Perfect It
Quick start
When invoked, answer this question before proceeding:
Am I 100% confident in this strategy?
If the answer is not strictly yes, run the loophole loop below. Do not claim 100% confidence unless the confidence gate passes.
Core rule
The target is strict 100% confidence. This does not permit fake certainty. If 100% cannot be reached with available evidence, say NOT 100% CONFIDENT, list the blockers, and specify the evidence or work required to reach 100%.
Loophole loop
Repeat until the confidence gate passes or a blocker remains:
Restate the strategy
- Define the exact strategy, scope, success criteria, non-goals, and failure modes.
- Convert vague claims into falsifiable checks.
Find loopholes adversarially
- Enumerate ways the strategy could fail: hidden assumptions, edge cases, missing requirements, security/privacy risks, data integrity issues, operational risks, UX risks, dependency/version risks, cost/performance risks, and rollback risks.
- For code or repo work, inspect the relevant files instead of relying on memory.
- For library/API claims, verify against authoritative current docs when behavior matters.
Fix properly
- For each material loophole, change the strategy so the loophole is impossible, detected mechanically, or explicitly accepted with rationale.
- Prefer simple, enforceable fixes over documentation-only promises.
Verify with evidence
- Run the strongest available checks: tests, typechecks, linters, code search, docs lookup, reproduction, invariants, examples, or peer/subagent review.
- Record the evidence. If a check cannot run, record why and treat it as a confidence blocker unless immaterial.
Re-score confidence
- If any material loophole, unverified critical assumption, or unresolved contradiction remains, confidence is not 100%; loop again.
Confidence gate
Only output 100% CONFIDENT when all are true:
- Success criteria are explicit and testable.
- All material loopholes found so far are fixed, made impossible, or proven irrelevant.
- Critical assumptions are backed by evidence, not vibes.
- Verification has run successfully, or skipped checks are proven immaterial.
- The revised strategy is simpler or at least not more fragile than the original.
Otherwise output NOT 100% CONFIDENT and the shortest path to reach 100%.
Output format
Use REPORT_TEMPLATE.md for substantial work. For short answers, use:
Confidence: 100% CONFIDENT | NOT 100% CONFIDENT
Strategy:
- ...
Loopholes found:
- ...
Fixes applied:
- ...
Evidence:
- ...
Remaining blockers:
- None | ...
Final strategy:
- ...
Validation
For report files, run:
python ~/.agents/skills/perfect-it/scripts/check-report.py path/to/report.md
1---2name: perfect-it3description: Stress-tests strategies, plans, designs, architectures, and uncertain answers until every material loophole is fixed or explicitly blocked. Use when the user says /perfect-it, "are you 100% confident", "perfect it", asks to find loopholes, asks for a higher-confidence strategy, or when the agent is not fully confident in its own strategy or answer.4---56# Perfect It78## Quick start910When invoked, answer this question before proceeding:1112> Am I 100% confident in this strategy?1314If the answer is not strictly yes, run the loophole loop below. Do not claim 100% confidence unless the confidence gate passes.1516## Core rule1718The target is strict 100% confidence. This does **not** permit fake certainty. If 100% cannot be reached with available evidence, say `NOT 100% CONFIDENT`, list the blockers, and specify the evidence or work required to reach 100%.1920## Loophole loop2122Repeat until the confidence gate passes or a blocker remains:23241. **Restate the strategy**25 - Define the exact strategy, scope, success criteria, non-goals, and failure modes.26 - Convert vague claims into falsifiable checks.27282. **Find loopholes adversarially**29 - Enumerate ways the strategy could fail: hidden assumptions, edge cases, missing requirements, security/privacy risks, data integrity issues, operational risks, UX risks, dependency/version risks, cost/performance risks, and rollback risks.30 - For code or repo work, inspect the relevant files instead of relying on memory.31 - For library/API claims, verify against authoritative current docs when behavior matters.32333. **Fix properly**34 - For each material loophole, change the strategy so the loophole is impossible, detected mechanically, or explicitly accepted with rationale.35 - Prefer simple, enforceable fixes over documentation-only promises.36374. **Verify with evidence**38 - Run the strongest available checks: tests, typechecks, linters, code search, docs lookup, reproduction, invariants, examples, or peer/subagent review.39 - Record the evidence. If a check cannot run, record why and treat it as a confidence blocker unless immaterial.40415. **Re-score confidence**42 - If any material loophole, unverified critical assumption, or unresolved contradiction remains, confidence is not 100%; loop again.4344## Confidence gate4546Only output `100% CONFIDENT` when all are true:4748- Success criteria are explicit and testable.49- All material loopholes found so far are fixed, made impossible, or proven irrelevant.50- Critical assumptions are backed by evidence, not vibes.51- Verification has run successfully, or skipped checks are proven immaterial.52- The revised strategy is simpler or at least not more fragile than the original.5354Otherwise output `NOT 100% CONFIDENT` and the shortest path to reach 100%.5556## Output format5758Use `REPORT_TEMPLATE.md` for substantial work. For short answers, use:5960```md61Confidence: 100% CONFIDENT | NOT 100% CONFIDENT6263Strategy:64- ...6566Loopholes found:67- ...6869Fixes applied:70- ...7172Evidence:73- ...7475Remaining blockers:76- None | ...7778Final strategy:79- ...80```8182## Validation8384For report files, run:8586```bash87python ~/.agents/skills/perfect-it/scripts/check-report.py path/to/report.md88```
arkan