Docs Cleaner

Docs Cleaner

Documentation goes wrong in recurring ways, and they need different work.

It rots. Something changed — a port, a path, a procedure, a decision — and the docs still describe the old world. The field's name for this is documentation rot or version drift: artifacts fall out of sync one at a time until people stop trusting the set.

It sprawls. The same topic accumulates three files, each partly right, none authoritative.

They trace to one root cause: a fact was written down in more places than it was defined. Everything below follows from fixing that.

Core principle (governs both modes)

Critical evaluation before deletion. Never blindly delete. Analyze each section's unique value before proposing removal. The goal is reduction without information loss.

This governs every deletion-shaped action in this skill, not just consolidation: archiving a document, dropping a derived value, and removing a cross-reference are all deletions and all owe the same analysis first. "It looked redundant" is not that analysis.

Which mode are you in?

They share the Drift Test below. When both apply, run Mode 1 first: consolidating docs that are also factually stale just produces one confidently wrong document.

A note on CLAUDE.md / AGENTS.md: both modes apply to them like any other document — govern them after a change, consolidate them when they duplicate each other. The one job that belongs elsewhere is restructuring a single oversized one — pushing low-frequency detail down into references while keeping the top level lean. That is what daymade-claude-code's claude-md-progressive-disclosurer skill is built for; prefer it when the problem is size and layering rather than truth or duplication. If it is not available, Mode 2 still handles the file — you just do the value analysis by hand.

The Drift Test (the core of both modes)

Before writing any value into a document — and when deciding whether an existing one earns its place — ask these questions in order, and stop at the first hit:

1. Can this be computed from details already recorded here or in the authoritative source? → It is a derived value. Do not write it. Compute it when someone asks. Examples: a count ("6 sections", "N groups covered"), a total, a status summarizing rows in a table below it, "last updated on" when the log underneath already says.

When you remove one, repair the sentence rather than leaving a hole: "the four checks below" becomes "the checks below". Deleting the number and leaving dangling grammar is a different defect, not a fix.

Exception — generated navigation. A table of contents restating the headings is a derived value by this test, and hand-maintained ones should indeed go. But leave it alone when it is produced by the docs toolchain at build time, or when the platform requires the marker to render navigation: those are computed on demand, which is what question 1 asks for. Check for a generator before deleting one.

2. Not computable, but the same fact is authoritatively defined somewhere else? → It is a copy. Link to that definition. Do not restate the value.

3. Neither derivable nor copied — it independently records what was true at a particular moment? → It may remain as an atomic historical observation: a dated change, original measurement, price observation or decision. Never replace its past value with a current value or a current-config link.

Apply the ordered test to each claim, including claims inside changelogs, postmortems and dated notes. A date does not exempt a locally derivable total, ratio or copied summary from questions 1 and 2. Remove that derived or copied restatement while preserving the independent event, measurement and causal context.

For example, a dated upload-limit change records the change at that time; today's config is a different fact. A region-addition event survives removal of a total computable from its recorded historical region details. Do not recompute that old total using today's regions.

A unique historical atom that passes questions 1 and 2 is not a drift repair. Its content value still receives Mode 2's analysis under the existing authorization. When the evidence does not establish whether a claim is derivable, copied or independently observed, retain it as unresolved and report the gap; do not guess a basis for deletion. New approval is needed only when the intended action exceeds the authorization already given.

Linking only solves case 2. Pointing a link at a derived value is a category error: there is no single authoritative cell to point at, so the link becomes scaffolding that goes stale on its own and creates a false impression that the two places are aligned.

Why this ordering matters: the mature answer to drift in API documentation is "generate it from the source, don't hand-copy it" (OpenAPI/Swagger exist for exactly this reason). Question 1 is that same instinct applied to prose — the surest way to keep a number correct is for it not to be written down twice.

