Auto Scope

Auto-Scope

Before diving into a substantive coding task, name the files that matter and the ones that don't — so the work doesn't open by reading the whole repo to orient.

The output is an in-context scope brief, not a file on disk and not a lock on the filesystem. Any file can still be read later; the brief steers the default, it doesn't enforce it. The win is that 5–15 named files replace an exploratory sprawl.

1. Self-gate — decide whether to fire at all

Run only when the task plausibly touches multiple files or an area not seen this session.

Skip silently — emit nothing, not even a note that you skipped — when the task is any of:

• a single-file or one-line edit
• a question, explanation, or review with no code change
• continued work in an area already scoped earlier this session

If you're skipping, say nothing about scoping and proceed normally. A "I decided not to scope this" line is itself noise.

2. Discover — use whatever search the environment has

Detect what's available before committing to a search method, then take the path that fits:

• grepika (mcp__*grepika* tools) present → use its search / refs / outline for ranked, snippet-level discovery.
• graphify (a graphify-out/ dir or the graphify skill) present → query the graph for structure and relationships.
• Neither → fall back to built-in Grep / Glob.

State in one line which method you used, so the choice is visible and correctable.

Whatever the tool, the discovery goal is the same — find:

• the entry point(s) for the feature or bug
• direct callers and callees of the symbols the change will touch
• the test(s) covering that area
• the config / schema / build files the change implies

Don't exhaustively crawl. Stop when the named set is enough to start; the brief's OPEN section carries the rest.

3. Degrade gracefully

If discovery can't complete — no search tool resolves the area, the codebase is unfamiliar, or the task description is too vague to localize — do not block and do not guess silently. Emit a partial brief: list what you did find, mark it partial, and put what you couldn't localize in OPEN. A half-scoped task started honestly beats a fully-scoped one built on guesses.

4. Emit the brief — fixed shape

Output exactly this block, nothing reformatted into prose:

## Scope: <task in one line>
Discovery: <grepika | graphify | grep/glob> [+ "partial" if incomplete]

IN — read / edit:
  - path/to/a.ext        (why: entry point)
  - path/to/b.ext        (why: caller)
  - path/to/b.test.ext   (why: coverage)

OUT — leave alone:
  - vendor/, build/, generated output
  - <unrelated module or area the task does NOT touch>

OPEN — unresolved scope:
  - <question, with the assumption you're proceeding on>

Rules for the block:

• IN holds the working set — keep it to the files the task actually needs, not everything tangentially related. If IN exceeds ~15 entries, the task likely needs decomposition; say so.
• Every IN entry carries a one-clause why.
• OUT names concrete areas worth explicitly excluding (the ones an unscoped pass would waste reads on), not a generic "everything else."
• OPEN states each open question with the assumption you'll proceed under, so silence from the user means the assumption stands.

5. Continue

After emitting the brief, proceed straight into the task against that scope — no confirmation gate. The brief stays on screen; if it mis-scoped, the user's next message corrects it. Treat a correction as authoritative and re-scope from it.

Boundaries

• This skill scopes file relevance. It does not break the task into sub-tasks (that's planning) and does not edit code itself.
• It produces no disk artifact. If a task is large enough to need a durable, reviewable scope across a /clear or restart, that's a different request — flag it rather than silently writing a file.