Ask — Context-Aware Q&A

Trigger

• Keywords: ask, quick question, context question, project question, 問一下, 想了解, 想知道

When NOT to Use

Scenario	Alternative
Code modification or implementation	/feature-dev
Code review or PR review	/codex-review-fast
Tech spec review	/review-spec
Document review	/codex-review-doc
Bug fixing	/bug-fix
Next step decision	/next-step
Deep multi-source research	/deep-research
Systematic code tracing	/code-explore

Procedure

Phase 0: Session Context Capture

Run these 4 commands in parallel to build session context:

#	Action	Tool
1	Current branch	Bash("git branch --show-current")
2	Feature detection	Bash("node scripts/resolve-feature.js") — the wrapper, not the CLI: it owns the failure payload, so a resolver that runs and fails — nonzero exit, signal, truncated write, off-contract payload — still yields the full shape with scan_error: true rather than {} or a truncated document. A missing node yields no JSON at all; treat that like any other unusable payload
2b	scan_error gate	scan_error !== false ⇒ the source sets are unknown, not empty — report and take the ⚠️ Need Human exit rather than answering from an empty set. Gate on !== false, not === true: a {} payload from a shell fallback has no such field, so a non-null key is not evidence the sets are complete
3	Changed files	Bash("git status --porcelain")
4	Recent commits	Bash("git log --oneline -5")

Untracked file fallback: If feature resolver returns key: null, derive feature from changed/untracked paths:

1. Parse git status --porcelain output for docs/features/<key>/ or skills/<key>/ patterns
2. Extract <key> as candidate feature
3. Use this for docs intent feature-first lookup when resolver fails

Phase 1: Intent Classification + Routing

1a. Conversation Context Integration

Before classifying intent, review prior conversation turns for:

• Active feature: feature being developed, files recently discussed or edited
• Ongoing task: what skill was last invoked, what phase we are in
• Implicit scope: if the user asks "why?" after a code change, the scope is that change

Use this context to disambiguate the question and select the right intent. See references/intent-patterns.md for edge cases.

1b. Intent Classification

Classify the question (LLM-inferred) into one or more intents:

Intent	Signal Examples	Context Actions
code	"function X 做什麼", file paths, module names	Grep → Read → trace 1 level
git	"最近改了什麼", "誰改的", "when"	git log / diff / blame
docs	"需求是什麼", "spec 寫了什麼"	Feature resolve → source set by question kind → fallback Glob
rules	"規則是什麼", "convention", "allowed"	Read rules/ files
skill	"有沒有 skill", "怎麼用 /X"	Glob skills/ → Read SKILL.md
arch	"系統架構", "整體設計"	CLAUDE.md + Explore agent
multi	Multiple intents mixed	Combine actions from each intent

1c. Skill Routing Check

Before gathering context, check if the question is action-oriented. See references/routing-table.md.

If a better skill is identified, suggest it: "這個問題更適合 /X，要改用嗎？" — do not auto-redirect.

Phase 2: Context Gathering

Execute per-intent tool call sequences. Hard limits apply.

code: Grep keywords (top 10 files) → Read most relevant (max 5) → trace imports (1 level)

git: git log --oneline -20 → git diff (if recent changes) → git blame (if specific lines)

docs: Resolve feature → pick the source set the question actually asks for → fallback Glob "docs/**/*.md" (top 5) → Read (max 3)

The question is…	Read	Not
"現在的行為是什麼"	code, rules/, current_authority	A tech spec — it records the design, not what shipped
"當初為什麼這樣設計"	design_records	—
"當初要求什麼 / 這張單結了嗎"	work_records	—
"這個決定是什麼時候做的"	history_records	—

Answering a current-behaviour question from a design record is the specific failure this split exists to prevent: the spec is the older artifact, so it reads as authoritative and is wrong. When a design record is the only source available, say which document the answer came from and that it may predate the code.

rules: Glob rules/*.md + .claude/rules/*.md → Grep keywords → Read + quote (max 3)

skill: Glob skills/*/SKILL.md → Grep keywords (top 5) → Read (max 3)

arch: Read CLAUDE.md + key entrypoints → dispatch Explore agent

multi: Combine steps from each intent. Parallel execution. Hard limit: max 8 file reads total.

Phase 3: Sub-Agent Dispatch (Optional)

Complexity	Criteria	Strategy
Simple	Single intent, clear target, < 5 files	Direct tools only (0 agents)
Medium	Multi-file, cross-module	1 Explore agent
Complex	Multi-intent, cross-cutting	2 agents parallel (hard max)

Dispatch when: Grep returns > 10 files across modules, or question involves architecture / cross-cutting concerns. Default: direct tool calls.

Phase 4: Answer Synthesis

Combine all gathered context into a structured answer. Follow the output format below.

Read-Only Enforcement

This skill is strictly read-only. The following git commands are prohibited:

git add | git commit | git push | git pull | git reset | git stash
git rebase | git merge | git checkout -- | git restore | git clean

allowed-tools does not include Edit, Write, or NotebookEdit.

Path Security

Control	Rule
Repo boundary	All Read/Glob within repo root (git rev-parse --show-toplevel)
Traversal rejection	Reject .. path segments, absolute paths outside repo, symlinks out of repo
Secret skip	Do not read .env, credentials.*, *secret* files
Output redaction	High-confidence secret patterns → [REDACTED]; medium-confidence → mask with 4 chars visible

Output Format

## Answer

{Direct, concise answer}

### Sources

| Type | Reference | Relevance |
|------|-----------|-----------|
| file | `path/file.js:42` | {why relevant} |
| commit | `abc1234 — message` | {why relevant} |
| command | `git log --oneline -5` | {result summary} |

### See Also

- `/code-explore` — for full trace
- {other relevant skill or doc}

Every claim must have at least one source evidence. Answer < 500 words unless user requests detail.

Verification

• Session context captured (branch, feature, changed files)
• Intent classified and context gathered per pipeline
• Source attribution present for all claims
• No Edit/Write/mutating git commands executed
• No secrets in output

References

• references/intent-patterns.md — Detailed intent examples and edge cases (read when classifying ambiguous questions)
• references/routing-table.md — Full skill routing decision table (read when checking if another skill is better)

Examples

Code Question

Input: /ask resolve-feature-cli.js 怎麼偵測 feature？
Phase 0: branch=feat/ask, feature=ask
Phase 1: Intent=code (file path mentioned)
Phase 2: Grep "resolve-feature" → Read scripts/lib/feature-resolver.js → trace exports
Phase 4: Answer with Sources (file evidence)

Git Question

Input: /ask 最近有什麼 commit？
Phase 0: branch=feat/ask
Phase 1: Intent=git ("最近", "commit")
Phase 2: git log --oneline -20
Phase 4: Answer with Sources (commit + command evidence)

Multi-Intent Question

Input: /ask auto-loop 的規則是什麼？有哪些 skill 會用到？
Phase 0: branch=feat/ask
Phase 1: Intent=multi (rules + skill)
Phase 2: Read rules/auto-loop.md + Grep "auto-loop" in skills/*/SKILL.md
Phase 4: Answer with Sources (file evidence from both rules and skills)