Enforces "prove before claim" discipline - validation, testing, and evidence
requirements before declaring work complete.
Triggers: completion, finished, done, working, should work, configured,
ready to use, implemented, fixed
Use when: claiming ANY work is complete, recommending solutions, stating
something will work, finishing implementations
DO NOT use when: explicitly asking questions or requesting clarification
DO NOT use when: work is clearly in-progress and not claiming completion
CRITICAL: This skill is MANDATORY before any completion claim. Violations
result in wasted time and eroded trust.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Enforces "prove before claim" discipline - validation, testing, and evidence
requirements before declaring work complete.
Triggers: completion, finished, done, working, should work, configured,
ready to use, implemented, fixed
Use when: claiming ANY work is complete, recommending solutions, stating
something will work, finishing implementations
DO NOT use when: explicitly asking questions or requesting clarification
DO NOT use when: work is clearly in-progress and not claiming completion
CRITICAL: This skill is MANDATORY before any completion claim. Violations
result in wasted time and eroded trust.
Philosophy: "Trust, but verify" is wrong. "Verify, THEN trust" is correct.
The Problem This Solves
Anti-Pattern:
"I've configured LSP for you. Just restart your session and it will work!"
Reality: Didn't test if cclsp starts, didn't verify tools are exposed, didn't check for known bugs.
Correct Pattern:
"Let me verify LSP configuration works..."
# Test cclsp starts
CCLSP_CONFIG_PATH=... npx cclsp@latest &
# Verify language servers respond
pylsp --help# Check for known issues
<web search for bugs in current version>
"Found critical issue: Claude Code 2.0.76 has broken LSP (Issue #14803). Here's proof..."
Core Principle
Before claiming completion, you must provide evidence that the solution actually works (tested), edge cases are handled (validated), claims are accurate (proven), and future verification is possible (reproducible).
When to Use
MANDATORY Usage (Non-Negotiable)
This skill is required before any statement like "this will work", "should work", or "is ready". Apply it before claiming configuration or setup is complete, before recommending solutions without testing them, before saying "done" or "finished", and before telling users to "try this" or "restart and test".
Red Flags (You're About to Violate This)
Thought Pattern
Reality Check
"This configuration looks correct"
Did you TEST it works?
"After restart it should work"
Did you VERIFY the restart fixes it?
"The setup is complete"
Did you PROVE each component functions?
"Just enable X and Y"
Did you CHECK for known issues?
"This will fix your problem"
Did you REPRODUCE the problem first?
"The language servers are installed"
Did you CONFIRM they respond correctly?
Required TodoWrite Items
When applying this skill, create these todos:
proof:problem-reproduced - Verified the issue exists
proof:solution-tested - Tested proposed solution works
proof:edge-cases-checked - Validated common failure modes
proof:evidence-captured - Logged evidence per evidence-logging skill
proof:completion-proven - Demonstrated acceptance criteria met
Validation Protocol
Step 1: Reproduce the Problem (proof:problem-reproduced)
BEFORE proposing a solution:
# Example: User says LSP doesn't work# FIRST: Verify current state
ps aux | grep cclsp # Is it running?echo$ENABLE_LSP_TOOLS# Is env var set?cat ~/.claude/.mcp.json | grep cclsp # Is it configured?
# Example: Testing if cclsp can start
CCLSP_CONFIG_PATH=/path/to/config npx cclsp@latest 2>&1 &
sleep 3
ps aux | grep cclsp | grep -v grep
# Did it actually start? Capture the outputtail /tmp/cclsp-test-output.log
Evidence Required:
Successful execution of proposed solution
Actual output (not assumed output)
Confirmation of expected behavior
Step 3: Check for Known Issues (proof:edge-cases-checked)
BEFORE recommending approach:
# Example: Research known bugs
<web search for"claude code LSP 2.0.76 issues">
<check GitHub issues for current version>
<verify language server compatibility>
## Definition of Done
User can successfully use LSP tools after following these steps:
### Acceptance Criteria- [ ] cclsp MCP server starts without errors
- [ ] Language servers (pylsp, typescript-language-server) respond
- [ ] LSP tools (find_definition, find_references) are callable
- [ ] No known bugs in current Claude Code version block functionality
### Test Evidence- [E1] cclsp process running: PASS (see evidence)
- [E2] pylsp responds: PASS (see evidence)
- [E3] LSP tools available: FAIL - **BLOCKED by bug #14803**### Conclusion
Cannot claim completion due to fundamental blocker identified. Can provide: diagnosis with evidence, workaround options, and next steps.
Integration with Other Skills
With scope-guard
Before implementation:
scope-guard: "Should we build this?"
proof-of-work: (not yet applicable)
After implementation:
proof-of-work: "Did we actually complete it?"
With evidence-logging
evidence-logging: HOW to capture evidence
proof-of-work: WHEN to capture evidence (always, before completion)
### Pre-Completion Validation- [ ] Problem reproduced with evidence
- [ ] Solution tested in actual environment
- [ ] Known issues researched (web search, GitHub, docs)
- [ ] Edge cases considered (what could go wrong?)
- [ ] Evidence captured in reproducible format
- [ ] Acceptance criteria defined and met
- [ ] User can independently verify claims
### Completion Statement Format
Instead of: "LSP is configured. Restart and it will work."
Required:
"I've verified your LSP configuration with these tests:
- [PASS] cclsp installed and configured in .mcp.json [E1]
- [PASS] Language servers installed and responsive [E2]
- [FAIL] LSP tools unavailable - discovered bug #14803 [E3]
**Proven Status:** Blocked by known issue in Claude Code 2.0.76
**Evidence:** See [E3] web search results
**Options:** 1) Downgrade to 2.0.67, 2) Wait for bug fix
**Cannot claim:** 'LSP will work after restart' (proven false)"
Red Flag Self-Check
Before sending completion message, ask:
Did I actually RUN the commands I'm recommending?
No → STOP, test them first
Yes → Capture output as evidence
Did I search for known issues with this approach?
No → STOP, research first
Yes → Document findings
Can the user reproduce my validation steps?
No → STOP, make it reproducible
Yes → Include steps in response
Am I assuming vs. proving?
Assuming → STOP, test your assumptions
Proving → Show the proof
Would I accept this level of validation from a coworker?
