PlanMerge

Turn fragmented AI planning drafts into one decision-traceable team plan.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for PlanMerge

Inspiration

Teams increasingly ask ChatGPT, Claude, Gemini, and other AI tools to draft the same product plan from different roles. The result is not one better plan; it is a pile of plausible documents. When someone finally merges them by hand, the team loses which ideas were selected, which alternatives were rejected, where opinions conflicted, and which source supported each decision.

PlanMerge was built to make that decision process visible.

What it does

PlanMerge turns multiple AI-generated planning drafts into one structured, reviewable team plan. It:

  • normalizes drafts from multiple AI tools into canonical ideas;
  • merges them into section-level Decision Blocks;
  • shows the selected option, rationale, alternatives, conflicts, and exact source excerpts;
  • scores section coverage, source coverage, and decision traceability through a Quality Gate;
  • lets teammates vote, leave anonymous opinions, override AI choices, and preserve the change in a Decision Log;
  • exports the final plan as Markdown or a portable JSON workspace.

A one-click verified sample loads 13 role-based drafts and produces a 12-section plan with complete source coverage and an intentional MVP-scope conflict, so judges can test the core value immediately.

How we built it

PlanMerge is a Next.js 16 and React 19 TypeScript application deployed on Vercel. A two-stage Responses-style pipeline first normalizes every draft, then generates structured Decision Blocks. Server-side validation cross-checks enums, IDs, and citations, retries malformed output with a repair prompt, and falls back to an offline deterministic harness when no API key is available.

Shared workspaces use Neon PostgreSQL and Prisma. The product is local-first for fast testing, with shared links, anonymous voting and opinions, link expiry and revocation, rate limiting, quality regression cases, and Playwright end-to-end coverage.

Codex helped turn product principles into working code: defining reliability invariants, implementing API and UI flows, creating quality and E2E harnesses, and auditing failure paths across the merge pipeline.

OpenAI Build Week extension — in progress

PlanMerge existed before OpenAI Build Week. During the submission period, we are meaningfully extending it with Codex and GPT-5.6 through a new Decision Room workflow.

For each unresolved conflict, GPT-5.6 will generate the smallest set of clarifying questions and a structured trade-off analysis. Team answers will trigger a section-level re-merge, producing a before-and-after decision diff while preserving every source and human decision. The repository will clearly separate the pre-event baseline from dated Build Week commits and Codex session evidence.

This section will be updated to completed tense with measured results before final submission.

Challenges we ran into

The hardest problem was not generating fluent text. It was preserving trust when the model normalizes, merges, rejects, or rewrites ideas. We had to prevent source-free options, validate cross-references, treat draft text as untrusted input, recover from malformed structured output, and keep the product usable when AI or database credentials are unavailable.

A second challenge was turning analysis into a coherent team workflow. Conflict detection only matters if people can inspect evidence, vote, override a choice, and export the reasoning alongside the polished document.

Accomplishments that we're proud of

  • Every selected option remains connected to its source idea and original excerpt.
  • The verified sample reaches 12/12 document sections, 13/13 source coverage, and 13/13 option traceability.
  • Quality gates can block approval and sharing when evidence is incomplete.
  • Human overrides and approvals survive export and sharing as decision records.
  • Judges can explore the main workflow immediately without providing credentials.

What we learned

Reliable AI collaboration requires more than a good prompt. It needs explicit schemas, provenance, repair paths, deterministic tests, human checkpoints, and a product interface that exposes uncertainty instead of hiding it.

We also learned that disagreement is valuable data. A strong planning tool should not silently average conflicting proposals; it should help a team understand and resolve them.

What's next

Complete the GPT-5.6 Decision Room, publish Build Week before-and-after evidence, expand templates beyond Korean service plans to PRDs and business plans, and evaluate whether conflict-resolution questions actually reduce review time and increase team agreement.

Built With

Share this project:

Updates