Consistency checks. Do the budget lines add up? Do milestone dates match the timeline? Does the total match the request?
• Rehearsing questions. Generating the questions a reviewer might ask in a call; see questions grant reviewers ask.
• Language. Polishing English if it is not your first language. Some programs require English, such as the Ethereum Foundation's PhD Fellowship (no 2027 call has been announced), while the CKB Community Fund DAO accepts English or Chinese.

What can AI not supply?

AI cannot supply anything a reviewer will verify, and in Web3 grants that is much of what matters.

• Evidence. Contract addresses, dashboards, downloads, users. Ink Spark requires that claims about users, volume or TVL be backed by data reviewers can verify. Prezenti Grants asks for evidence for every usage claim; as of October 29, 2026 its site says both that Season 3 is open until December 29 and that it has concluded, so check with Prezenti first.
• Track record and references. The OpenSats General Fund needs at least two written references for general grants, and Maelstrom at least one.
• Real costs. Hours, rates and line items come from your plans, not from a model's sense of what sounds reasonable. See the Web3 grant budget guide.
• Relationships. Walrus ecosystem grants go to builders already active in the ecosystem, ideally recommended by someone on the team.
• Current program facts. Check every program detail on the program's own site; each program page in our directory shows when we last verified it.
• Commitments. Milestones in a proposal become terms in a grant agreement. Only you can promise them.

The stale-facts problem is real. ESP, the Ethereum Foundation's Ecosystem Support Program, stopped open proposals in November 2025 and had no Wishlist or RFP items open as of October 1, 2026. Gitcoin Grants has not run a round since GG24 in October 2025 and is folding quadratic funding into a local-money pilot, with no builder application yet. A model trained earlier may describe both as open.

A workflow that keeps your voice

The order matters: facts first, AI second, you last.

1. Write a fact sheet. Links to code, contracts, metrics, team profiles, past grants, costs. Nothing goes in the proposal that is not on this sheet.
2. Write a rough draft yourself. Ugly is fine. It sets the voice and the claims.
3. Ask AI to restructure it to the program's criteria and length limits.
4. Ask AI to critique it as a skeptical reviewer.
5. Rewrite in your own words, and delete any sentence you cannot back with something on the fact sheet.
6. Check every program fact against the official page.
7. Read it aloud. Anything you would not say to a reviewer on a call comes out.

Start from a structure built for grants rather than a blank chat: our Web3 grant proposal template and how to write a Web3 grant proposal give one.

Prompts that work

Good prompts ask the model to organize, check or cut, and tell it not to invent. These work with any capable model:

1. Here are the program's eligibility rules and review criteria: [paste].
   Turn them into a proposal outline, one heading per criterion.
   Do not write any content.

2. Here is my draft: [paste]. List every claim that has no number,
   date, name or link, and say what evidence would support it.

3. Act as a grant reviewer who has read 50 applications this week.
   Name the three weakest points of this proposal and the questions
   you would ask on a call.

4. Check this budget: do the line items add up to each milestone,
   and do the milestones add up to the total request? Show the math.

5. Cut this section to 150 words. Keep every number and name.
   Do not add anything.

6. Rewrite this paragraph in plain English for a non-native reader.
   Keep every fact. Do not add new claims.

Prompt 2 is the most useful of the six: it points the model at the vague claims in your draft instead of asking it to produce new ones.

Disclosure and program rules

Follow the program's rules on AI when it has them, and answer honestly when a form asks. The Compound Grants Program is the clearest case in our directory: its rules say low-quality or LLM-generated proposals are rejected, and its Questbook page is where proposals are filed. Note that its main Dapps and Ideas domain has been paused since April 14, 2026, and the other domains fund only exceptional teams.

If a form asks whether you used AI, say how: "used for editing and translation; all content and data are ours" is a complete answer. Also keep confidential material out of tools whose data terms you have not read: unreleased code, partner names under NDA, identity documents.

FAQ

Do grant reviewers reject AI-generated proposals?

Some grant reviewers reject AI-generated proposals outright: the Compound Grants Program's rules say low-quality or LLM-generated proposals are rejected. Most programs in our directory publish no AI rule, but generic AI text still tends to fail because it lacks the numbers, links and track record reviewers look for. A proposal where AI only edited your own specific draft does not have that problem.

Should you disclose that you used AI on a grant proposal?

Disclose AI use on a grant proposal when the program asks, and describe it accurately, for example "used for editing and translation". When a program does not ask, disclosure is your choice, but every fact in the proposal must still be yours and verifiable. Never let a model invent metrics, users, partners or references, since reviewers check them and grant agreements may rely on them.

Can AI write a grant budget?

AI can check a grant budget but should not write one. A model does not know your hourly rates, hosting costs or audit quotes, so any numbers it produces are guesses. Build the budget yourself from line items, then ask AI to verify that the lines add up, that milestone amounts match the milestone table, and that the total equals the amount requested.

Is it acceptable to use AI to translate a grant proposal?

Using AI to translate or polish a grant proposal is generally acceptable and can help non-native English speakers meet language requirements, such as the English-only rule of the Ethereum Foundation's PhD Fellowship, whose next call has not been announced. Check the translation yourself or with a fluent reader, because a model can change the meaning of technical terms, numbers or commitments, and you will be held to the translated text.