Record Replay Skill Miner

Record Replay Skill Miner

Use this skill when a task has been repeated enough that it should become durable Srednoff OS capability.

Workflow

1. Define the repeated job, trigger phrase, expected output, and definition of done.
2. Collect only the minimum evidence needed: Record & Replay notes, command transcript, diff, runbook, issue, PR, or automation memory.
3. Scrub secrets, tokens, cookies, private keys, personal data, paid-account details, and production identifiers before extracting patterns.
4. Separate stable procedure from one-off project details.
5. Choose the smallest durable surface:
   • AGENTS.md for project policy.
   • SKILL.md for reusable workflow.
   • scripts/ for deterministic fragile steps.
   • evals/ for regression fixtures.
   • hooks for opt-in enforcement.
   • automation memory for recurring task continuity.
6. Draft the skill with name and description only in frontmatter, a concise imperative workflow, checklist, and guardrails.
7. Add or refresh agents/openai.yaml when the skill should be discoverable in the app.
8. Add a validation path: prompt trigger check, quick skill validation, selector eval, script test, or manual replay scenario.
9. Reject the conversion if the workflow is still vague, low frequency, duplicate, unsafe, license-risky, or too project-specific.

Skill Quality Checklist

• Name is lowercase hyphen-case and under 64 characters.
• Description front-loads exactly when to use and not use the skill.
• Body is shorter than the repeated workflow it replaces.
• Scripts are included only when deterministic execution beats instructions.
• References are linked only when the details are too large for SKILL.md.
• Guardrails cover destructive, paid, production, security, legal, medical, financial, hiring, and privacy-sensitive actions when relevant.
• Validation is explicit enough for the next automation run to catch drift.

Output Shape

Return:

• proposed durable surface;
• skill or eval name;
• source evidence used;
• what was generalized;
• what was intentionally excluded;
• validation command or replay scenario.

Guardrails

• Do not store raw prompts, secrets, private logs, customer data, or full proprietary traces in a skill.
• Do not copy third-party workflow text or code without license review.
• Do not create broad prompt-spam skills; merge with an existing narrower skill when overlap is high.
• Do not add hooks that perform destructive or externally visible actions.