Say what the software should do.
You keep prompting your coding agent. Kibi writes that intent down beside the code, carries it onto every branch, and treats “done” as a claim that needs fresh evidence.
Prompt the intent. Kibi makes the agent remember it—and prove the implementation.
What usually gets lost
- The prompt. The session ends, and the reason for yesterday’s change is gone.
- The link to the code. A ticket can say what you wanted. It rarely names the function that was supposed to do it.
- The meaning of green. A passing suite does not say which product behavior is still true on this commit.
What you actually do
-
State the behavior
Prompt the way you already do. “Draft edits must save when someone navigates away.” You do not fill in a requirements form.
-
Approve real decisions
The agent proposes a plan before it writes project knowledge. You approve it, or you correct the product call. The bookkeeping stays with the agent.
-
Read the result
The health report names what is proven, what still has a gap, and what contradicts itself. A green test run is a different fact.
Where to go next
Use Kibi on a repository
Install it, connect the agent you already use, and learn how to read the first health report. Written for the person prompting, not for the agent.
- What is Kibi?
- Quick start
- Connect your coding agent
- Read the health report
Look up the exact contract
Commands, tool schemas, entity fields, and the checks Kibi actually runs. Dry on purpose, and rendered from the same files as the repository.
- CLI reference
- MCP tools
- Entity schema
- Error reference
Add Kibi to the repository
Install the three packages the project will run, then create the local .kb/ directory. That command does not invent what the product does.
npm install --save-dev kibi-core kibi-cli kibi-mcp
npm exec -- kibi init
pnpm add -D kibi-core kibi-cli kibi-mcp
pnpm exec kibi init
yarn add -D kibi-core kibi-cli kibi-mcp
yarn exec kibi init
Requires Node.js 22+ and swipl (SWI-Prolog 9.0+) on your PATH. Bun commands and per-platform setup are in the installation guide.
Questions people ask first
Do I write the requirements myself?
No. You describe the behavior. Your agent turns that into requirements, scenarios, tests, and links to the code. You step in when the product decision is actually ambiguous.
Does this replace my issue tracker?
No. Keep the tracker you have. Kibi connects a decision to the code that implements it, and to evidence that the code still does it.
What counts as proven?
A requirement is proven only when a test that claims to verify it has fresh end-to-end evidence for the current code. A passing unit test, a coverage percentage, or an old receipt does not count.
Will Kibi decide the product for me?
No. kibi init creates the directory and the Git hooks. It does not infer behavior. Checks prove what was encoded. Missing knowledge stays visible.
This repository publishes its own report
Every push to the default branch republishes requirement health for Kibi itself: how many current requirements are fully proven, which ones contradict each other, and which evidence has gone stale.