Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when you need maximum precision on a critical task — production deployments, security-sensitive code, financial calculations, or any work where mistakes are unacceptable.
.claude/skills/sharpdeveye-zero-defect/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | 152% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 439% | 0% |
| case-10 | ✓→✗ | ▼ Worse | -36% | 0% |
| case-09 | ✓→✓ | = Same ✓ | 56% | 0% |
| case-16 | ✓→✓ | = Same ✓ | 168% | 0% |
Invoke /agent-workflow — it contains workflow principles, anti-patterns, and the Context Gathering Protocol. Follow the protocol before proceeding — if no workflow context exists yet, you MUST run /teach-maestro first.
Activate maximum precision mode for the current session. This command establishes execution-time discipline — not for the workflow design, but for how the agent itself operates. Every response, every code change, every claim must pass the zero-defect gate.
This is model-agnostic. It works with Claude, Gemini, GPT, Codex, or any AI agent.
Follow these for the entire session after this command is invoked:
| # | Rule | Why | |---|------|-----| | 1 | Read before writing — Re-read the relevant code/context before every modification | Prevents edits based on stale mental models | | 2 | Verify before claiming — Run verification before any success claim | Prevents false completion reports | | 3 | One logical change at a time — Avoid sprawling multi-file edits that compound errors | Reduces cascading failures | | 4 | State uncertainty explicitly — Say "I'm not sure about X" instead of guessing | Prevents confident hallucination | | 5 | Check every import and reference — Verify that every function, variable, and module exists | Prevents "symbol not found" errors | | 6 | Dry-run mentally before committing — Trace the code path for both happy path and edge cases | Catches logic errors before they ship | | 7 | Never hallucinate APIs — Only use functions, methods, and parameters that exist in the codebase or documentation | Prevents non-existent API calls | | 8 | Re-derive, don't recall — For math, logic, or complex reasoning, work it out fresh instead of from memory | Prevents confident but wrong answers |
Before claiming ANY work is complete, pass every item:
| Sloppy pattern | What to do instead | |---------------|-------------------| | "This should work" without testing | Run the test, show the output | | Editing code without re-reading the file first | View the file, then edit | | Assuming a function exists because it sounds right | Grep the codebase to confirm | | Making 5+ file changes in one shot | Break into sequential, verifiable steps | | Saying "Done!" before verification | Run build/test, paste the result | | Guessing at API parameters | Read the actual function signature | | Fixing a bug by changing something nearby | Trace the actual root cause first | | "I'm confident this is correct" | Confidence is not evidence — verify |
This command applies to every interaction for the remainder of the session. There are no exceptions. The rules apply to:
If you catch yourself about to violate a rule, stop and correct course before responding.
After the critical work is done, run /evaluate to review the output quality, or /refine for a final polish pass.
NEVER:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,753 | 4,382 | +17% | 1 | 1 | 0% | 608 | 1,279 | +110% | 0 | 0 | — |
case-02 | fail→fail | 3,868 | 4,727 | +22% | 1 | 1 | 0% | 233 | 1,193 | +412% | 0 | 0 | — |
case-03 | fail→fail | 7,486 | 2,451 | -67% | 1 | 1 | 0% | 1,273 | 1,030 | -19% | 0 | 0 | — |
case-04 | fail→fail | 5,128 | 11,574 | +126% | 1 | 1 | 0% | 761 | 2,908 | +282% | 0 | 0 | — |
case-05 | fail→fail | 13,784 | 8,538 | -38% | 1 | 1 | 0% | 3,107 | 2,542 | -18% | 0 | 0 | — |
case-06 | fail→pass | 5,052 | 8,082 | +60% | 1 | 1 | 0% | 939 | 2,362 | +152% | 0 | 0 | — |
case-07 | fail→fail | 4,099 | 14,039 | +242% | 1 | 1 | 0% | 764 | 3,437 | +350% | 0 | 0 | — |
case-08 | fail→fail | 1,941 | 6,805 | +251% | 1 | 1 | 0% | 323 | 2,028 | +528% | 0 | 0 | — |
case-09 | pass→pass | 8,050 | 8,753 | +9% | 1 | 1 | 0% | 1,744 | 2,717 | +56% | 0 | 0 | — |
case-10 | pass→fail | 9,629 | 2,032 | -79% | 1 | 1 | 0% | 1,959 | 1,247 | -36% | 0 | 0 | — |
case-11 | fail→fail | 6,792 | 10,204 | +50% | 1 | 1 | 0% | 498 | 2,777 | +458% | 0 | 0 | — |
case-12 | fail→fail | 9,017 | 5,383 | -40% | 1 | 1 | 0% | 1,400 | 1,731 | +24% | 0 | 0 | — |
case-13 | fail→fail | 8,696 | 6,714 | -23% | 1 | 1 | 0% | 1,276 | 2,010 | +58% | 0 | 0 | — |
case-14 | fail→fail | 5,437 | 1,519 | -72% | 1 | 1 | 0% | 1,040 | 1,110 | +7% | 0 | 0 | — |
case-15 | fail→pass | 3,539 | 17,520 | +395% | 1 | 1 | 0% | 666 | 3,590 | +439% | 0 | 0 | — |
case-16 | pass→pass | 8,982 | 15,682 | +75% | 1 | 1 | 0% | 1,432 | 3,834 | +168% | 0 | 0 | — |
case-17 | fail→fail | 5,394 | 16,090 | +198% | 1 | 1 | 0% | 874 | 3,561 | +307% | 0 | 0 | — |
case-18 | fail→fail | 5,971 | 1,877 | -69% | 1 | 1 | 0% | 1,114 | 1,201 | +8% | 0 | 0 | — |
case-19 | fail→fail | 12,113 | 6,001 | -50% | 1 | 1 | 0% | 2,343 | 1,317 | -44% | 0 | 0 | — |
case-20 | pass→pass | 20,793 | 24,514 | +18% | 1 | 1 | 0% | 3,542 | 5,101 | +44% | 0 | 0 | — |
case-21 | pass→pass | 15,237 | 25,803 | +69% | 1 | 1 | 0% | 2,754 | 5,027 | +83% | 0 | 0 | — |
case-22 | pass→pass | 10,922 | 11,984 | +10% | 1 | 1 | 0% | 1,735 | 2,896 | +67% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +5 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.