| name | scope |
| description | Fast parallel change-scope analyzer. Launches 5+ agents concurrently to identify
what files need changing, what could break, and the recommended approach.
Lighter than /specify โ no interview, no spec.json, just a quick scoped report.
Use this skill whenever the user wants to understand the blast radius of a change
before diving into implementation. Also use when the user has a bug or feature
request and wants to know where to look and what to touch.
Trigger phrases: "/scope", "scope this", "impact analysis", "change scope",
"what needs to change", "blast radius", "what would break",
Korean: "๋ณ๊ฒฝ๋ฒ์", "์ค์ฝํ ๋ถ์", "๋ญ ๋ฐ๊ฟ์ผ ํด", "์ด๋๋ฅผ ๊ณ ์ณ์ผ ํด",
"์ํฅ๋ฒ์", "์ด๋ ๊ฑด๋๋ ค์ผ ๋ผ", "๋ญ๊ฐ ๊นจ์ง ์ ์์ด"
|
| allowed_tools | ["Read","Grep","Glob","Bash","Agent","Skill"] |
| validate_prompt | Must produce a Scope Analysis report containing all 6 sections:
1. Change Map (table with File/Module, Change Type, Risk, Reason)
2. Impact Radius (Direct, Indirect, Tests, Docs)
3. Risk Summary (HIGH/MED/LOW counts)
4. Gaps & Warnings
5. Recommended Approach (with suggested order)
6. Must NOT Do
Must include a Confidence tag (HIGH/MED/LOW with N/6 agents completed).
Report must be under 80 lines. All file paths must be absolute or repo-relative.
Must NOT: generate spec.json, start implementing changes, interview beyond one question.
Must have handled Step 0 prerequisites (empty project, scope size, input validation).
|
/scope โ Parallel Change-Scope Analyzer
Analyze a requirement and produce a change-scope report by running multiple analysis
agents in parallel. Two rounds of concurrent execution, then synthesis.
The whole point is speed: instead of sequential analysis that takes 5 minutes,
launch everything at once and get results in ~1 minute.
Step 0: Prerequisites and Early Exits
Before launching any agents, verify the environment is analyzable.
0a. Check for source code
if project root contains no source files (no .ts, .js, .py, .go, .rs, .java, .sh, etc.):
Output: "No source code found in this project. /scope requires an existing
codebase to analyze. Create some code first, then run /scope again."
STOP โ do not proceed to Phase 1.
0b. Scope size gate
Perform a quick keyword search (Grep/Glob) to estimate how many files/modules the
requirement touches. If the estimate exceeds 20 files across 5+ directories:
Output: "This change appears to affect N files across M directories.
/scope is designed for focused changes. For large-scale rewrites, consider:
- /specify โ full spec generation with interview and structured planning
- Breaking this into smaller scoped changes and running /scope on each"
Ask the user whether to proceed anyway or switch to /specify. If they choose to
proceed, continue but add a Confidence: LOW (broad scope) tag to the final report.
0c. Validate input
The user provides a requirement, bug description, or feature request. Examples:
- "I want to refactor the hook system"
- "Add auto-fix feature to the check skill"
- "spec.json validation is too slow"
Ambiguous Input Recovery: If the input is too vague to search for (e.g., "make it
better", "improve things", "fix it"), attempt auto-detection before asking:
- Check
git diff HEAD โ if non-empty, offer to scope the uncommitted changes
- Check
git log --oneline -5 โ if recent commits suggest a theme, offer that
- If still unclear, ask exactly ONE clarifying question with specific options
(e.g., "Did you mean: (a) the auth module, (b) the API layer, (c) something else?")
Do not loop โ one auto-detect attempt, then one question at most, then proceed.
Phase 1: State Discovery
Two steps: first parallel agents, then /check skill.
Step 1A: Parallel Agents (3 concurrent)
Launch all three agents in a single message so they run concurrently.
Agent 1: code-explorer
Find all code directly related to the requirement.
Prompt template:
"Find all code related to: {requirement}
Focus on: entry points, core logic, data flow, and test files.
Project root: {project_root}"
Agent 2: docs-researcher
Find internal documentation, ADRs, and conventions relevant to the change.
Prompt template:
"Find internal documentation relevant to: {requirement}
Look for: architecture decisions, conventions, constraints, related past changes.
Project root: {project_root}"
Agent 3: code-explorer (git state)
Understand current git state and recent changes in the affected area.
Prompt template:
"Analyze the current git state for areas related to: {requirement}
1. Run: git log --oneline -10 for recent changes in relevant areas
2. Run: git diff --name-only to find uncommitted changes
3. Identify any in-progress work that might interact
Project root: {project_root}"
Implementation: Use three parallel Agent tool calls with subagent_type set to
code-explorer for Agents 1 and 3, and docs-researcher for Agent 2.
Step 1B: /check Skill
After the three parallel agents complete, invoke the /check skill via Skill tool.
This runs the full rule-based verification against .dev/rules/ to surface any
cascading change requirements or rule violations in the affected area.
Skill("check")
Wait for /check to complete. Its PASS/WARN results feed into Phase 2 as
check_results.
Skip condition: If .dev/rules/ does not exist, skip Step 1B and proceed
directly to Phase 2 (note "no rules configured" in the report).
Phase 2: Deep Analysis (Parallel)
After Phase 1 (Step 1A + 1B) completes, launch three more agents in a single
message. Inject Phase 1 findings as context into each prompt.
Agent 4: gap-analyzer
Prompt template:
"Analyze gaps for this change:
Requirement: {requirement}
Code findings:
{agent_1_results}
Documentation findings:
{agent_2_results}
Current state:
{agent_3_results}
Check results (rule violations/warnings):
{check_results}
Focus on: missing requirements, edge cases, things that must NOT be changed."
