| name | ai-code-tells |
| description | Scan a code diff for the consistent "tells" of unreviewed AI-generated code — narrating comments, generic names, hollow over-mocked tests, signature-restating docs, and unjustified decisions. Read-only; reports each finding with file:line, why it's a tell, and the concrete fix. Use when the user says "check for AI tells", "review this diff for AI smells", "ai-code-tells", or invokes /ai-code-tells. Works standalone; pairs well with any always-on code-quality rules you define in your own CLAUDE.md. |
| metadata | {"author":"Big Emotion","version":"1.0.0"} |
AI-Code Tells Review
A focused, read-only review that catches the five consistent signatures of AI code that nobody read before accepting it. It does not fix anything by default — it reports findings so a human keeps the mental model. Apply fixes only when the user asks.
The five tells below are the complete rule set — the skill needs no external configuration. If your user-level or project CLAUDE.md defines equivalent always-on code-quality guardrails, this skill acts as their on-demand review pass; honor any project-level tightening in the repo's own CLAUDE.md (e.g. hard gates).
When to Activate
- The user invokes
/ai-code-tells or asks to "check for AI tells / AI smells".
- After generating a non-trivial chunk of code, before presenting it as done.
- Before opening a PR, as a self-review pass.
Scope — what to review
Resolve the target in this order, stop at the first that applies:
- An explicit path, file list, or PR number the user named.
- Staged changes:
git diff --cached. If empty →
- Working-tree changes:
git diff. If empty →
- The current branch vs its merge-base with the default branch:
git diff $(git merge-base HEAD origin/main 2>/dev/null || git merge-base HEAD main)...HEAD.
Review only the changed lines and enough surrounding context to judge them. Do not audit the whole repo. If nothing is in scope, say so and stop.
The five tells
For each, scan the diff and record findings. A finding is file:line — <tell> — why it's a tell — concrete fix.
-
Narrating comments. Comments that restate what the code literally does (// increment counter, // loop over the list and process each item, // return the result). The fix: delete it, or replace it with the why (a decision, constraint, or gotcha) if one exists. A method whose every line has a play-by-play comment is the strongest signal.
-
Generic, domain-blind names. Identifiers like data, result, item, value, obj, temp, process(), handle(), doStuff(), Manager, Helper, Util where a domain word exists. The fix: rename to what it is in the problem domain. Flag only where a more meaningful name is genuinely available — loop indices and trivially-scoped temporaries are fine.
-
Hollow / over-mocked tests. Tests that (a) mock or stub the very thing under test, (b) reach into private internals via reflection / as any / @ts-ignore / accessing _private members instead of the public interface, (c) assert only that a mock was called rather than on real output/behavior, or (d) are elaborate scaffolding that would pass even if the logic were wrong. The fix: test the public interface the way a caller uses it; assert on observable output or state; mock only true external boundaries (network, clock, fs).
-
Docs that restate signatures. Doc comments / JSDoc / docstrings that just re-list the parameters and return type the signature already shows, with zero rationale. The fix: keep docs that explain why, constraints, edge cases, and tradeoffs; drop the rest.
-
Decisions with no recorded "why". A non-obvious choice (an algorithm, a data shape, a dependency, a workaround, a magic number) introduced with no comment, commit-body line, ADR, or PR note explaining the reasoning. The fix: add the rationale where the project records it (inline comment for local choices, ADR / changelog / PR body for architectural ones).
Output format
Report grouped by tell, most severe first. For each finding:
[Tell N — short label]
path/to/file.ext:LINE
found: <the offending snippet, one line>
why: <why this is a tell>
fix: <the concrete change>
End with a one-line verdict: a count per tell and whether the diff reads like code a human reviewed. If the diff is clean on a given axis, say so explicitly rather than omitting it — "0 narrating comments" is a useful signal.
Rules
- Read-only by default. Report findings; do not edit. If the user then says "fix them", apply the minimal change per finding and nothing else (scope discipline).
- No false-positive padding. A name is only flagged when a better domain name actually exists. A comment is only flagged when it adds nothing. Precision over volume — a noisy report trains the user to ignore it, which defeats the point.
- Be able to defend every finding. This skill exists to rebuild the reading habit, not to nag. If you can't say why a flag matters, drop it.
- Respect project tightening. If the repo's CLAUDE.md promotes any tell to a hard gate (e.g. "no reflection in tests", "names from the domain glossary"), report violations of it as blocking, not advisory.
