ai-explore git:20260901.deb68d3A

Immutable. This exact content is served forever at /api/v1/blob/93a1dfe9c641bc2a.

---
name: ai-explore
description: >-
  Answers questions about this repository by reading it, anchored to file:line, and tours
  an unfamiliar area for somebody who has just arrived. Trigger for "where does X live",
  "how does this work", "why does it do that", "what depends on Y", "walk me through",
  "map this module", "trace this import chain", "onboard me". Not for evidence from
  outside this repository — use /ai-research. Not for diagnosing a failure — use
  /ai-debug. Not for judging a diff — use /ai-verify.
license: Apache-2.0
---

# Read this repository and answer

## What it produces

An answer anchored to real paths. Every claim points at `file:line`, or it is marked as a
guess.

## Steps

1. Match the depth to the question, not to a flag. "Where is X" gets a path and a sentence.
   "How does this work" gets the flow. "Onboard me" gets a tour, and a tour is longer than
   five sentences by definition.
2. Match the words to who is asking. Somebody who does not code gets what the thing does and
   what it costs them when it breaks, with the file paths kept as evidence rather than as
   the answer; somebody who does gets the path first. Ask which if the question does not
   say, and never answer a business question with a call graph.
3. Find the real entry point before reading anything else. Working outward from the wrong
   file produces a confident answer about the wrong subsystem.
4. Follow one real path end to end and say what happens at each hop. Never summarise a flow
   from file names alone — that is the single most common way this goes wrong, and it reads
   exactly like a correct answer.
5. When the shape is the answer, draw it. Under 70 columns, and every box is a real file
   that exists:

   ```
   settings.json ──> chain.py ──> self_protect ──> exit 2
                        └───────> loop_guard
   ```
6. Explain what is here, not the pattern in general. If a textbook name applies, one clause
   is enough; the reader came for this codebase.
7. Name the pitfall by pointing at a line in this repository. A generic warning helps
   nobody; "this returns None on line 84 and the caller does not check" does.
8. If the answer is "it does not exist here", say that in the first sentence.

## Done when

- Every claim has a path, or is explicitly marked as unverified.
- The person could reach your conclusion by opening the files you named, in that order.
- Nothing was changed. This skill reads; it never writes.

## What this is not

Not a design review, not a refactor, and not an opinion about quality. If you spot
something genuinely dangerous, say it in one line at the end and move on.

- "I read the file names, so I can summarise the flow" — a flow summarised from file names
  alone is the single most common way this goes wrong; follow one real path end to end.

## Routing

In scope — routes here:

- "where does X live" — a path and a sentence answer it.
- "how does this work", "why does it do that" — the flow, followed end to end through real
  files.
- "what depends on Y" — dependents mapped with anchors to the real call sites.
- "walk me through", "onboard me" — a tour of the module for somebody who has just arrived.
- "map this module", "trace this import chain" — the shape of the code, drawn from real
  paths.

Not for:

- Evidence from outside this repository — use /ai-research, which goes and gets findings we
  do not have yet and cites them.
- Diagnosing a failure — use /ai-debug, which names a cause at `file:line` and writes the
  check that fails for it.
- Judging a diff — use /ai-verify, which passes verdicts on completed work, not tours.
- Deciding what to build — use /ai-plan, which needs options, a recommendation and the
  authority to proceed.
- Saving a finding we just made — use /ai-note, which records what we learned; exploring
  only reads.