This applies regardless of how small the request seems. This skill runs UPSTREAM of brainstorming, planning, and TDD skills — invoke it first.
When to use this skill (TRIGGERING)
You MUST invoke this skill on ANY user message that could be one of these:
| Type | Chinese examples | English examples |
|---|---|---|
| Feature request | "add X", "add Y", "implement Z", "make one", "help me write" | "add X", "create Y", "build Z", "implement" |
| Modification | "change X", "modify", "adjust", "switch to", "change X to Y" | "change X", "modify", "switch to", "turn X into Y" |
| Optimization | "optimize", "improve", "enhance", "perfect", "better", "faster" | "optimize", "improve", "enhance", "better", "faster" |
| Refactor | "refactor", "rewrite", "clean up" | "refactor", "rewrite", "clean up" |
| Design question | "how should this be designed", "how about the architecture", "what's the approach", "what's the plan" | "how should we design", "what's the architecture", "how would you approach" |
| Multi-item asks | anything with "also", "as well as", "additionally", "at the same time" | anything with "also", "and", "plus" |
| Polite imperatives | "help me X", "can you X", "need X", "by the way X" | "can you X", "please X", "I need X" |
| Vague improvement | "could this be a bit better here", "this isn't quite right" | "this could be better", "something feels off" |
Red Flags — STOP, you're rationalizing
If you catch yourself thinking any of these, STOP and invoke the skill:
| Rationalization | Reality |
|---|---|
| "This is too small to need intent analysis" | Small tasks are where scope creep starts. 3 lines of analysis costs less than 3 rounds of rework. |
| "I already know what they mean" | Knowing ≠ outputting. User needs to see the deep goal to confirm or correct it. |
| "Let me just start, I can extrapolate mid-way" | Mid-way extrapolation = scattered, not batched. User pays cost twice. |
| "The request is very specific" | Specific surface ≠ specific deeper goal. Check Layer 3 anyway. |
| "They said quick change" | If they didn't prefix quick: or skip:, treat as normal. |
| "It's obvious from context" | Obvious to you ≠ obvious to them. The gate is for them, not you. |
| "This is a follow-up question in the same conversation" | See "Multi-turn behavior" below — may still need fresh analysis. |
When to skip this skill (HARD EXEMPTIONS)
Skip ONLY if the message is CLEARLY one of:
- Pure factual question — "what is X", "X is what", "what is X", "how does X work", ends in
?asking for explanation not change - Single-line typo/wording fix — "change a typo", "spelling is wrong", "fix a typo"
- Short reply / confirmation — "good", "right", "mm", "ok", "continue", "yes", "no", "no problem", "fine"
- User explicit override — message starts with
skip:,quick:,直接:,快速: - Pasted error/log for diagnosis — user pasted stack trace or error log, asking "what's this error" / "what's this error"
- Request to modify Claude's previous text reply (not code) — "reword the previous reply"
- Handoff / take-over, scope already defined — the message hands off / resumes work AND already gives task + order/boundaries + stop/go (decision gate) → follow it directly; do NOT re-run analysis or re-confirm scope (e.g. "take over Task 5 per handoff / continue the handoff / pick up task board #X"). ⚠️ Exception: DO run Steps 1-3 (scoped to just the new part) if the taker adds a NEW independent feature, changes a boundary, or hits an outcome-changing choice — a handoff exempts continuation, not new scope.
When skipping, answer directly without invoking this skill.
The protocol (MANDATORY when invoked — every step)
Step 1 — Three-layer intent analysis
Output this table before any other content:
## 意图分析
| 层级 | 内容 |
|------|------|
| 1. 表面需求 | [literal restatement of the request] |
| 2. 直接目的 | [the immediate outcome they want] |
| 3. 深层目标 | [the underlying goal/value they're optimizing for] |
If Layer 3 contradicts or suggests a better path than Layer 1, flag it explicitly with a "⚠️ deep goal and surface request may have a better path" note.
Step 2 — extrapolation (2-8 extrapolations)
Based on Layer 3 (deeper goal), propose 2-8 additional items the user did NOT ask for but that serve the same deeper goal.
Quantity rule: Stop when quality drops. 2 excellent extrapolations beat 6 padded ones.
## 举一反三
| # | 补充建议 | 为什么服务同一深层目标 | 推荐/可选 |
|---|---------|------------------------|-----------|
| 1 | ... | ... | 推荐 |
| 2 | ... | ... | ... |
Step 3 — STOP and get scope confirmation
Write the confirmation prompt in the same language as the user's message:
- User wrote in Chinese → "The above is the intent analysis and supplementary suggestions. Which ones do you want all done? Or just the original request?"
- User wrote in English → "Above is the intent analysis and extrapolations. Which items should I tackle — all, a subset, or just the original request?"
- Mixed/unclear → Chinese by default (per user's global CLAUDE.md)
Do not write code, do not make file edits until the user confirms scope.
Step 4 — Execute agreed scope in one batch
Once user confirms, execute all agreed items together. Don't re-ask for each item.
Multi-turn behavior
- First invocation in a thread — run full Steps 1-3, UNLESS it's an already-scoped handoff take-over (exemption 7 → follow the handoff directly).
- User adds a new independent ask mid-conversation ("add another X", "also add Y", "by the way do Z") — run full Steps 1-3 scoped to just the new item.
- User refines agreed scope ("remove 4", "swap the priorities of 1 and 3") — NO re-trigger. Just confirm and execute.
