naman10parikh/agentgrid / .claude/skills/skill-creator
skill-creator
VerifiedAuto-generate new skills from learnings, approved patterns, and repeated user preferences. The meta-skill that makes the harness self-improving. Core to AutoLab.
When to Use
- After discovering a repeatable pattern across 2+ sessions
- When the user explicitly approves a workflow ("always do X")
- When LEARNINGS.md has 3+ entries about the same topic
- When a sub-agent finds a new capability worth capturing
- During AutoLab nightly improvement cycles
Process
1. Source Mining
Scan these sources for skill candidates:
memory/LEARNINGS.md— repeated corrections = skill gapsmemory/MEMORY.md— confirmed patterns = skill candidatesmemory/daily/*.md— workflows that worked = codify themresources/read/*.md— external techniques worth internalizing- Session transcript — what did the user approve? correct? praise?
packages/web/features.json— feature patterns that recur
2. Skill Candidate Evaluation
For each candidate, score on 4 axes:
| Axis | Question | Threshold |
|---|---|---|
| Frequency | How often would this skill be triggered? | 2+ times/week |
| Complexity | Does it need multi-step instructions? | 3+ steps |
| Error-prone | Do we keep getting it wrong without the skill? | 2+ mistakes |
| Reusable | Does it apply across agents, not just one? | Cross-agent preferred |
Score >= 3/4 axes → create the skill.
3. Skill Template
---
name: { skill-name }
description: { one-line description, under 100 chars }
---
## When to Use
- {trigger condition 1}
- {trigger condition 2}
## Process
### Step 1: {Action}
{What to do, what to check}
### Step 2: {Action}
{What to do, what to check}
## Output
- {What the skill produces}
- {How to verify it worked}
## Model Tier
- {Haiku | Sonnet | Opus} for each step
4. Skill Creation Checklist
- Write
SKILL.mdto.claude/skills/{skill-name}/SKILL.md - Description must be under 100 chars (for ~70 token routing)
- Add trigger patterns that the skill-routing system can match
- If the skill involves code patterns, include inline examples
- If the skill involves external tools, document the setup
- Update CLAUDE.md skills count
- Test: verify the skill is discoverable via
/skill-routing
5. Skill Categories
| Category | Purpose | Examples |
|---|---|---|
| Build | Code generation patterns | app-factory, agent-runtime |
| Meta | Self-improvement, research | self-improve, deep-think |
| Ops | Setup, deployment, infra | mcp-setup, secrets-setup |
| Agent | Agent-specific skills | Per-agent skills in agents/{name}/skills/ |
| Capture | Knowledge capture | transcript-to-insight, resource-digest |
6. For Agent-Level Skills
Skills can also be created for individual agents (not just the harness):
agents/{agent-name}/skills/{skill-name}.md
These are loaded into the agent's system prompt via the memory-loader. They follow the same template but are scoped to that agent's domain.
Integration with AutoLab
The Skill Creator is a key component of the AutoLab nightly cycle:
- AutoLab analyzes traces from the day's agent runs
- Identifies patterns where agents struggled or succeeded
- Calls Skill Creator to codify successful patterns
- New skills are available to all agents via Group Evolution
YouTube/Article Transcript Pipeline
When a YouTube video or article needs to be processed:
- Get the URL from user or resource queue
- For YouTube: Use computer use to visit a transcript service (e.g., tavio.io, kome.ai, or youtubetranscript.com)
- Enter the URL, download the transcript text
- Clean and structure the transcript
- Extract insights relevant to Energy platform
- Save to
resources/read/{topic}.md - If insights warrant a new skill, create it
X/Twitter Content Pipeline
- Use computer use to browse the tweet/thread
- Extract the full text content
- If it's a thread, compile all tweets
- Extract actionable insights
- Save to resources and update relevant docs
Skill Testing & Evals (from Anthropic's skill-creator enhancements, March 2026)
Two Kinds of Skills
- Capability uplift — helps Claude do something the base model can't do consistently (e.g., document creation, complex workflows). May become unnecessary as models improve.
- Encoded preference — sequences steps according to your team's process (e.g., NDA review, weekly updates). Durable, value depends on fidelity to actual workflow.
Writing Evals
For each skill, define test cases:
## Evals
### Test 1: {scenario}
- **Prompt:** "{test input that should trigger the skill}"
- **Expected:** {description of correct output}
- **Files:** {any test files needed}
### Test 2: {scenario}
...
Evals check that:
- The skill triggers when it should (no false negatives)
- The skill doesn't trigger when it shouldn't (no false positives)
- Output quality meets the bar defined in "Expected"
Benchmark Mode
Run standardized assessment across all evals:
- Track eval pass rate, elapsed time, token usage
- Run after model updates or skill edits
- Store results locally or integrate with CI
Multi-Agent Eval Execution
Spin up independent agents to run evals in parallel:
- Each eval in a clean context (no cross-contamination)
- Own token and timing metrics per eval
- Use
isolation: "worktree"for code-modifying skills
A/B Comparisons
Use comparator agents to judge:
- Skill v1 vs v2 (after edits)
- Skill vs no skill (is it still needed?)
- Judges don't know which output is which → unbiased
Description Tuning
As skill count grows, description precision is critical:
- Too broad → false triggers on unrelated prompts
- Too narrow → never fires when needed
- Analyze description against sample prompts
- Suggest edits that cut both false positives and false negatives
The Future: Skills → Specifications
As models improve, skills evolve from "how to do X" (implementation) to "what X should produce" (specification). Evals already describe the "what" — eventually the eval IS the skill.
Verification
After creating a skill:
- Confirm it appears in skill count
- Confirm skill-routing can find it with relevant keywords
- Confirm the description fits in ~70 tokens
- Test with a simulated trigger to verify steps are clear
- Write at least 2 evals per skill
- Run benchmark to establish baseline metrics
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: 469k | 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: 321k | 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: 321k | 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: 321k | Updated | Claude CodeCodex | Verified | Install |