Integrate with a coding agent
Coding agents integrate projects with QA-Board well when they know its conventions: what run() returns, how metrics
are declared, what a batch is, and that a change is better only once its runs beat the reference's. Two skills in the
repository say that, in the agents' format. Point your agent at them instead of pasting the docs.
The skills are new. Tell us what your agent got wrong, and we'll fix the skill.
Install the skills
Copy them into your project, where your agent looks for skills (.claude/skills/ for Claude Code), or reference them
from your AGENTS.md:
git clone --depth 1 https://github.com/Samsung/qaboard.git /tmp/qaboard
mkdir -p .claude/skills
cp -r /tmp/qaboard/skills/qaboard-integration /tmp/qaboard/skills/qaboard-results .claude/skills/
| Skill | What the agent learns |
|---|---|
qaboard-integration | To wrap a project: the shape to pick (wrap, import, evaluate a model), qaboard.yaml, run() and postprocess(), metrics with Metric/define_metric, visualizations, batches, the CI job and the gate, and a checklist before saying it's done. |
qaboard-results | To read results: qaboard.runs(url) and qa runs for the runs a page shows, the compare API and qa gate --json to check a change against the reference, and what not to touch on the storage. |
Then ask: "Integrate this project with QA-Board; the inputs are in data/slices, the CI should run the 4 small
slices", or "Did my change improve the detector? Check with QA-Board."
Tell agents how your project uses QA-Board
The integration skill ends by adding a section to your project's AGENTS.md or CLAUDE.md: how to run the batch,
the metrics that matter and their direction, and that results must be compared with the reference before a change is
called better. Start from
templates/AGENTS.md.
What agents can read
Agents use the same APIs as scripts: qaboard.runs(url) for the runs a web app page
shows, the compare API and qa gate --json for a verdict against the
reference, and the HTTP API for everything else. An MCP server and JSON output from every qa
command are on the roadmap.
- Project setupWhat the agent will create: qaboard.yaml and the entrypoint.
- Results in scriptsThe runs a page shows, from Python and the shell.
- Quality gateFail the CI when metrics regress.