Inspiration
The hardest part of writing a technical weekly update or design review isn’t drafting—it’s standing behind every claim. Drafts often sound polished but thin: strong outcome language, weak or missing evidence. After a few AI rewrites, outdated details and empty phrasing get mixed in.
Discussions around multi-agent review point to the same failure mode: when two agents check each other without a hard rule that objections need evidence, they tend to agree anyway. We narrowed that insight into a small product wedge—block unsupported technical claims before the draft is treated as done.
What it does
OpenFlow ClaimGate is built for technical weekly reports, design notes, and review drafts.
- Paste a draft
- Extractor pulls out claims and outcome statements
- Evidence labels each claim: supported / weak / unsupported
- Gate returns two outputs:
- a submission-ready version (unsupported outcome language removed or rewritten)
- a rejection list (what was blocked and why)
It is not a general writing assistant. It does one job: force key technical statements through an evidence gate.
How we built it
- A fixed three-agent pipeline (no dynamic routing)
- Claims normalized into structured JSON: claim / evidence_status / reason
- One hard Gate rule: unsupported outcome claims cannot enter the final draft
- UI kept to three panels: source draft, claim list, final draft + blocked items
- Model calls capped with token limits and timeouts so a full run is demo-friendly
Challenges we ran into
- “What counts as evidence” is subjective—so we reduced it to checkable labels instead of free-form judgment
- Models treat rhetoric as fact—Extractor is intentionally strict and over-includes rather than misses claims
- Scope creep toward a platform—every feature request was cut unless it served the evidence gate
Accomplishments that we're proud of
- A real writing pain point compressed into a demo that can finish in minutes
- Output that is not “a smoother essay,” but a gated draft with explicit rejection reasons
- Proof that a small agent set with hard constraints beats a larger unconstrained agent team for this task
What we learned
Multi-agent value is less about role count and more about enforceable handoff rules. For technical writing, one rule—no evidence, no claim—is more useful than a general-purpose agent stack.
What's next for OpenFlow ClaimGate
- Accept commit / PR links as evidence sources
- Turn “weak evidence” into a one-click “add evidence” interaction
- Ship a team weekly-report template while keeping the same evidence gate (still not a platform)
Built With
- css
- grok
- node.js
- react
- start
- tailwind
- tanstack
- typescript
- xai
- zustand
Log in or sign up for Devpost to join the conversation.