Position gives no exemption

A fact is a fact wherever it sits. All of these are body text for the purposes of the test above, and all of them are where stale values actually survive audits:

frontmatter · YAML/JSON metadata · parenthetical asides · a description column in an index table · checkbox state · introductory framing before the real content · cross-document footnotes · an extra column added to a table · commit-message-shaped notes left in prose

What counts as a fact

Anything that changes over time and has an authoritative definition somewhere: numbers, status, ownership, relationships, classifications, rules, decision sources, dates, counts. Business rules count too — "all quotes route through the finance lead" is defined once and referenced, never restated in each downstream doc.

Mode 1 — Post-change governance

Run these in order. The order is the point: fixing derived docs before establishing what is authoritative just multiplies the work.

1. Scope it from the change, not from the repo. Enumerate the facts this change altered — this port, this path, this procedure name, this default. That list is what bounds the work. The list is bounded; the search for each item on it is not — a fact's quiet copies live precisely in files the change never touched, so each fact gets searched repo-wide in step 5. What is forbidden is the other thing: opening files to look for unrelated problems. An unbounded sweep is how a doc task turns into an unreviewable refactor.

2. Identify the authoritative source (SSOT) for each affected fact. Which file defines this port / path / procedure — as opposed to mentioning it? Update that first. Everything downstream either points at it or is derived from it.

The practical test for "defines" is where a change has to be made for reality to change: the file the deployment actually reads, the schema the code loads, the runbook a human follows step by step. A file that recites the value while explaining something else is mentioning it. When two files both look declarative, prefer, in order: the one a machine consumes over one only humans read; the one nearest the thing it describes; the one other docs already cite. Then say in your report which you picked and on which of those grounds — an arbitrary pick becomes permanent once step 2 points everything else at it, so it should be a stated decision rather than an accident of which file you opened first.

3. Resolve implementation disagreements from established intent and authorization. Treat code, scripts, config, and observed runs as evidence of current behavior, not as automatic authority over intended behavior. When a requirement, authoritative source, or test establishes the intended behavior and the user has already authorized the scoped implementation fix, repair and verify the implementation first, then update the docs to match. Keep that repair within the facts and files already authorized by the task.

When intended behavior is ambiguous or the implementation change is not already authorized, report the disagreement and obtain the missing decision or authority before editing code. Do not make the documentation describe wished-for behavior, and do not silently turn a documentation pass into an implementation project.

4. Decide each affected doc's disposition before editing it.

Do not patch a document whose premise is gone — you get something internally consistent and globally wrong, which is harder to spot than an obviously outdated file. A "this section is superseded" note left inline is not a third option. Maintainer bookkeeping does not belong in a reader's path; readers lack the context that makes it meaningful and only get confused. Move it out.

"Archive" means: out of the reader's path, still in history. Concretely: look for a convention the project already has (_archive/, archive/, deprecated/, or whatever it uses) and follow it; if there is none, create _archive/ at the same level as the file. Either way move it with git mv so history follows, then repoint or remove every inbound reference. Do not delete it outright, and do not settle for adding status: deprecated frontmatter while leaving the file where readers still find it — a retired document still in the reader's path is still being read.

Ask before archiving anything you did not create, for the same reason step 3 of Phase 4 needs agreement. Running unattended and cannot ask: leave the document untouched and report it as needing retirement. Do not patch it instead — a document whose premise died gets more misleading, not less, when you update its facts.

5. Find every copy of each changed fact — before you change it. This is the step that gets skipped, and it is where the classic failure lives: the obvious occurrence gets updated and three quiet ones do not. See Verifying below for the exact commands.

The highest-risk shape is renumbering or inserting an item into a sequence, because you are not changing one fact but a whole family of mutually-referencing ones: body text saying "see step 7", cross-references in sibling files, and — the nastiest — prose that baked the count into a sentence ("the four checks below"). That last one contains none of the numbers you are searching for, so a search by step number will never surface it.

