mega-eval-durability
Phase 1F only: AI/agent durability audit from an Evaluation Brief + primary HTTPS URL, output phase1f-durability-raw.md. Use for 'mega eval durability audit', 'AI resilience for eval', or Phase 1F without the full mega-eval pipeline.
- License 许可
- MIT
- 最近更新
- 2026-10-11
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- mega-eval-durability
- description
- Phase 1F only: AI/agent durability audit from an Evaluation Brief + primary HTTPS URL, output phase1f-durability-raw.md. Use for 'mega eval durability audit', 'AI resilience for eval', or Phase 1F without the full mega-eval pipeline.
03 SKILL.md 正文
Mega-eval — Phase 1F: AI durability audit only
You are running only Phase 1F of the mega-eval pipeline: a report-only audit of how resilient the subject’s AI / agent surface is to model, provider, and capability change. This is not a general business or platform-risk review.
Shared methodology location
Resolve references/ per skills/README.md. Read: references/durability-audit-template.md, references/subagent-prompts.md (Prompt ID phase1f.durability.v1), and references/pipeline-checklist.md (Phase 1 / Phase 2 expectations).
Full orchestration rules: Phase 0 (durability audit gate), Phase 1F, and Phase 2 in full SKILL.md (../../SKILL.md or ../mega-eval/SKILL.md).
Inputs and outputs
| Reads | eval-brief.md or pasted Evaluation Brief; shared primary https:// URL from the brief’s audit section or user message |
| Writes | <workspace>/phase1f-durability-raw.md |
Opt out: If MEGA_EVAL_DURABILITY_AUDIT is off / 0 / false, do not run — tell the user to enable it or use full mega-eval with the env unset.
If the subject has no meaningful AI / LLM / agent surface, write the short N/A stub described in references/durability-audit-template.md and stop cleanly.
Use run_id and run_log paths in the subagent prompt per references/subagent-prompts.md. If logging is enabled, follow the full SKILL.md run-log rules for Phase 1.
Next steps
Run mega-eval-synthesis once Phase 1A–1C are complete; it merges phase1f-durability-raw.md when present and omits the exec-summary posture section when the risk band is N/A. Use mega-eval for the full pipeline.
04 同仓库更多 Skills
mega-eval-synthesis
Phase 2 only: synthesize phase1a/1b/1c raw outputs into phase2-synthesis.md (critical fixes, design issues, next steps); merges phase1d-design-raw.md when present. Use when Phase 1 is done and you want 'mega eval synthesis only', 'merge critique competitive strengths', or judgment-heavy synthesis without Phase 3–4.
mega-eval-strengths
Phase 1C only: strengths and opportunities analysis from an Evaluation Brief, output phase1c-strengths-raw.md. Use for 'mega eval strengths only', 'upside analysis', 'growth angles without critique track', or Phase 1C standalone.
mega-eval-security
Phase 1E only: observation-only security posture audit from an Evaluation Brief + primary HTTPS URL, output phase1e-security-raw.md. Use for 'mega eval security audit', 'security posture for eval', or Phase 1E without the full mega-eval pipeline.
mega-eval-prompt-derivation
Decompose any user request into a structured Prompt Derivation Engine spec — a JSON object that captures intent, task type, context, output contract, instructions, eval checks, and execution plan. Use this skill when you need to convert a vague or complex request into a precise, schema-validated prompt blueprint before execution. Trigger on: 'derive a prompt for', 'prompt spec', 'decompose this request', 'prompt derivation', 'structured prompt', 'prompt blueprint', or when an upstream orchestrator needs a machine-readable task breakdown.
mega-eval-hater
Phase 1A only: 12-persona critical feedback (hater mode) from an Evaluation Brief, output phase1a-hater-raw.md. Use for 'mega eval hater only', 'roast this product', '12 lens critique', 'adversarial feedback without full eval', or competitive run of Phase 1A alone.
mega-eval-design
Phase 1D only: optional live-site design audit from an Evaluation Brief + primary HTTPS URL, output phase1d-design-raw.md. Report-only (no code fixes). Use for 'mega eval design audit', 'UX teardown of this URL', 'visual audit for eval', or Phase 1D without full mega-eval.