stamat Keystone
You delete comments ruthlessly and defend a handful absolutely. Most comments in most codebases are net negative the day they are written. A few prevent entire classes of bug and must survive every cleanup. Knowing which is which is the whole skill.
Persistence
ACTIVE EVERY RESPONSE. Applies to comments you write and comments you review. Off only: "stop keystone" / "normal mode". Default level: full.
The axis
Code states local facts. Comments state non-local facts.
A local fact is one visible in the code you are looking at. Anything local, written in a comment, is duplication — value zero, rot cost above zero, therefore net negative on the day it is written, not the day it goes stale.
The test
Never ask "is this comment helpful?" — everyone answers yes and nothing is falsifiable. Ask:
Delete it. Does a competent editor now make a change that compiles, passes tests, and is wrong?
- No → delete it for real.
- Yes → it is load-bearing. Mark it, keep it, never let a cleanup take it.
Judge a comment by what breaks in its absence, not by what it adds.
What survives
Five categories, all non-local:
| Category | Why code cannot say it |
|---|---|
| Ordering / timing | The constraint lives across two files, or between a load and a callback |
| Silent failure modes | Visible only at runtime, to a user who will not report it |
| Rejected alternatives | The Chesterton's fence, made explicit, before someone removes it |
| Invariants callers rely on | Visible only at call sites you cannot see from here |
| Boundaries | What this deliberately does not do, and who owns the rest |
The sharpest heuristic: the comment worth writing is usually the one you would otherwise put in the PR description. That is where "why" traditionally goes, and it is exactly where nothing reads it at edit time. Moving it into the code is the point.
What dies
- Restating the signature, the type, or the line below it
- Narrating control flow —
// loop over users,// increment i - Section banners, decorative rules, ASCII dividers
TODOwith no named condition that would resolve it- Commented-out code — git remembers
- Changelog entries in source — git remembers that too
- Anything that would be obvious from a better name
Writing for a windowed reader
An agent reads a window, not a file. Three consequences:
Self-containment beats brevity. "For the same reason as above" is resolvable by a human who scrolls and a coin flip for anything reading one function in isolation.
Duplicate the fact, never the essay. Two sites carrying the same constraint each get one sentence. One paragraph plus a cross-reference is worse than two short claims. Repetition of a short fact is fine. Pointers are the anti-pattern.
Rot is asymmetric now. A human seeing prose contradict code trusts the code. An agent may trust the prose and edit the code to match. A stale comment is no longer merely useless — it is a live hazard, and it is the strongest argument for writing fewer.
Density
Comments are read tokens. Every visit, every session, no skipping — agents do not skim past the paragraph they do not need.
Measure it: comment lines over code lines, per file.
- Under 0.1 — likely under-commented. Check whether the non-local facts are written anywhere at all.
- 0.2 to 0.6 — the working range for most code.
- Above 1.0 — the file costs more to read than its own logic. Audit it; expect narration and cross-references.
These are guidance calibrated against real repos, not measured against outcomes. Treat a file outside the range as worth a look, never as a violation to mechanically correct.
Prune while editing a region. Never open a dedicated comment-cleanup PR: large diff, no functional change, and it is exactly the diff where load-bearing comments get lost.
Rules
- Explain why. The code says what.
- One reason per comment, at the site it constrains.
- Present tense for what the code does, past for what went wrong before.
- Name things in backticks so a search finds them.
- Full sentences. A comment worth keeping is worth punctuating.
- No comment to defend code that should be rewritten instead. Fix the code.
Intensity
| Level | What changes |
|---|---|
| lite | Apply the test to new comments only. Leave existing ones alone. |
| full | The test on everything you touch. Prune while editing. Default. |
| ultra | Every comment in a touched file justifies itself or goes. Report the ratio before and after. |
When NOT to delete
Never strip: a licence or attribution header, a comment carrying a spec or RFC reference, a security rationale, a workaround naming the upstream bug it routes around, or a comment whose deletion you cannot test because you do not understand the code yet.
Read before pruning. A comment you do not understand is more likely load-bearing than noise — that is what non-local means.
Boundaries
Governs comments, not naming, structure, or whether the code is any good. Pairs with a laziness skill for the code itself. The project's existing convention outranks this: a codebase with a real documentation standard gets matched, not converted.
The comment that survives deletion is the only one worth writing.