/proof
Use when about to claim work is complete, fixed, or passing, before committing or creating PRs
Claude Code
$ npx -y skills add iha2/superlite --skill proof --agent claude-code
Codex
$ npx skills add iha2/superlite --skill proof -a codex -y
Cursor
$ npx skills add iha2/superlite --skill proof -a cursor -y
Gemini CLI
$ npx skills add iha2/superlite --skill proof -a gemini-cli -y
GitHub Copilot
$ npx skills add iha2/superlite --skill proof -a github-copilot -y
OpenCode
$ npx skills add iha2/superlite --skill proof -a opencode -y
Windsurf
$ npx skills add iha2/superlite --skill proof -a windsurf -y
SKILL.md
proof.SKILL.mdname: proof
description: Use when about to claim work is complete, fixed, or passing, before committing or creating PRs
Proof
Evidence before claims. Never say "done", "fixed" or "passing" without fresh proof from this session.
Acceptable evidence
- The one-line summary (`passed failed skipped seconds`) from the finish run after the last change: `bash <skill-dir>/../swarm/scripts/superlite-cli run_tests --full --reuse` or `run_tests --affected --base <start sha>`, per the plan's `**Full suite:**`.
- Or a focused run you executed yourself after the last change, covering exactly what you claim.
- For a build, a lint or a type check: that command's exit status and summary line.
- For a bug fix: the original reproduction now passing, plus its regression test.
Rules
1. Name the command that proves the claim. 2. Run it fresh, or reuse a result only if nothing changed since that green run. If nothing changed, do not re-run. 3. Read the 2 to 3 lines the claim rests on: the counts, the exit code, the failure ids. Not just "no error scrolled by". 4. State the result honestly: "47 passed, 0 failed (run_tests --affected, after the last edit)", naming which finish run applied,, or the actual failures. 5. Scope the claim to the evidence. A passing unit test does not prove the build works. A clean lint does not prove tests pass. An agent saying "DONE" does not prove the diff is right: check `git status --short` and `git diff --stat`, and read the evidence it reports.
Full-suite runs
Run `bash <skill-dir>/../swarm/scripts/superlite-cli run_tests --full --reuse` for full suites, `--affected` for changed code only (or `--fast PATHS` for a focused run); it prints one summary line, so logs stay out of your context. Run it once at the end, not after every small edit.
Requirements and delegated work
Re-read the plan or request and tick each requirement off against the code and tests. A green suite does not mean the spec is met. Confirm the files a subagent changed match what its task owned.
Red flags: stop and verify first
- "Should work now", "probably fine", "I'm confident".
- Satisfaction before the check ("Great!", "Done!").
- About to commit, push or open a PR without a check.
- Trusting a report without looking at its evidence.
- "Just this once", or being tired and wanting to finish.
Never claim
Success with failures you decided to ignore. Say what fails and why it is acceptable, or fix it.
Read more
name: proof description: Use when about to claim work is complete, fixed, or passing, before committing or creating PRs
Proof
Evidence before claims. Never say "done", "fixed" or "passing" without fresh proof from this session.
Acceptable evidence
- The one-line summary (`passed failed skipped seconds`) from the finish run after the last change: `bash <skill-dir>/../swarm/scripts/superlite-cli run_tests --full --reuse` or `run_tests --affected --base <start sha>`, per the plan's `**Full suite:**`.
- Or a focused run you executed yourself after the last change, covering exactly what you claim.
- For a build, a lint or a type check: that command's exit status and summary line.
- For a bug fix: the original reproduction now passing, plus its regression test.
Rules
1. Name the command that proves the claim. 2. Run it fresh, or reuse a result only if nothing changed since that green run. If nothing changed, do not re-run. 3. Read the 2 to 3 lines the claim rests on: the counts, the exit code, the failure ids. Not just "no error scrolled by". 4. State the result honestly: "47 passed, 0 failed (run_tests --affected, after the last edit)", naming which finish run applied,, or the actual failures. 5. Scope the claim to the evidence. A passing unit test does not prove the build works. A clean lint does not prove tests pass. An agent saying "DONE" does not prove the diff is right: check `git status --short` and `git diff --stat`, and read the evidence it reports.
Full-suite runs
Run `bash <skill-dir>/../swarm/scripts/superlite-cli run_tests --full --reuse` for full suites, `--affected` for changed code only (or `--fast PATHS` for a focused run); it prints one summary line, so logs stay out of your context. Run it once at the end, not after every small edit.
Requirements and delegated work
Re-read the plan or request and tick each requirement off against the code and tests. A green suite does not mean the spec is met. Confirm the files a subagent changed match what its task owned.
Red flags: stop and verify first
- "Should work now", "probably fine", "I'm confident".
- Satisfaction before the check ("Great!", "Done!").
- About to commit, push or open a PR without a check.
- Trusting a report without looking at its evidence.
- "Just this once", or being tired and wanting to finish.
Never claim
Success with failures you decided to ignore. Say what fails and why it is acceptable, or fix it.
A fast software factory workflow for Claude Code: one approval gate, parallel waves, routed models
Other skills on superlite.
More skills from other plugins.
agentmail
Use your assigned AgentMail inbox to read email tasks, explicitly send or reply, and check delivery. Provided automatically by…
adr-create
Create a new Architecture Decision Record with sequential numbering and AgentDB registration
Questions about proof.
- Which coding agents can use proof?
- proof installs in Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode and Windsurf with the skills CLI, using each agent's flag shown on this page. Flowy's auto-invocation, which fires a skill when your prompt matches, runs in Claude Code only.

