| name | add-technique-from-reference |
| description | Two modes: (1) failure-mode insight → new satirical technique, (2) production disaster → new entry in Real-World AI Failures. MUST trigger when the user provides an external URL and says "add this," "收录," "加个技法," "add technique," "add to README," etc. Also force-trigger for ANY link about AI coding agents, harness engineering, prompt engineering, LLM safety, agent-caused production outages, vibe coding failures, or AI agent security — even without an explicit "add" request. Core workflow: fetch → check relevance → deduplicate → draft → review with user (output draft, ask for feedback) → insert → verify.
|
-
Create a todo list (4–5 high-level steps). Do NOT create granular subtasks.
-
Fetch the URL. Use anysearch extract, web_fetch, or curl fallback. Read the full content.
-
Check relevance. If the article's topic is clearly AI coding agents / harness failures / production incidents, proceed. If it's unrelated (general tech news, non-AI topics), ask the user with ask_user_question whether to proceed. Do NOT silently skip or force-fit.
-
Deduplicate. Read all existing ### N. techniques in README.md. For each insight: (a) same core joke as an existing technique? Stop — enrich or skip. (b) same failure mode, different example? Stop — existing technique owns this. When uncertain, ask the user with ask_user_question. If nothing is novel, add the reference to references/index.md and stop.
-
Draft both EN and ZH versions. Follow the technique structure and writing style in .agent/references/writing-style.md. Write English first, then translate to Chinese.
-
Review gate — do NOT edit README files yet. Output both drafts to the conversation. Use ask_user_question (header: "Review draft", multiSelect: true) with these modification directions:
- "Section leads feel flat" — replace 首先/其次/然后 or First/Second/Then with vivid conversational leads
- "Strengthen the epigraph" — make the opening quote sharper, more quotable
- "Ending is formulaic" — replace the "说真的…" template with a story beat (timeline/scene/reveal)
- "Deepen the absurdity" — push a weak paragraph further with more concrete, absurd detail
- "'When caught' excuse needs layers" — make it a multi-layered corporate deflection cascade
- "Chinese translation needs work" — tone, epigraph, or 被发现了就说 doesn't match EN
- "Looks good — insert as-is"
Apply revisions, ask again. Loop until user approves. Only then proceed to insertion.
-
After approval: add reference to references/index.md. Insert technique into both READMEs following AGENTS.md cross-reference rules (insertion point, count update, layer assignment, defense objectives). Verify grep -c "^### [0-9]" README.md README-zh.md returns identical counts.
Writing reference
Technique structure, writing style rules, Chinese translation conventions, and anti-patterns are in .agent/references/writing-style.md. Read it when drafting.
Insertion Rules
Follow AGENTS.md § Cross-Reference Rules § A (insertion point, count update, layer assignment, defense objectives). Verify with grep -c "^### [0-9]" README.md README-zh.md.