ai-explore
Answers repo questions with file:line evidence and tours unfamiliar modules. Trigger: "where does X live", "how does this work", "walk me through", or "onboard me". Not for external evidence (/ai-research), failures (/ai-debug), or diff correctness (/ai-verify).
Install
npx skills add https://github.com/arcasilesgroup/ai-engineering/tree/main/skills/ai-explore
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install arcasilesgroup-ai-engineering@llmmart
git clone https://github.com/arcasilesgroup/ai-engineering.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole arcasilesgroup/ai-engineering collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
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
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.
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.
Find the real entry point before reading anything else. Working outward from the wrong file produces a confident answer about the wrong subsystem.
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.
When the shape is the answer, draw it. Under 70 columns, and every box is a real file that exists:
settings.json ──> src/chain/mod.ts ──> self-protect ──> exit 2 └────────> loop guard (src/guards/loop.ts)
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-orchestrator, 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.
## Lifecycle
Lane: any
Writes: nothing — it is a lens, not an author
Read by: n/a
Dies: n/a — its citations land in other skills' artifacts
Next: none
Files (ai-engineering)
-
SKILL.md 3.8 KB
--- name: ai-explore description: >- Answers repo questions with file:line evidence and tours unfamiliar modules. Trigger: "where does X live", "how does this work", "walk me through", or "onboard me". Not for external evidence (/ai-research), failures (/ai-debug), or diff correctness (/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 ──> src/chain/mod.ts ──> self-protect ──> exit 2 └────────> loop guard (src/guards/loop.ts) ``` 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-orchestrator, 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. ## Lifecycle Lane: any Writes: nothing — it is a lens, not an author Read by: n/a Dies: n/a — its citations land in other skills' artifacts Next: none
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.