How to Use AI for Grant Writing in 2026: Prompts, Templates, and a 7-Step Process

Grant writing is one of the slowest, most repetitive forms of professional writing. A typical foundation proposal runs 10-25 pages with appendices. A federal proposal can hit 100. Grant writers across nonprofits, universities, healthcare systems, and small research labs spend 30-60 hours producing a single proposal — most of which is structuring evidence, summarizing prior work, and rewriting boilerplate that funders have seen a hundred times.

AI is reshaping this work, but in a narrow way. Tools like ChatGPT, Claude, and Perplexity cut drafting and editing time dramatically. They cannot replace the relationships, the sourced evidence, or the program-officer judgment that wins grants. Used well, they free up the hours grant writers should be spending on the parts of the job that actually move the needle.

This guide covers the exact 7-step workflow our team has tested across nonprofit foundation grants, federal research proposals, and SAMHSA program applications. It assumes you already know your program area and have a real project to fund. If you don't, AI can't help — and the proposal won't win regardless.

What AI Can and Cannot Do in a Grant Proposal

Before opening any tool, get clear on the boundaries. The grant writers who get burned by AI all share one mistake: they let it generate the substance, not just the language.

What AI does well: structuring sections to match RFP requirements, drafting prose from your bullet-point notes, summarizing literature you've already read, formatting logic models and tables, writing alternative versions of an executive summary, generating boilerplate organizational descriptions, and pressure-testing your evaluation plan by asking what could go wrong.

What AI does badly: inventing statistics, fabricating citations, generating "evidence-based practices" that don't exist in the literature, writing the project narrative when it has no information about your actual program, and producing the kind of community-specific detail that program officers use to separate real applicants from copy-pasted ones.

The rule that solves this: AI handles structure and language; the human owns evidence and judgment. Every statistic, citation, partner name, and methodology detail must trace to a source the writer can defend in a phone call with the program officer. If you can't defend it, cut it.

The 7-Step AI Grant Writing Workflow

This is the workflow our team uses for proposals between $50K and $5M. Smaller grants compress the steps; federal R01 and NSF proposals expand the evaluation and budget phases significantly.

Step 1: Build the Evidence Pack (Human only — 4-8 hours)

This is the step grant writers most often try to skip. Before opening ChatGPT or Claude, build a single Evidence Pack document that contains every fact your proposal will rely on. The Evidence Pack is for you — it never gets shown to the funder — but it becomes the source material for every AI prompt that follows.

The Evidence Pack should contain:

• Problem statistics: 5-10 sourced statistics about the problem you're addressing, each with a working URL, the year of the data, and the geographic specificity (national vs. your service area).
• Literature on the proposed approach: 6-12 peer-reviewed citations supporting your intervention, with author, year, journal, and a one-sentence summary of what each source actually demonstrates.
• Organizational track record: 3-5 specific outcomes from prior programs — number served, retention rate, measurable result — with the year and the funding source that paid for the work.
• Partner commitments: name and role of each partner organization, what they're contributing in cash and in-kind, and whether you have a Letter of Support, MOU, or just a verbal agreement.
• Budget framework: total request amount, salary lines with FTE percentages, indirect rate, contractor lines, equipment, and travel — the actual numbers, not estimates.
• Logic model bones: the inputs, activities, outputs, short-term outcomes, and long-term outcomes you'll claim. Bullet form is fine.

Aim for 5-10 pages of dense, sourced content. If you can't fill out the Evidence Pack, you're not ready to write the proposal — and AI cannot rescue an underdeveloped project.

Step 2: Reverse-Engineer the RFP (Use Claude — 30 minutes)

Most grant writers read the RFP once and miss half the requirements. Use Claude to extract a structured checklist before drafting anything.

Use this prompt:

Below is the full RFP for [funder name and program]. Extract every distinct requirement into a numbered checklist organized by: (1) eligibility requirements, (2) required proposal sections with word or page limits, (3) required attachments and supporting documents, (4) evaluation criteria with relative weights, (5) deadline and submission format, and (6) any explicit "do not do" instructions. For each item, quote the exact language from the RFP. If anything is ambiguous, flag it with [CLARIFY].

The output is the source of truth for the rest of the proposal. Print it. Tape it to your monitor. Every section you draft must satisfy a specific item on the list — and if a paragraph doesn't, cut it.

