Purpose

Developer-facing comments and doc comments in this repository are read by contributors, months or years after they were written, with none of the context you have right now. This skill defines who that reader is, what each kind of comment is for, and which patterns are banned.

Scope boundary: rustdoc inside declare_lint_rule! / declare_assist_rule! blocks is end-user documentation generated into the website. Load this skill for comment hygiene, but use lint-rule-development for the audience, content structure, examples, and option documentation. Its content rules take precedence for those blocks.

The Reader

For developer-facing comments, write for a Biome contributor who is competent in Rust but has no access to your current context: not this conversation, not the pull request, not the issue, not the diff. They see only the repository at HEAD.

Two consequences follow directly:

1. Never narrate change history. Words like "now", "previously", "no longer", "the new approach" are meaningless at HEAD, where only one approach exists. State how the code works, not how it came to be.
2. Never address the reviewer. A comment that argues your change is correct ("this properly handles X") belongs in the PR description, not in the source. The comment must justify the code as it stands, permanently.

Three Kinds of Documentation, Three Different Jobs

Do not mix the jobs. Implementation details do not belong in /// docs — put them as // comments inside the body. The contract does not belong scattered across inline comments — put it on the item.

The Deletion Test

Before writing any comment, ask: does this state something the reader cannot recover from the code itself?

• If the information is already carried by names, types, or structure, do not write the comment. If the name fails to carry it, improve the name.
• Information that legitimately needs a comment: an invariant, a rationale, a coupling to code elsewhere, a workaround with a link, surprising behavior of a dependency, a term of art the module defines.

When editing later, the same test applies in reverse: a comment that no longer passes it should be deleted, not left to rot.

Behavior Documentation

Write documentation for a human reader, not as a translation of the implementation.

• Start with a plain-language description of what the function returns or accomplishes.
• Use short or medium-length sentences. Keep one main idea per sentence.
• Avoid internal jargon. If a technical term is necessary, explain what it means in the same paragraph.
• Describe business-logic caveats that can surprise callers. Examples include fallback behavior, work limits, ambiguous results, overload ordering, and conditions that return None, Unknown, or an indeterminate result.
• Do not describe implementation details unless callers need them to understand the behavior.

Add an example when the behavior depends on relationships that the function signature cannot show clearly. Common cases include:

• overload selection;
• mapping arguments to optional or rest parameters;
• following imports or re-exports across files;
• fallback behavior for ambiguous or incomplete information;
• a result whose meaning is not obvious from its type.

Introduce the example before the code block. State what the example demonstrates and what result is expected.

Keep snippets minimal and self-contained.

Module documentation should describe a durable concept or design reason. Do not list individual functions or queries merely to summarize the file. Such lists become stale as items are added or renamed. If the module has no durable concept to explain, use a brief one-line description.

Banned Patterns

Narrating the next line. Delete these on sight:

// Increment the generation counter
generation += 1;

Change-history narration. Rewrite as present-tense rationale:

// BAD: We now intern types instead of cloning them.
// GOOD: Interning avoids cloning these types on every lookup.

Reviewer-addressed justification. Move the argument to the PR:

// BAD: This correctly handles the overload case from the bug report.
// GOOD: Overloads are matched by arity before parameter types, so a
//       partial-arity call cannot select the wrong candidate.

Restated rustdoc. A /// doc that rewords the item name says nothing:

// BAD:
/// Handles the type inference.
fn infer_types(...)

// GOOD:
/// Infers the type of `expr` in the scope of `module`, returning
/// `TypeData::Unknown` when the expression references an unresolved import.
fn infer_types(...)

Vague hedging. "Some cases", "various reasons", "handles edge cases", "etc." — either name them or drop the sentence.

Ad-hoc section banners (// ----- helpers -----, // ==== TYPES ====). For grouping in long files, use the region comment pattern below instead.

Region Comments

Long files group related items with paired region markers:

// #region FILE-LEVEL METHODS
...
// #endregion

This is an established convention across the codebase (biome_service, biome_module_graph, biome_rowan, the parsers). The Workspace trait in crates/biome_service/src/workspace.rs uses it to group its methods (PROJECT-LEVEL METHODS, FILE-LEVEL METHODS, SEARCH-RELATED METHODS). Editors fold on these markers, which is the point: they exist for navigation, not documentation.

Rules:

• Every // #region has a matching // #endregion. An unpaired marker breaks editor folding silently.
• The name states what the group contains. It can be a plain label (Shared helpers) or anchored to a function (#region parse_thematic_break_parts) when the region holds one entry point and its private support code.
• Use regions only where they earn their keep: files or impl/trait blocks long enough that folding helps. A file that fits on two screens does not need them.
• A region name is organization, not documentation. It never substitutes for rustdoc on the items inside it.

Editing Existing Code

• Preserve existing doc comments. If your change alters behavior, extend or correct the specific prose — never replace it with generic text. Deleting hard-won context is worse than leaving a comment slightly stale.
• Match the surrounding density. A heavily documented module deserves the same level on new items; do not blanket a sparse module with comments.

Exemplar

The //! module docs at the top of crates/biome_service/src/workspace.rs show the target register. They define a term the rest of the module depends on ("open documents") and give its meaning in both the LSP and CLI contexts; they explain a design decision the signatures alone would make confusing (the workspace is stateful, yet every method takes &self, because the trait must be thread-safe and caching happens internally); and they state the error philosophy once, at the top, instead of repeating it on every method. Everything is present tense; nothing mentions how the design evolved or defends a change.

Self-Check Before Finishing

After completing any task that touched comments, re-read only the comments in your diff, in isolation from the code changes:

1. Does each one pass the deletion test?
2. Does any reference the conversation, the change itself, or the reviewer?
3. Would a reader without access to the diff understand each one?

Fix or delete what fails. Deletion is the default; a missing comment is cheaper than a misleading one.

References

• Diátaxis — the framework behind the explanation / reference / rationale split above.
• lint-rule-development — for rule rustdoc, which is end-user documentation.