| name | clarify-scope |
| description | Decide whether to ask before coding or proceed with stated assumptions. Small/clear tasks → do it. Large/ambiguous tasks → state assumptions and ask at most 1 critical question. Run at the start of any task. |
Clarify Scope
Before starting a task, apply the following decision tree.
Decision tree
Small, clear task → do it immediately
Signals:
- Know exactly which files need to be touched
- Goal cannot be interpreted in multiple ways
- Scope does not extend beyond 1 module
Action: do not ask, proceed immediately.
Large or ambiguous task → state assumptions + ask at most 1 question
Signals of a large/ambiguous task:
- Goal can be interpreted in multiple ways
- Touches > 1 module or > 1 layer
- Success criteria are unclear
- May affect existing behavior
Response format:
Assumptions:
- [what I assume about the scope/approach]
- [what I assume about constraints]
- ...
→ Proceeding with this approach. Correct me if anything is wrong.
If there is 1 thing that, if wrong, would send the work in a completely wrong direction — ask that one thing, nothing else.
"Worth asking" vs "not worth asking"
Worth asking (if wrong, heads in the opposite direction):
- "Is this a new endpoint or modifying an existing one?"
- "Is this feature for all users or only admins?"
- "Does this need to be backward-compatible or is a breaking change allowed?"
Not worth asking (if wrong, only requires a small fix):
- "What should the variable be named?"
- "What should the error message say?"
- "Should I use
const or let?"
Examples
Task: "Fix bug where login doesn't redirect correctly"
→ Small, clear → do it immediately, no questions.
Task: "Refactor the auth system"
→ Large, ambiguous. Response:
Assumptions:
- Keep the existing API surface (backward-compatible)
- Only refactor internal logic, no DB schema changes
- Scope: src/auth/ and files that call into it
→ Proceeding with this approach. Correct me if anything is wrong.
Single question: are there any active sessions on production that need to be migrated, or is it acceptable to invalidate all of them?