Step 3: Statement of Need (Use Claude + Perplexity — 2 hours)

The Statement of Need is the section grant writers most often pad with fluff and where AI most often invents statistics. Use Perplexity to gather sourced data, then Claude to turn the data into prose.

Step 3a — Gather sourced inputs in Perplexity:

For [problem area, e.g., "youth opioid use disorder in rural Appalachia"], find 2024-2026 published statistics on: prevalence rates by demographic, geographic distribution, treatment access gaps, and economic impact. Cite every claim with a working URL and the publishing organization. Prioritize CDC, SAMHSA, peer-reviewed journals, and state health department data over advocacy-organization estimates. Where data is unavailable, say so explicitly.

Step 3b — Convert to proposal prose in Claude:

Below is my Evidence Pack and Perplexity research with sources. Write the Statement of Need section, [word limit] words, structured as: (1) the problem nationally with one or two anchor statistics, (2) the problem in our specific service area with sourced local data, (3) the gaps in current responses and why our population remains underserved, and (4) the consequences of inaction in human and economic terms. Use only the facts and sources I've provided. Where I have a community quote or testimonial in the Evidence Pack, weave one in to ground the section.

The community-quote rule is what separates a real Statement of Need from a generic one. AI-generated Statements of Need without specific local voice all sound the same. A single quote from a real client, partner, or community member transforms the credibility of the section.

Step 4: Project Description and Logic Model (Use Claude + ChatGPT — 3 hours)

The Project Description is mostly translation work — converting your Evidence Pack and logic model bones into structured prose. Run Claude for the narrative, then ChatGPT to format the logic model as a clean table.

Claude prompt for the narrative:

Using my Evidence Pack and the RFP requirements checklist, draft the Project Description, [word limit] words. Cover: (1) project goal and measurable objectives in SMART format, (2) target population with eligibility criteria, (3) the intervention in concrete operational detail — who does what, when, and how often, (4) how the intervention is grounded in the evidence-based practices cited in my literature review, and (5) the implementation timeline organized by quarter. Use only the activities, partners, and citations from my Evidence Pack. Do not invent staff roles, partner contributions, or evidence-based-practice names.

ChatGPT prompt for the logic model:

Convert the following logic model bones into a formatted Logic Model table with columns for Inputs, Activities, Outputs, Short-Term Outcomes (0-12 months), Intermediate Outcomes (12-24 months), and Long-Term Outcomes (24+ months). Use SMART criteria for every outcome. Output as a markdown table I can paste into Word or Google Docs.

The "no invented partner contributions" rule is critical. Multiple program officers have told us they've seen proposals list partners who later said they were never asked to be involved. This is one phone call away from being caught — and it ends future eligibility with that funder.

Step 5: Evaluation Plan (Use Claude — 2 hours)

The Evaluation Plan is where federal proposals are won or lost. It is also the section AI handles surprisingly well, because the structure is highly conventional.

Using my logic model and the RFP's evaluation criteria, draft the Evaluation Plan, [word limit] words. Cover: (1) evaluation questions tied to each project objective, (2) data collection methods with specific instruments where applicable (validated scales, administrative data, qualitative interviews), (3) data collection schedule and responsible staff, (4) data analysis approach for both process and outcome measures, (5) how findings will be used for continuous quality improvement, and (6) any planned dissemination. If the RFP requires an external evaluator or specific evaluation framework (CDC Framework, RE-AIM, etc.), structure the section accordingly.

Then pressure-test it:

Now critique the evaluation plan you just drafted. What are the three weakest measurement choices? Where is the data collection burden likely too high? Which outcomes will be hardest to attribute to the intervention versus to outside factors? Be ruthless — assume a federal program officer who reviews 80 proposals a year is reading this.

The critique output is what improves the section. Most evaluation plans get rejected for measurement-burden problems and weak attribution logic — both of which AI flags accurately when prompted.

Step 6: Budget Justification and Sustainability (Use ChatGPT — 2 hours)

Switch to ChatGPT for the budget. Its data analysis tools handle the salary, fringe, and indirect calculations more cleanly than Claude.

Build a 3-year budget justification for the following request: total ask [$X], distributed across personnel ([list FTE positions and percentages]), fringe at [X%], travel ([list trips and amounts]), supplies ([categories]), contractual ([list contractors and amounts]), and indirect at [X%]. For each line, write 2-4 sentences justifying the cost in plain language tied to specific project activities. Output as a structured budget narrative I can paste into the proposal.

