# Triage Rules > Two-axis triage protocol for any AI agent. When facing a batch of findings, bug reports, audit results, or todo items — MUST sort by (confidence × blast-radius) before proposing actions. Confident + reversible/local damage = self-execute (Lane A). Confident + irreversible/colossal = recommend-and-ask (Lane B1). Uncertain regardless of risk = explain options and defer (Lane B2). Output as a scannable decision board, not a wall of text. Activates whenever an agent has just produced or received N≥2 findings/proposals/todos and is about to suggest actions. Also enforces a YAGNI self-check on items that introduce new mechanism, defensive code, future abstraction, or replace working systems. - Skill: `motiful/triage-rules` (Agent Skill) - Install (CLI): `npx skillmds@latest add motiful/triage-rules` - Raw SKILL.md: https://api.skillmd.com/api/skills/motiful/triage-rules/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: motiful (https://skillmd.com/u/motiful) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/motiful/triage-rules --- # Triage — Two-Axis Decision Lanes A `*-rules` hard-constraint skill (loaded by the rules-as-skills meta protocol). MUST/NEVER statements are binding, not advisory. ## Trigger Activates the moment any of these is true: - Just finished an audit, survey, code-walk, or research phase that produced ≥2 findings - Received a batch of bug reports, issues, or todo items from the user - About to propose actions on a list of items - About to introduce a new mechanism, hook, dependency, configuration field, or abstraction - About to add defensive code "just in case" or "for future flexibility" - About to replace a working mechanism with a more elegant but more complex alternative If the agent is about to write a multi-paragraph proposal touching ≥2 distinct items, this skill MUST run first. ## MUST - **MUST classify EACH item by two axes BEFORE writing any proposal:** - **Confidence** — how sure are you in the diagnosis / direction? (高 / 中 / 低) - **Blast radius** — if you act and you are wrong, what breaks? (reversible / local damage / colossal-or-irreversible) - **MUST output a decision board** (短表格 / 短列表 with one line per item), NOT a long narrative. - **MUST place each item in exactly one lane:** - **Lane A** — auto-execute (high confidence + low blast radius) - **Lane B1** — recommend & escalate (high confidence + high blast radius / irreversible) - **Lane B2** — discuss with options (low confidence regardless of blast radius) - **MUST give Lane B1/B2 items concrete options + tradeoffs in plain language.** No undefined jargon. Rank options when possible. State your recommendation explicitly. - **MUST tell the user what is in each lane** before executing anything, so they have ~30 seconds to redirect a misclassified item. - **MUST run a YAGNI self-check on any item that:** - Introduces a new mechanism (hook, socket, IPC field, config option, parameter) - Adds defensive code "in case X happens" - Adds an abstraction "we might need later" - Replaces a working mechanism with a more elegant but more complex alternative The single question: **"Would a real user fail RIGHT NOW without this? Or am I anticipating a problem that hasn't happened?"** - Answer YES → ship it. - Answer NO or "not sure" → backlog it, do not ship. - YAGNI failure with high confidence + low risk → reflect, drop it. - YAGNI failure with uncertainty or high risk → escalate to user. ## NEVER - **NEVER batch-execute Lane A items without first showing the user the triage board.** The board is the contract; without it the user cannot redirect a misclassification. - **NEVER let a B1/B2 item slip into Lane A "because it seemed easy".** Easy ≠ low blast radius. If the action touches shared state, public APIs, irreversible operations, or multi-module behavior, it is B1 even if you are 100% confident. - **NEVER produce a wall-of-text audit when a triage board would convey the same information.** Length is not depth — it is friction. - **NEVER use jargon in B1/B2 explanations without inline definitions.** If the user has to look up a term to make a decision, the explanation is broken. - **NEVER conflate "I am confident" with "this is safe to do".** These are two axes, not one. A 100%-correct destructive action is still high-blast-radius. - **NEVER add new mechanism / hook / abstraction without YAGNI passing.** "Sounds clever" is not the same as "user fails without it". The monkey-paw failure mode begins here. - **NEVER skip the triage step when the user gave a one-line task that contains hidden batches** ("clean up the progress dir", "fix the bugs in this module" — these are batches in disguise). ## How to Apply When the trigger condition fires: 1. **Freeze before proposing.** No "let me suggest..." until the board is written. 2. For each finding/item, ask: - (a) How confident am I in the diagnosis or direction? (高/中/低) - (b) If I act and I am wrong, what breaks and is it recoverable? (reversible / local damage / colossal) 3. Run the YAGNI self-check on any item introducing new mechanism / defensive / future / replacement. 4. Sort into Lane A / B1 / B2. 5. Write the board: each item one or two lines, no jargon, in a table or short list. 6. Hand to the user. Wait for sign-off on Lane A items in batch + targeted decisions on B1/B2. 7. Execute only what the user has authorized. ## Example board format ``` | 项 | 描述 | Lane | 备注 | |----|------|------|------| | 1 | xxx | A | 自信高 + 改 1 个文件 | | 2 | yyy | B1 | 自信高 + 删 21 个 worktree（不可逆）— 推荐 dry-run 先过目 | | 3 | zzz | B2 | 不确定 root cause；选项：(a) ... (b) ...，倾向 (a) | ``` Then explicitly: - "Lane A 我准备直接干（项 1）" - "Lane B1 等你拍板（项 2）" - "Lane B2 等你选（项 3）" ## Why **Past failure (booth project, 2026-05-07)**: An agent doing a long audit dumped a 7-file 5000-word report. The user could not read it, could not make decisions, and the audit produced zero downstream value. The same agent, switched to triage-board mode mid-session, produced a 1-page table — the user scanned it in 60 seconds, signed off Lane A items in batch, and had focused discussion on Lane B items. Same depth of work, 10× the delivered value. **Past failure (same project, same day)**: Agent proposed three "hook chemistry reactions" (PostToolUse / SessionStart / UserPromptSubmit) as elegant replacements for working polling code. None of them survived a YAGNI self-check on close inspection — one (UserPromptSubmit) was even based on a misunderstanding of hook semantics. All three were retracted. The lesson: "sounds clever" is the precursor to monkey-paw bugs. YAGNI gates this. The lesson: AI agents are good at producing exhaustive content. Users are bandwidth-limited. Triage is the bandwidth-matching protocol. YAGNI is the trap-avoidance protocol. Both bind by default; opting out requires an explicit reason. ## Relationship to other rules - **booth-dj-rules** (project-specific) — manages booth bug-handling flow. Compatible: that skill governs research-first; this skill governs the post-research bucketing. - **rules-as-skills meta** (machine-wide) — auto-loads any `*-rules` skill including this one as MUST-level. - Pairs with the `booth-dj` Failure Modes catalog (the catalog names the patterns, this skill prevents them at proposal time).