AI Grant Writing: Rewrite Last Year's Grant to Funder Word Limits

AI grant writing works best as a rewrite, not a blank page: give the model last year's application, the new funder's guidelines, and their funder questions with each word limit, all together. Then work in four passes. First, map every question to the old text and list what nothing answers. Second, rewrite each answer inside its word or character count, using the funder's own terms. Third, check the required attachments against what you have. Fourth, keep watching the funders you follow for new calls. A person still owns the voice, every number and the submission. Some funders restrict AI-written proposals, so read their rules first.

Most advice on this topic stops at "build a boilerplate bank and paste sections into ChatGPT, one question at a time." That works for one answer. It falls apart on the real job, which is below.

What is the real job when you reuse a grant application?#

You have one strong application, usually the one that got funded. You have twelve funders on the list for this year. Each one asks for the same program in a different shape:

• One wants a 500-word statement of need. Another wants "the problem you address" in 1,500 characters, spaces included.
• One says "participants", another "beneficiaries", a federal call "the target population".
• One wants a logic model, one wants three SMART objectives, one wants neither.
• Each has its own attachment list: the IRS determination letter, the latest 990, an audit or reviewed financials, a board list, a budget in their template.

So the work is less writing than translation and bookkeeping. You match their questions to what you have already said well, find what you never said, cut to their counts, and make sure nothing required is missing on the day. AI handles that kind of job well, if you give it everything at once and not one paste at a time.

What do you need before you start?#

Put these in one folder per funder, and work on copies:

1. Your source application. The funded version, plus any reports you sent that funder. Final reports hold your best outcome numbers.
2. The new funder's guidelines or RFP, as a PDF or saved page. Include the FAQ and the scoring rubric if they publish one. The rubric tells you what gets points.
3. The question list with limits. Portals often show questions only after you log in. Copy each question with its limit and its unit, word or character, into a plain text file.
4. Your attachments folder. Determination letter, 990s, audit, budget, board list, key staff bios, letters of support, with dates in the file names.
5. A facts file. One page with the numbers you are allowed to use: people served, outcomes, budget, staff, and the source and year of each. This is what stops the model inventing figures (more on that below).

Pass 1: Which funder questions does the old narrative already answer?#

Don't start by rewriting. Start with a map, because it shows the gaps before you spend an hour polishing answers.

Read guidelines.pdf and questions.txt for [Funder], and our application
source-application.docx. Make a table with one row per funder question:
question number, question text, limit and unit, the section(s) of our
application that answer it (quote the first sentence of each), and a
coverage rating: full / partial / none. Use the funder's own words for the
question. After the table, list every question rated partial or none, and
say exactly what information is missing. Do not write any answers yet.

The "none" rows are the valuable part. Common ones:

• Evaluation with specific measures, when your old funder only asked for outcomes.
• Sustainability after the grant ends.
• Equity or community voice: how the people served shaped the program.
• Organisational capacity: who runs it and what else they manage.
• Their priority area by name, for example "rural access" or "workforce".

For every gap, get the facts from a person before any drafting. A model asked to fill a gap on its own will write a confident paragraph about an evaluation plan you do not have.

Pass 2: How do you rewrite to each funder question's word limit?#

Now draft, one question per row of the map, and keep three rules.

Say the unit, and aim under it. "500 words" and "3,000 characters including spaces" are very different lengths. Ask for about 90% of the limit, so edits have room.

Count with code, not by eye. Language models count their own words badly. They read text in tokens, not words, and a draft that "feels" like 480 words may be 560. If your tool can run code, ask for the count to be computed. Claude Code, Codex and ChatGPT's data analysis can all do this. If it cannot, paste the draft into a word counter. Always do a final paste into the portal's own box, because some portals count line breaks and some trim spaces.

Use the funder's terminology. Reviewers score against their own language. If the RFP says "families with low incomes", do not answer about "underserved households".

Using the coverage table, draft the answer to question 4.
Limit: 1,500 characters including spaces. Aim for 1,350.
Use only facts from source-application.docx and facts.md. Use this
funder's terms: [list from guidelines]. Keep our voice: plain, first
person plural, no adjectives we did not use ourselves. After the draft,
count the characters with code and print the count. Mark any sentence
that states a number with [check] and name the fact it came from.

The [check] tags make review quick. Every number in the draft should trace back to a line in facts.md. A number that traces to nothing is a number the model made up.

When everything is drafted, run one more pass across the whole application. Programs described twice in different words, or a budget figure that changed between answers, are what reviewers notice.

Pass 3: What attachments does the funder require, and which do you have?#

This pass takes minutes and prevents the worst outcome, an application rejected before anyone reads it.

From guidelines.pdf, list every attachment or document the funder
requires or accepts, with any rules (format, page limit, date range,
template, signature). Compare against the files in /attachments.
Make a checklist: required item, rule, the file we have (or "missing"),
and whether our file meets the rule (e.g. "990 is FY2023; they ask for
the most recent filed year").

Look at the date rules carefully. "Most recent audited financials", "board list for the current year" and "budget on our template" are where an old file quietly fails.