For sustainability, use Claude:

Draft a Sustainability section, [word limit] words. Address: (1) plans to embed proven activities into existing organizational operations after the grant, (2) potential additional funding sources we will pursue with realistic probability assessments, (3) cost-reduction strategies as the program matures, and (4) commitments from leadership and board. Use only the facts in my Evidence Pack. If I have not provided concrete sustainability plans, write [SUSTAINABILITY GAP] in brackets — do not invent commitments.

The [SUSTAINABILITY GAP] flags are the most useful output here. They're a checklist of conversations you need to have before submission.

Step 7: The Mandatory Human Review (Use yourself — 3 hours)

Before submission, walk through the entire proposal manually with this checklist. AI produces drafts that look correct but contain landmines — your job is to catch them.

• Every statistic traces to a source. Pick five random statistics in the proposal and verify each one in your Evidence Pack or original source. If any number can't be traced, delete it or replace with a sourced figure.
• Every citation is real. Open Google Scholar and verify each citation by author, year, and journal. AI fabricates citations more often in grant proposals than in any other document type because the literature density is high.
• No invented partners or evidence-based practices. Re-read every named organization and intervention model. Each must be a real entity in the relationship the proposal describes.
• RFP checklist is satisfied. Walk through the Step 2 checklist line by line. Every requirement gets checked off, every word limit is honored, every required attachment is included.
• Logic model is internally consistent. Each output produces an outcome. Each outcome traces to an evaluation method. Each method has a responsible staff member with allocated FTE in the budget.
• Budget math reconciles. Salary plus fringe plus everything else equals the total request, and indirect is calculated on the right base. ChatGPT gets this right 95% of the time — but the 5% kills proposals.
• Plain-English read-aloud test. Read the Statement of Need and Project Description aloud. If a sentence sounds like AI corporate filler, rewrite it. Program officers can hear the difference.

Common AI Grant Writing Mistakes

Mistake 1: Asking AI for the literature review

"Find me 10 evidence-based practices for adolescent substance use prevention" is the worst possible prompt. ChatGPT and Claude will produce a list that looks authoritative but mixes real interventions, paraphrased descriptions, and outright fabrications. Use Perplexity for sourced literature with working URLs, or — better — search PubMed and SAMHSA's Evidence-Based Practices Resource Center yourself.

Mistake 2: Letting AI invent the local need

If you ask AI "describe the unmet need for [program] in [my county]," it will produce a confident, specific-sounding paragraph that is essentially fiction. Local data must come from your county health rankings, state health department, or community needs assessments. AI's training data does not contain reliable county-level statistics for most programs.

Mistake 3: Generic boilerplate organizational description

AI happily writes a 400-word organizational description that could apply to any nonprofit. Program officers reading their 50th proposal of the week skim it in 8 seconds and forget it. A strong organizational description has at least three specific, sourced outcomes from prior programs — numbers, dates, and a sentence on what was learned. Write that yourself, then have AI tighten the prose.

Mistake 4: Skipping the RFP reverse-engineering

Grant writers who jump straight to drafting and try to align with the RFP at the end produce proposals that miss requirements. The Step 2 checklist takes 30 minutes and prevents the most common rejection reason: ineligibility on a technicality.

Tools That Pair Well With This Workflow

A few interactive tools accelerate this work:

• Claude vs ChatGPT comparison — pick the right primary tool before you start.
• Perplexity vs ChatGPT — covers when to use each for sourced research.
• AI Tools Comparison Builder — broader comparison if you want a third tool in the stack.
• AI Business Plan Workflow — overlapping prompts for the budget and sustainability sections.
• AI for Standard Operating Procedures — useful for the implementation-detail sections of large federal proposals.
• Claude for Financial Analysis — extended prompt patterns for the budget justification.

What to Do With the Time You Save

A grant proposal is not the actual deliverable — it is a thinking artifact that opens the door to a relationship with a funder. What program officers fund is the conversation it enables: a 30-minute call, a site visit, a track record they can vouch for to their board.

Use the AI workflow above to compress the writing time. Use the time you saved to do the parts AI cannot do: meet with three potential partners, attend the funder's pre-application webinar, call a prior grantee for advice, and tighten your own organizational evidence pack so the next proposal is even faster. The proposals that get funded are the ones where the writing reflects relationships and rigor the writer actually built.

For more on positioning AI work in your career, see our guide to talking about AI experience and our best AI certifications roundup. If you're a grant writer evaluating which AI skills to add to your resume, the AI Skills Checker can identify gaps before your next role transition.