Agent 5: tradeoff-analyzer
Prompt template:
"Analyze tradeoffs for this change:
Requirement: {requirement}
Code findings:
{agent_1_results}
Focus on: risk per file/module, simpler alternatives, over-engineering warnings.
Skip decision_point YAML โ just give the risk table and alternatives."
Agent 6: codex-strategist
Prompt template:
"Synthesize a strategic view of this change:
Requirement: {requirement}
Code findings:
{agent_1_results}
Documentation context:
{agent_2_results}
Gap analysis:
{agent_4_results โ if available, otherwise pass Phase 1 summary}
Focus on: blind spots, cross-cutting concerns, architectural fit."
Note on Phase 2 ordering: Agents 4, 5, and 6 can all launch together because
they each receive Phase 1 results directly. Agent 6 (codex-strategist) works with
Phase 1 results โ it does not need to wait for Agent 4's output. The codex-strategist
provides independent strategic synthesis, not a review of the gap analysis.
Phase 3: Synthesis
After all agents complete, synthesize results into a single report.
Do this yourself โ no subagent needed.
Output Format
## Scope Analysis: {requirement_title}
**Confidence: {HIGH|MED|LOW}** โ {N}/6 agents completed
### 1. Change Map
| File/Module | Change Type | Risk | Reason |
|-------------|-------------|------|--------|
| path/to/file.ts | MODIFY | LOW | {why} |
| path/to/other.ts | MODIFY | HIGH | {why} |
| path/to/new.ts | CREATE | MED | {why} |
### 2. Impact Radius
- **Direct**: {files that must change}
- **Indirect**: {files that might need adjustment}
- **Tests**: {test files to update/create}
- **Docs**: {documentation to update}
### 3. Risk Summary
| Risk | Count | Key Items |
|------|-------|-----------|
| HIGH | N | {list} |
| MED | N | {list} |
| LOW | N | {list} |
### 4. Gaps & Warnings
- {gap or warning from gap-analyzer}
- {over-engineering warning from tradeoff-analyzer}
- {blind spot from codex-strategist}
### 5. Recommended Approach
{1-3 sentences: the simplest path to implement this change}
**Suggested order**:
1. {first thing to do}
2. {second thing to do}
3. {third thing to do}
### 6. Must NOT Do
- {prohibition from gap-analyzer}
- {prohibition from tradeoff-analyzer}
Output Rules
- Every file path must be absolute or repo-relative (no vague references)
- Risk levels must be justified (not arbitrary)
- Keep the report under 80 lines โ brevity is the point
- Always include the Confidence tag (see Error Handling section)
- If any agent was SKIPPED, note which agent and why in section 4 (Gaps & Warnings)
- Do not rate risk for areas only covered by failed agents โ mark as "UNKNOWN (agent N skipped)"
When to Use /scope vs Other Skills
| Skill | Purpose | Output | When to use |
|---|
| /scope | Quick blast-radius analysis | Markdown report (Change Map + Risk) | Before starting work, to understand what to touch |
| /bugfix | Root-cause diagnosis + fix | spec.json + executed fix | When a bug needs to be fixed, not just analyzed |
| /check | Rule-based verification | PASS/WARN per rule | After changes, to verify compliance |
| /specify | Full spec generation | spec.json v5 with interview | When building a feature from requirements |
Error Handling
Agent Failure Tiers (N-of-M)
/scope launches 6 agents across two phases. Not all need to succeed for a useful report.
| Scenario | Action | Confidence |
|---|
| 1 agent fails | Mark that agent as SKIPPED in the report. Synthesize from remaining 5. Note: "Agent N (role): SKIPPED โ [reason]" in section 4. | HIGH (5/6 coverage) |
| 2 agents fail (same phase) | Synthesize from remaining agents. Add confidence warning: "Partial analysis โ [missing perspectives] were not evaluated." | MED (4/6 coverage) |
| 2 agents fail (cross-phase) | Same as above but note which phase was degraded. | MED (4/6 coverage) |
| 3+ agents fail | Produce a partial report with available results only. Prepend: "LOW CONFIDENCE โ only N/6 agents returned results. Consider re-running or using /specify for thorough analysis." Do NOT issue risk ratings for areas covered by failed agents. | LOW (<=3/6 coverage) |
| All Phase 1 agents fail | Cannot proceed to Phase 2 (no context to inject). Report the failure and suggest retry or manual exploration. | NONE โ abort |
| All agents fail | Report the failure. Suggest the user retry or use a different approach. | NONE โ abort |
When an agent is SKIPPED, treat its coverage area as unknown risk in the synthesis.
Do not assume the area is safe just because no agent flagged it.
Confidence Tag
Every report must include a confidence tag after the title:
## Scope Analysis: {requirement_title}
**Confidence: {HIGH|MED|LOW}** โ {N}/6 agents completed{, broad scope | if applicable}
Confidence is the minimum of:
- Agent coverage: HIGH (5-6/6), MED (3-4/6), LOW (1-2/6)
- Scope breadth: if Step 0b flagged broad scope, cap at LOW
Other Error Cases
- No relevant code found: If code-explorer finds nothing, report "No directly related code found" in the Change Map and flag it as a HIGH risk gap.
.dev/rules/ missing: Skip Step 1B entirely and note "No rules configured โ /check skipped" in section 4.
- Vague input: Follow the Ambiguous Input Recovery procedure in Step 0c.
Constraints
- Do NOT generate spec.json โ this is a lighter tool than /specify
- Do NOT start implementing changes โ only analyze
- Do NOT interview the user beyond one clarifying question
- Do NOT run agents sequentially when they can run in parallel
- Maximum wall-clock time target: ~90 seconds (two parallel rounds)