Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Generates overkill, maximally-optimized prompts for coding agents and agentic services. Use this skill when you need to optimize system prompts, design custom agentic instructions, configure tool permissions, or construct robust multi-agent orchestration instructions. Triggers on requests containing: "overengineer prompt", "optimize agent instructions", "build system prompt", "agentic prompt engineering", "make agentic workflow instructions". Exclude when: requests are for general programming, s
.claude/skills/hcswint7-prompt-overengineer/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 31% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 88% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 194% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 150% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 79% | 0% |
This skill instructs the agent on how to build maximally-optimized, production-grade system prompts for coding agents and agentic workflows. It leverages advanced techniques in context engineering, token discipline, hooks integration, and structured output contracts to design prompts that maximize task completion rate and safety while reducing inference cost.
When tasked with generating or optimizing an agent prompt, follow these six sequential steps:
Analyze the target agent's goals and runtime environment. Identify:
To keep the active context size lean, do not read all reference files upfront. Dynamically view only the reference files matching the task requirements:
Select at least seven (7) optimization concepts from references/concept-index.md that directly benefit the target agent. You must justify your selection (e.g., "Selecting PREFIX_CACHING and TIERED_MEMORY because...").
CoT or ReAct), structured outputs (OUTPUT_CONTRACT), context discipline (TOKEN_BUDGET), and safety boundaries (SCOPE_BOUNDARIES).Draft the optimized prompt by matching the structure of assets/overkill-template.md.
Evaluate your drafted prompt by running the evaluation script:
powershellpython "${CLAUDE_SKILL_DIR}/scripts/score-prompt.py" --prompt-file <path_to_drafted_prompt>
If the prompt scores below 8.0 out of 10.0, read the JSON output's suggestions, revise the prompt, and re-run the scorer. Do not output the prompt to the user until it meets or exceeds the 8.0 threshold.
Provide the finalized prompt to the user as a markdown code block or file, accompanied by a brief summary explaining:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 17,020 | 6,879 | -60% | 1 | 1 | 0% | 2,783 | 1,552 | -44% | 0 | 0 | — |
case-02 | fail→fail | 21,942 | 32,999 | +50% | 1 | 1 | 0% | 3,737 | 7,150 | +91% | 0 | 0 | — |
case-03 | fail→fail | 21,197 | 10,743 | -49% | 1 | 1 | 0% | 3,548 | 1,843 | -48% | 0 | 0 | — |
case-04 | fail→pass | 20,063 | 28,351 | +41% | 1 | 1 | 0% | 3,715 | 4,855 | +31% | 0 | 0 | — |
case-05 | fail→pass | 16,507 | 23,685 | +43% | 1 | 1 | 0% | 2,782 | 5,230 | +88% | 0 | 0 | — |
case-06 | pass→fail | 16,355 | 7,520 | -54% | 1 | 1 | 0% | 3,085 | 1,362 | -56% | 0 | 0 | — |
case-07 | fail→pass | 10,957 | 26,485 | +142% | 1 | 1 | 0% | 1,948 | 5,721 | +194% | 0 | 0 | — |
case-08 | fail→pass | 13,226 | 31,703 | +140% | 1 | 1 | 0% | 2,082 | 5,202 | +150% | 0 | 0 | — |
case-09 | fail→fail | 13,350 | 32,108 | +141% | 1 | 1 | 0% | 2,263 | 6,246 | +176% | 0 | 0 | — |
case-10 | fail→fail | 15,166 | 33,491 | +121% | 1 | 1 | 0% | 2,634 | 6,345 | +141% | 0 | 0 | — |
case-11 | fail→fail | 22,896 | 22,743 | -1% | 1 | 1 | 0% | 2,125 | 5,013 | +136% | 0 | 0 | — |
case-12 | fail→fail | 17,417 | 23,802 | +37% | 1 | 1 | 0% | 1,401 | 5,233 | +274% | 0 | 0 | — |
case-13 | fail→pass | 17,309 | 28,290 | +63% | 1 | 1 | 0% | 2,964 | 5,313 | +79% | 0 | 0 | — |
case-14 | fail→pass | 11,365 | 16,708 | +47% | 1 | 1 | 0% | 1,896 | 4,081 | +115% | 0 | 0 | — |
case-19 | fail→pass | 14,234 | 2,553 | -82% | 1 | 1 | 0% | 2,489 | 1,429 | -43% | 0 | 0 | — |
case-15 | fail→pass | 28,451 | 29,264 | +3% | 1 | 1 | 0% | 2,012 | 5,268 | +162% | 0 | 0 | — |
case-16 | fail→fail | 16,520 | 27,349 | +66% | 1 | 1 | 0% | 2,935 | 5,786 | +97% | 0 | 0 | — |
case-17 | pass→pass | 10,996 | 23,255 | +111% | 1 | 1 | 0% | 1,858 | 5,552 | +199% | 0 | 0 | — |
case-18 | fail→fail | 13,364 | 8,144 | -39% | 1 | 1 | 0% | 2,224 | 2,492 | +12% | 0 | 0 | — |
case-20 | pass→pass | 12,390 | 12,100 | -2% | 1 | 1 | 0% | 2,340 | 3,190 | +36% | 0 | 0 | — |
case-21 | pass→pass | 13,596 | 26,806 | +97% | 1 | 1 | 0% | 3,165 | 7,100 | +124% | 0 | 0 | — |
case-22 | pass→pass | 9,419 | 8,255 | -12% | 1 | 1 | 0% | 1,902 | 2,696 | +42% | 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 19 counted toward the lift figure. The other 3 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 +32 percentage points is the difference between those two pass rates over the 19 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.