Stop starting every grant narrative from a blank page.
Grant staff re-type the same mission statement, program description, and outcomes data into a new narrative every time a notice comes in — and reviewers have no way to tell which sentences are backed by a real source and which an AI tool invented. Draft Generator pulls directly from your organization’s own knowledge base and its narratives proven by real award outcomes, and hands you an editable first draft you finish and save — not a document that leaves the building on its own.
The actual Draft Generator
Not a mockup — this is the live 4-step Draft Generator wizard, captured from a running instance of the application on September 8, 2026.
Watch a draft get built
Each step below mirrors a real stage in the generation code (src/lib/drafts/generator.ts).
How it actually works
Draft Generator is a real 4-step wizard — Select Opportunity, Customize, Generate, Review & Export — with genuine step state, not a cosmetic progress bar. You can jump between steps freely, and nothing you’ve entered is lost moving back and forth.
Generation itself is a multi-source retrieval pipeline, not a single prompt to an AI model. Before drafting, the code pulls your organization’s profile (mission, service area, population served), your knowledge-base entries scoped to whichever categories that template actually needs, and — where your service area supports it — real Census, HUD, BLS, and CDC need-statement data for your county or state, with an explicit instruction to the model not to fabricate a statistic beyond what was actually retrieved.
Narratives proven by real award outcomes get a separate, prioritized query: up to five of your organization’s own narratives that funded a past award in this opportunity’s funder category, ranked by a real win/loss effectiveness score and injected as high-weight examples — not a generic “proven” label with nothing behind it.
Every knowledge-base entry, proven narrative, and outside data point actually supplied to the model is tracked and shown back to you in a source list underneath the draft — the same list the model was given, not a reconstruction after the fact.
Inputs and outputs
- One saved, scored opportunity from your pipeline
- One of six templates: grant narrative, donation request letter, budget narrative, impact statement, letter of inquiry, full proposal
- Your organization profile and knowledge-base entries, scoped to the template
- Up to 5 of your own narratives proven by real award outcomes in this funder category
- Real Census/HUD/BLS/CDC need data for your service area, when available
- An editable draft with a real confidence score based on how much source material backed it
- A source list showing exactly which knowledge-base entries and proven narratives were used
- Optional humanized rewrite and rubric-based rescoring, on request
- A saved draft version and a normal pipeline entry once you click Save
Where a person is in the loop
Writing a draft and submitting a draft are two separately-gated actions in the code, not two names for the same button. In the wizard, the only forward action a draft can take is Save, which writes the text onto a normal pipeline item. Nothing on this page calls AutoApply or any portal-submission endpoint.
Drafts can also be generated with no person writing them — an agent can produce a full draft on its own from a queued opportunity. That path is structurally blocked from submitting: the code that creates those drafts fails outright unless it marks the row pending review, and it never sets a submission timestamp. The review screen built for those drafts can only dismiss one from the queue, never submit it.
The one real path from a draft into AutoApply is the Draft Queue, and it always requires a person to click Approve or Submit. An organization can optionally set a confidence threshold so an approved draft above that threshold skips a second, separate Submit click — but Approve itself is always a person clicking a button, never something the system does on its own.
Quantified results
We don’t have a published draft-acceptance rate, time-saved-per-draft, or win-rate-lift number for Draft Generator yet. Those figures depend on real usage over a full grant cycle, and we’d rather show you an honest “pending” than a number we backed into. This section will be replaced with real pilot data once we have it.
What it actually connects to
Both the initial draft and the optional “Humanize” rewrite are generated by the same underlying model.
- Claude (Anthropic) — drafting and rewriting
Pulled for your service area’s county or state, when a draft’s template calls for a need statement.
- Census, HUD, BLS, CDC — aggregated county/state need-statement figures
Not yet connected: the Email Draft button in Review & Export is visible but disabled pending a Gmail connection — we’re disclosing that here rather than implying it already works.
Security and reliability
Every knowledge-base and proven-narrative query is scoped to your organization, every generated draft is saved as a version you can trace, and the source list shown under a draft is the same list the model was actually given — not a reconstruction after the fact. For the platform-wide security and governance model — including how the AutoApply approval gate works downstream of a saved draft — see the Trust and Governance page.
Customer story
COMING SOON
We don’t have a published customer story for Draft Generator yet. Once an organization has used it through a real grant cycle and agreed to be named, their story will go here — not a composite or hypothetical example.
FAQ
Run the free Funding Potential Scan first — a short, no-signup check of the kind of funding your organization is likely to qualify for.
Analyze a Grant Narrative
Connect your knowledge base and let Draft Generator turn a scored opportunity into an editable first draft — sourced from your own facts and narratives proven by real award outcomes.