Pass 4: How do you watch the funders you follow for new calls?#

The last pass is the one that pays for the other three. You keep a short list of funders you fit: their grants pages, their newsletters, a federal search on Grants.gov for your categories. Once a week, something needs to check them.

Every Monday at 8am: open each page in funders.csv. For each, list any
call that opened or changed since last week: name, deadline, amount,
eligibility, and a fit note against our facts.md (fit / maybe / no, and
why in one line). Add new rows to calls.csv. Do not start applications.

"Fit notes" matter more than alerts. Twenty alerts a week is noise. Three rows that each say why you might qualify is a pipeline. Scheduling a recurring AI job that reads local files covers how this works in Claude, ChatGPT and a desktop agent, and where each one stops. Funder portals behind a login are a separate problem, covered in letting an agent work on sites that need your login.

How do you keep your organisation's voice?#

Reviewers read a lot of proposals, and AI-smoothed prose has tells: every paragraph balanced, every sentence the same length, "holistic" and "transformative" throughout. Three habits keep your voice:

• Rewrite from your own text, not from a blank page. Pass 2 only rearranges and cuts sentences you already wrote. That is why reuse beats generation.
• Give it a voice sample and a banned-word list. Paste two paragraphs you like and name the words you never use.
• Read it aloud before it goes in. If a sentence is not something your executive director would say to a program officer, cut it.

What the model must never supply:

What do funders say about AI-assisted proposals?#

Policies differ, and they change. As of September 2026, the major public ones say:

• NIH (notice NOT-OD-25-132): "NIH will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be original ideas of applicants." If AI use is identified after an award, NIH may refer the matter to the Office of Research Integrity. It applies to receipt dates from September 25, 2025, together with a cap of six applications per principal investigator per calendar year. NIH's peer reviewers are separately barred from putting applications into generative AI tools (NOT-OD-23-149).
• NSF (December 2023): proposers "are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used", and are "responsible for the accuracy and authenticity" of what they submit. Reviewers may not upload proposals to non-approved AI tools.
• UKRI: since a December 2024 update, applicants are "expected to be transparent where they have used generative AI tools in the development of an application."
• Private and community foundations: most have no public policy. Some ask for disclosure in the application form. Read the guidelines and FAQ for each call, and if they are silent, a one-line email to the program officer is cheaper than guessing.

The workflow above sits in the defensible part of this range. Your facts, your prior writing and your decisions go in, and the model reorganises and cuts. Under NIH's rule, though, even a heavily AI-rewritten section may count as "substantially developed by AI". For federal research grants, use AI for the map and the checklists, and write the science yourself.

Should you buy a grant writing tool or use your Claude or ChatGPT plan?#

Dedicated tools bundle three things: a funder database, a library of your past content, and an AI drafting screen. General plans give you the model and nothing grant-specific. Prices as of September 2026, from each vendor's own pricing page:

How to choose:

• Buy a database tool if finding new funders is the bottleneck. A general model cannot search an index it does not have. Instrumentl or Grantboost can surface funders a web search misses.
• Buy a drafting tool if several people write and you want a shared content library with no setup.
• Use the plan you already pay for if you know your funders and the work is rewriting and checking. That covers most small shops applying to the same thirty foundations each year.

Not every tool says which model writes the draft. Ask the vendor. If the answer is Claude or GPT, the extra you pay buys the grant-specific wrapper: the database, the library, the workflow. That can be worth it. Just know that is what you are buying. For what your own subscription can be used with, see using your Claude or ChatGPT subscription in AI agent apps.

Where Universe fits#

Universe is not a grants database, and it has no funder index or grant-specific scoring. It is a Mac app that runs Claude Code or Codex on your computer, signed into the Claude or ChatGPT plan you already have. It does the four passes above as one job, not as a stack of chat pastes. The asks on the grant writing page are exactly these:

• "Rewrite our last application to this funder's questions, inside their word counts."
• "Which of their questions does our draft not answer?"
• "List every attachment they ask for and which ones we already have."
• "Every Monday, check the funders I follow for new calls and tell me which ones we fit."

What that looks like in practice:

• It works in the folders you give it. The coverage map, the drafts and the attachment checklist come back as files, a document and a datasheet you can sort, beside the PDFs they came from.
• It can count with code. Claude Code and Codex can run commands, so the limit check can be a computed number, not a guess.
• It uses a real Chrome signed in as you, copied from your own profile, so it can read funder pages and portals you log into. Sending, paying and posting wait for you. You submit.
• Schedules run the Monday scan. They run on a Mac, so that Mac has to be on.
• Version history keeps earlier versions of each draft on the Mac, so the cut that went too far can be restored.
• Skills can hold your voice rules and your banned-word list once, for every application.

On data: the work happens on your Mac and the model runs on your own account, so the text goes to Anthropic or OpenAI under your plan's terms. When you are signed in, Universe syncs files and session transcripts to its cloud, because sharing and a second Mac need them. The app also works without an account.

Universe is free to start, and Pro is $19 a month, on top of your model plan (pricing). For the organisation's side of the work, finding calls, keeping standing documents current and reporting on time, see Universe for nonprofits. The full feature list is at /features. To try the four passes on an application you have already submitted, download it for Mac (Apple silicon or Intel, macOS 13 or later).