- SKILL.md 110 lines · 3.7 KB
Witness: evidence instead of "done"
When to use
- Before the work starts: fix the acceptance criteria and capture a baseline, so the criteria cannot later be bent to fit the result.
- The agent finished a task and is about to say "done", "fixed", "passing".
- The user asks whether the feature really works.
Hard rule
Do not claim a task is complete until
witness runexits 0 in the same message. Manual criteria (skip) do not count as proof.
If the check was not run, say "not verified", not "should work".
Workflow
Check the CLI.
bashwitness --versionIf the command is missing, stop. Do not install anything and do not run a package manager: point the human to the install instructions in the project README and wait for them to confirm. Installing software is the human's decision, not the agent's.
Initialise.
bashwitness initIf
.witness/already exists, do not overwrite the settings.Acceptance criteria — with the human.
bashwitness claim "<task>"Show the criteria and ask the human to approve them. Do not use
--yesunless the human explicitly approved: the point of the product is that a human approves the criteria, not the model.Baseline before the work.
bashwitness run --baselineThe app must be running. For a non-default address use
--base-url http://127.0.0.1:<port>. The baseline is what separates a real check from a tautology.If the work already started or finished before this skill was loaded, an honest baseline no longer exists. Say so explicitly, run the checks anyway, and report "no baseline: before/after comparison unavailable". Never run
--baselineafter the fact: that would present the finished app as the starting point and fake the differential.After the work.
bashwitness runRead the
detailof failed criteria and thedifferentialblock, not just the summary.Evidence.
bashwitness report # .witness/report.html witness badge # .witness/badge.svg + badge.json
If the agent runs as a single command, use witness wrap -- <command>: it performs the work
and immediately runs the check. The agent's code matters more than the check's code.
Forbidden
- Weakening criteria after a run, deleting checks, setting
required: falsejust to get green. - Running
--baselineafter the work to fabricate a before/after comparison. - Treating
skip(manual steps) as proof of completion. - Checking a neighbouring route when the task names a different one.
- Summarising the agent's report instead of the
witness runoutput. - Saying "tests passed" when no run happened in that message.
Response format
Verification: witness run → exit 0
Criteria: 4, machine-checked: 3, manual: 1
Passed: 3/3 machine checks
Tautologies: none
Report: .witness/report.htmlIf exit != 0:
Verification: witness run → exit 1
Failed:
- <criterion>: <detail>
Next step: <concrete action>If the checks ran without a baseline, keep the same format and append
(no baseline: before/after comparison unavailable) to the verification line.
Reply in the language the user writes in.
Boundaries
Witness checks only what the criteria say. It is not a code review, does not hunt for vulnerabilities and does not judge code quality. A green badge means "the approved checks passed", not "the product is good".
Scanned by
NVIDIA SkillSpector
Automated static analysis (no LLM) · v2.12.0 · 2026-11-20
Partial analysis: some checks could not run on this component.
0 observations
SkillSpector looks for patterns such as shell commands, credential access, network calls or unpinned dependencies. An observation is something worth reading in the files before you install, not a confirmed vulnerability. Components that run code locally, like hooks, MCP servers and security tooling, naturally show more of them. Open one to see the lines it points at.
No observations.
Learn more about NVIDIA SkillSpector or how components are reviewed.
Agent compatibility
Checked by NVIDIA SkillEvaluator Harness Check · v1.0.1 · 2026-11-20
How each agent reads this skill once it is installed, from the agents' documented loading rules. No agent was run. A note means the agent loads the skill but changes part of its text, or drops it.
Claude Code
loadsCodex
loadsCopilot CLI
loadsOpenCode
loadsHermes
loadsOpenClaw
loads
Every agent loads this skill as written. No notes.
Learn more about NVIDIA SkillEvaluator.