mgiovani/ai-arsenal / skills/create-skill
create-skill
VerifiedCreate a new agent skill (or Claude Code slash command) from a plain-language description, using live spec fetching, pattern research, and an approval-gated blueprint before any files are written. Use whenever the user wants to build, scaffold, or author a new skill, subagent capability, or slash command, including phrasings like 'make a command for X', 'create a slash command', 'turn this into a reusable skill', or 'package this workflow as a skill'. Not for editing CLAUDE.md/AGENTS.md memory rules (use create-rule) or discovering/installing existing third-party skills (use find-skills).
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 20.3 KB |
| evals/evals.json | config | 3.5 KB |
| evals/trigger-eval.json | config | 2.2 KB |
| references/agent-prompts.md | markdown | 4.5 KB |
| references/eval-system.md | markdown | 5.6 KB |
| references/frontmatter-fields.md | markdown | 2.8 KB |
| references/schemas.md | markdown | 6.3 KB |
| references/skill-anatomy.md | markdown | 10.2 KB |
| references/specification-urls.md | markdown | 2.6 KB |
| scripts/generate_report.py | script | 11.5 KB |
| scripts/improve_description.py | script | 16.7 KB |
| scripts/package_skill.py | script | 4.5 KB |
| scripts/quick_validate.pyexecutable | script | 11.1 KB |
| scripts/run_eval.py | script | 8.1 KB |
references/agent-prompts.md
View on GitHub# Agent Prompts
Prompts for grader agents used in the eval system. Load this reference when running Phase 6 eval execution.
## Contents
- [Grader Agent](#grader-agent)
- [Batch Grader Prompt](#batch-grader-prompt)
- [Description Improvement Prompt](#description-improvement-prompt)
## Grader Agent
Used to score eval results with blind comparison. Receives both outputs (labeled A and B, not "with-skill" and "baseline") to prevent bias.
### Prompt Template
```
You are a skill quality evaluator. You will receive two outputs from the same prompt,
labeled Output A and Output B, and a list of assertions about what good output looks like.
Your job:
1. Check each assertion against both outputs
2. Score Output A on a 1-5 scale
3. State which output is better overall (or if they're equivalent)
4. Provide one specific, actionable suggestion for improvement
PROMPT THAT WAS SENT:
{eval_prompt}
ASSERTIONS (what good output should satisfy):
{assertions_formatted}
OUTPUT A:
{output_a}
---
OUTPUT B:
{output_b}
---
Respond with ONLY valid JSON in this exact format:
{
"expectations": [
{
"assertion": "<assertion text>",
"output_a_pass": <true|false>,
"output_b_pass": <true|false>,
"reasoning": "<one sentence explanation>"
}
],
"summary": "<2-3 sentence explanation of overall quality difference>",
"score": <1-5 integer for Output A>,
"comparison": "<output_a_better|output_b_better|equivalent>",
"eval_feedback": "<one specific, actionable suggestion to improve either the skill or the eval assertions>"
}
Scoring guide for Output A:
1 - Completely wrong or unhelpful
2 - Partially correct but major issues
3 - Acceptable but notable gaps
4 - Good, minor issues only
5 - Excellent, satisfies all assertions
Important: You do not know which output came from a skill and which is baseline.
Grade purely on quality relative to the assertions.
```
### Usage in Phase 6
When spawning the grader as a Task agent:
```
Grader Agent:
- subagent_type: "general-purpose"
- model: "sonnet"
- prompt: [Fill template above with eval_prompt, assertions_formatted, output_a, output_b]
```
Randomly assign which output is A vs B for each grader call. Track the mapping separately so you can interpret the results correctly.
### Interpreting Results
**score field**: Quality of Output A (1-5). If you assigned with-skill as A, this scores the skill output.
**comparison field**: Which is better overall:
- `output_a_better`: A won
- `output_b_better`: B won
- `equivalent`: No meaningful difference
**eval_feedback field**: Actionable suggestion. Can target:
- The skill (improve instructions, add missing steps)
- The description (improve trigger precision)
- The evals (improve assertion specificity)
### Score Interpretation
| Score | Meaning | Action |
|-------|---------|--------|
| 5 | Excellent | Ship it |
| 4 | Good | Minor tweaks optional |
| 3 | Acceptable | Identify the gap, fix, re-eval |
| 2 | Needs work | Major revision needed |
| 1 | Broken | Rethink the approach |
Target: score ≥ 4 AND comparison = `output_a_better` (assuming with-skill = Output A).
## Batch Grader Prompt
For grading multiple evals in one agent call (reduces Task tool invocations):
```
You are a skill quality evaluator. Grade the following {n} eval results.
For each eval, you receive:
- The prompt sent
- Output A and Output B (you don't know which is which)
- Assertions about expected output
Return a JSON array with one grading object per eval.
EVALS TO GRADE:
{evals_formatted}
Each grading object must follow this schema:
{
"eval_id": "<id>",
"expectations": [...],
"summary": "...",
"score": <1-5>,
"comparison": "<output_a_better|output_b_better|equivalent>",
"eval_feedback": "..."
}
Return ONLY the JSON array, no other text.
```
Use the batch prompt when grading 3+ evals to reduce latency and cost.
## Description Improvement Prompt
Used by `improve_description.py` to generate an improved description after seeing failures:
```
Current skill description:
{current_description}
The description was tested against 20 trigger queries (10 that should activate the skill,
10 that should not). Here are the failures:
Failed should-trigger queries (skill didn't activate when it should have):
{failed_should_trigger}
Failed should-not-trigger queries (skill activated when it shouldn't have):
{failed_should_not_trigger}
Write an improved description that:
1. Covers the missed trigger phrasings
2. Avoids false positives for out-of-scope queries
3. Stays under 1024 characters
4. Remains assertive and specific
Return ONLY the new description text, no other content.
```
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |