Reviewing Readability Agent Skill

Reviewing Readability

Review all comments, docstrings, and names (functions, classes, variables, tests) in the target code for clarity and conciseness, and flag non-obvious logic or unclear function purpose that lacks documentation.

This skill only evaluates: it produces findings with concrete proposed rewrites and does not apply them. The caller decides whether to apply the rewrites or present them as feedback.

Audience

The reader is a developer unfamiliar with the implementation context who is trying to understand the logic in the location they are currently reading. They have general Python/TypeScript expertise but don't know the history of why things were built this way.

Principles

• Explain intent, not mechanics — don't restate what the code does; explain *why* or *what would go wrong without it*.
• Lead with the main idea — the first sentence should state the rule or intent; put mechanics, edge cases, and exceptions after. A comment can be concise and accurate yet still bury the point by opening with the mechanics.
• Concise wins — shorter comments are easier to understand. If a 4-line comment can be 2 lines, make it 2.
• Use a list for multiple cases — when a comment enumerates several conditions, outcomes, or steps, a bulleted list (-) is usually easier to scan than the same content packed into prose. Lead with a one-line summary, then list the cases.
• Avoid jargon without context — if a term is project-specific (e.g. "delta path", "fragment path", "DG"), either define it briefly or use a more descriptive phrase.
• Names should stand alone — a test name or function name should communicate what it does without needing to read the docstring.
• Comment non-obvious logic, not the obvious — skip comments that restate the code (# increment counter), but flag genuinely complex or non-obvious logic that has *no* explanatory comment. Likewise, flag a function whose purpose isn't clear from its name and signature and that lacks a brief docstring; leave self-explanatory functions undocumented.
• Comments that say "unreachable" or "no-op" should explain why — the reader needs to know why the case can't happen or why no action is needed.
• Prefer active voice; name the actor — passive constructions ("the id is assigned", "completions that are reported") force the reader to infer who does what. Say who acts on what ("the runner assigns a new id", "the frontend reports completions"). This is easy to miss because passive prose can still be accurate and concise — check for it explicitly.

Evaluation Process

• Collect all comments, docstrings, class names, function/method names, and test names in the target scope.
• For each item, ask:
• Would a newcomer understand this on first read?
• Is there jargon that isn't defined nearby?
• Could it be shorter without losing meaning?
• Does it explain the "why" or just the "what"?
• Does the first sentence state the main idea, or does it bury it under mechanics?
• If it enumerates several cases, would a bulleted list scan better than prose?
• Is it in passive voice? Would naming the actor and switching to active voice read more directly?
• For names: does it communicate the purpose without reading the body?
• Also scan for missing documentation: is there complex or non-obvious logic with no explanatory comment, or a function whose purpose isn't clear from its signature and that has no docstring? Any comment or docstring you propose adding must itself follow the principles above — lead with the intent, stay concise, use active voice, and don't narrate the obvious.
• Report the findings per the Output Format below.

How much to flag

Readability fixes are cheap — a comment reword or a rename takes seconds, so don't spend effort ranking findings by importance or deciding what's "worth it."

• Flag everything that makes the code clearer or more concise. The only thing you skip is a change where it's genuinely ambiguous whether it improves readability (a lateral rewrite that's just a matter of taste). If a change is a clear improvement, include it no matter how small.
• Don't categorize by priority or severity. Leave alone only what's already clear and concise.

Output Format

Produce findings, grouped by file. For each item, give the location (file and line or symbol), the issue, and a concrete proposed rewrite (or, for missing documentation, the comment/docstring to add).

Common Patterns to Flag

• Complex or non-obvious logic with no explanatory comment
• A function whose purpose isn't clear from its name and signature and that lacks a brief docstring
• Comments that open with mechanics or edge cases instead of leading with the main point
• Comments that explain the implementation history instead of current behavior
• Docstrings that list every parameter's type when the signature already has type annotations
• Test names that use internal abbreviations (e.g. test_dg_inside_fp instead of test_write_within_fragment_scope)
• "Pass through" / "falls through" without saying what happens instead
• Passive voice that hides the actor (e.g. "a new id is received", "completions that are reported") where active voice would read more directly
• Multi-line comments where one line would suffice
• Several conditions/outcomes packed into prose that would scan better as a bulleted list
• Comments that were correct when written but now describe deleted/changed behavior
• Reference comments (spec, RFC, issue number) that aren't needed to understand the code, or that point somewhere a reader can't reach (dead links, private/internal tickets or docs) — flag them, proposing to drop the unneeded ones and repoint the rest to a public GitHub issue or an in-repo spec

What NOT to Change

• Type annotations (those aren't documentation)
• Inline comments that mark a subtle correctness constraint (e.g. ordering dependencies)
• Reference comments (spec, RFC, issue number) that a reader needs to understand the code and point somewhere accessible (a public GitHub issue or an in-repo spec) — keep the identifier intact rather than trimming or vague-ifying it
• Legal headers