- User says "continue" / "go ahead" during Step 4 execution — NO re-trigger. Continue executing the batch.
- New session / post-compaction — treat as fresh thread, re-trigger on first qualifying message.
- Pure debug/test of already-agreed feature — NO re-trigger. Just work on the bug.
Thinking dimensions for extrapolation
When generating extrapolations, scan these dimensions:
- Other similar issues on the same page/module — change one button → should other buttons in the same module also be changed
- Friction points before and after in the user journey — add a login page → signup/forgot password/session expiry
- Cross-module consistency — frontend changes a field name → backend/DB/API docs/tests
- Edge cases — empty data / errors / insufficient permissions / network failure
- Efficiency multiplier — can repetitive tasks be automated/batched/templated
- UX details — loading / error messages / confirmation dialogs / keyboard shortcuts
- Data and state — persistence / cache invalidation / concurrency / migration
- Observability — Logging / error reporting / performance monitoring / event tracking
- Performance & Scale — N+1 / batch / pagination / caching
Anti-patterns (DO NOT)
- ❌ Writing code or editing files before outputting Step 1-2
- ❌ Skipping "draw inferences from one instance" and only restating the surface request
- ❌ Proposing extrapolations unrelated to Layer 3 (scattered suggestions)
- ❌ Padding to 8 items when only 2-3 are genuinely relevant (quality > quantity)
- ❌ Asking one-by-one about each extrapolation (batch the decision instead)
- ❌ Jumping to Step 4 without user confirming scope
- ❌ Running brainstorming/plan/tdd skills before this one (those are downstream)
Concrete examples
✅ Correct activation — "Add a batch order export feature to the e-commerce admin backend"
## 意图分析
| 层级 | 内容 |
|------|------|
| 1. 表面需求 | 电商后台加"批量导出订单"功能 |
| 2. 直接目的 | 一次性下载多个订单,避免逐个操作 |
| 3. 深层目标 | 让运营/财务能高效做数据分析和对账 |
## 举一反三
| # | 补充建议 | 为什么 | 推荐 |
|---|---|---|---|
| 1 | 按日期/状态/渠道筛选后导出 | 深层目标是"高效分析",筛选是前提 | 推荐 |
| 2 | 导出格式支持 Excel/CSV/JSON | 下游工具需求不同 | 推荐 |
| 3 | 大数据量异步任务 + 飞书通知 | 避免前端阻塞 | 推荐 |
| 4 | 导出记录存档(谁/何时/何字段)| 合规审计 | 可选 |
| 5 | 保存导出模板 | 重复导出零配置 | 可选 |
| 6 | 敏感字段脱敏选项 | 外部分享安全 | 推荐 |
以上是意图分析和补充建议。要全做哪几条?还是只做原需求(基础批量导出)?
✅ Correct activation — "help me optimize this query performance"
## Intent Analysis
| Layer | Content |
|------|------|
| 1. Surface | Optimize the query's performance |
| 2. Direct | Make the query run faster |
| 3. Deeper | Keep the overall page/API responsive under production load |
## Extrapolations
| # | Suggestion | Why it serves the deeper goal | Rec |
|---|---|---|---|
| 1 | Add DB index on filtered columns | Root cause of slow query | Recommended |
| 2 | Cache hot query results (Redis / in-memory) | Page responsiveness, not just this query | Recommended |
| 3 | Add `EXPLAIN ANALYZE` regression test | Prevent regression on future schema changes | Recommended |
| 4 | Monitor p95/p99 query latency | Responsiveness is a distribution, not an average | Optional |
Above is the intent analysis and extrapolations. Which items should I tackle — all, a subset, or just the original request?
✅ Correct skip — "What's the difference between React 19 and 18"
Answer directly. Do NOT invoke this skill.
✅ Correct skip — "Fix the typo on line 23"
Fix the typo. Do NOT invoke this skill.
✅ Correct skip — handoff take-over (scope already defined)
User pastes a handoff prompt with task + boundaries + stop/go (decision gate) already given → execute directly, no intent tables, no scope re-confirmation. (Trigger only if a NEW feature or boundary change then appears.)
❌ Wrong — jumping to code
User: "Add a batch export feature to the e-commerce admin backend" Wrong response: immediately writes code for basic export Correct: Invoke this skill, do Steps 1-3, wait for confirmation.
❌ Wrong — unrelated extrapolations
User: "Optimize the login page's loading speed" Wrong extrapolations: "Also add a registration page/user profile page while you're at it" (unrelated to the deeper goal "loading speed") Correct extrapolations: "image lazy loading / code splitting / SSR pre-rendering / critical resource preload" (all serve "loading speed")
Rationale (why this matters)
From real-team user feedback (TikTok / e-commerce collaboration practice):
"Users consider this the most efficient way to collaborate — a single discussion can identify all optimization points around the same goal, avoiding multiple back-and-forth rounds."
Missing this skill causes:
- User must keep coming back with follow-up requests
- The agent implements narrow surface request, missing obvious improvements
- Scope creeps across multiple rounds
- Wastes user's time and context budget
Self-check before responding
Before ANY response to a user message, ask yourself:
"Is this a feature/modification/optimization/refactor/design question/multi-item ask?"
- YES → Invoke this skill. Output Step 1-3. STOP.
- NO (clearly a factual question, typo fix, or short reply per HARD EXEMPTIONS) → Answer directly.
- UNSURE → Default to invoking this skill. See
<EXTREMELY-IMPORTANT>at top.