We've added docs-kit to Labs — a design-doc manager for AI-assisted coding
We have published docs-kit, built in-house, on Labs: a design-doc management tool for AI-assisted coding. An AI does not question a design document, so a stale one acts as a wrong instruction. docs-kit judges freshness from git history and lets an AI open only the documents that matter — cutting both rework and tokens.
We have published our own docs-kit on Labs: a design-doc management tool for AI-assisted coding. The user guide and the binaries are up with it.
Once an AI writes most of the implementation, the bottleneck moves onto the design docs (→ The paradigm shift AI coding brings).
An AI doesn't question a design doc
A person notices within a few lines. "That function name can't exist any more" — and stops to check the code.
An AI doesn't doubt. If it says status: current, it treats it as current, and because it is obedient it brings the implementation towards the document. It starts a new feature on an API name we removed. It rewrites working code back into an old design. Every one of those is the AI following instructions faithfully, so no amount of prompt-tuning fixes it. The cause is that the information we hand it is out of date, so the information is what has to be fixed.
git does the judging
Each document's front matter records the code it is responsible for, and the commit it was last reconciled against.
scope: [src/core/editing/**] # code this doc covers
verified_at: 8f3a1c2e0b7d4a91c3f6e2d8b5a704c1e9f3d602 # last reconciled commit
If code inside scope moved after verified_at, the document is stale.
$ docs-kit stale
SPEC-editing-model stale code-changed: 3 files changed since 8f3a1c2
ARCH-overview ok (verified at 8f3a1c2)
ok 1 / stale 1 / broken 0 / skipped 0
The verdict rests on a fact in the repository, not on a self-reported "last updated" line. It fails CI, and the built-in MCP server hands the same verdict to the AI.
Less to read means fewer tokens
Throwing the docs away and handing over only code means making the AI read the whole codebase every time — which is expensive. Handing it every design doc is no better.
docs-kit's MCP server does not carry the prose. It returns paths, metadata and a 200-character summary. The AI rules out what is irrelevant from that, and opens only the documents it actually needs. Measured on a real repository: one reverse lookup returned 15 documents, and only 4 were opened.
The rework also disappears — no more rewriting code to match a stale document and reverting it later. Judging freshness is a way of saving tokens.
Ready to use
- Version — 1.1.0
- Platforms — macOS / Linux / Windows. A single Go binary;
gitis the only requirement - Format — plain Markdown with front matter. Drop docs-kit and the docs still read fine
- Download — Installation
- Documentation — the user guide, in Japanese and English
The design docs for the product we are working on now — EXP—002 on Labs — are managed with docs-kit too. An open-source release is in preparation.
A design-doc management tool for AI-assisted coding. It records scope and verified_at in front matter and judges freshness from git history. A single Go binary; git is the only requirement.
Rework caused by stale documents disappears, and because the MCP server carries no prose, the AI opens only what it needs. Less to read means fewer tokens.