There is no mechanical sweep for this, and you should not pretend otherwise. Two obvious scopings were measured and both fail:

• Search count words repo-wide → 224,330 lines / ~550 MB on a real documentation repo (17.7 MB scoped to a single project directory). Nobody runs that, and an instruction that guarantees its own bypass is worse than no instruction.
• Search count words only in files that mention the subject → bounded, but wrong in the direction that matters. Prose that bakes a count into a sentence is exactly the prose that refers to the subject by alias, translation, pronoun, or bare link — "the four checks below", 「预检有四个检查」 — and never by its title. Measured: a file containing a genuine stale count matched zero searches for the subject's name or filename. The filter is anti-correlated with the target.

So do this instead, and report what it does not cover:

1. Read the file you renumbered, in full. Most stale self-references live in it.

2. Build the file list, then search count words inside it. Nothing earlier in Mode 1 produced this list — do not reach for the inbound-reference search in Updating references, which belongs to Mode 2 and answers a different question ("what pointed at this file", not "what discusses this subject"). Produce it here:

   ROOT=$(git rev-parse --show-toplevel 2>/dev/null) || { echo "not in a git repo — pass the docs root explicitly" >&2; exit 1; }
   rg -l -F "<the subject's name>" "$ROOT" --no-ignore --hidden -g '!.git'

   That gives you files naming the subject outright; step 3 covers the ones that do not. It is a small set, so name the files directly rather than piping a list into xargs (an empty list makes GNU xargs run the search with no path at all, which searches the current directory instead — a wrong scope that looks like output):

   rg -n -i -H \
     -e '\b(one|two|three|four|five|six|seven|eight|nine|ten|both) [a-z]+\b' \
     -e '[一二三四五六七八九十两](个|条|步|项|点)' -e '[0-9]+(个|条|步|项|点)' \
     -e '\b[0-9]+ [a-z]+\b' \
     path/to/one.md path/to/two.md

   Two details, both measured: -H because rg omits the filename when given exactly one file, and "which file" is the entire answer here. And the pattern deliberately takes any following word rather than a list of nouns — a list will not contain the noun your project uses ("four gates", "three lanes") and the miss is silent. Over a handful of files the extra noise is cheap; the missed line is not.

3. List the subject's aliases yourself and search those, then the count words in whatever that turns up. This is the step no tool does for you.

Then say what you checked: "read the renumbered file; searched count words in the N files that reference it by name; prose referring to it by other names was not mechanically covered." That sentence is worth more than a green check mark that quietly meant one of the two failures above.

6. Clean derived values you meet along the way — but only in files you were already editing. Removing a stale count in a file you have open is hygiene, not scope creep. Opening unrelated files to hunt for more is scope creep. That boundary is what keeps this a reviewable change.

7. Report what you did in the structure under What to report.

Mode 2 — Consolidation

Consolidate redundant documentation while preserving all valuable content. The Core principle above governs every deletion in this mode.

Phase 1: Discovery

1. Identify all documentation files covering the topic, and state the boundary you searched within (a directory, a suite, one project) before you search — the same bounded-scope rule Mode 1 step 1 uses. Two agents given "the docs" and no boundary consolidate different file sets. A topic-keyword sweep inside that boundary is the usual method — note -F, since topic terms like C++, .NET or v1.0.0 are literals and as patterns they match almost everything: rg -l -i -F -e '<topic term>' -e '<synonym>' <boundary> --no-ignore --hidden -g '!.git' This list scopes the entire consolidation, so a garbage list means a garbage analysis.
2. Count total lines across files — wc -l over the list from step 1
3. Map content overlap: for each heading in each file, note which other files cover the same ground. The artifact is a table of heading → files-that-also-cover-it; sections appearing in more than one file are the consolidation candidates. This step decides overlap only — it does not decide keep-or-delete. A unique section can still be dropped in Phase 2 for a reason that has nothing to do with duplication (self-evident, superseded, a byproduct), so read this as "nothing here licenses deleting a unique section as redundant", not as "unique means it survives".

Phase 2: Value analysis

For each document, build a section-by-section table:

The Delete example names an authoritative duplicate. Historical claims receive the same ordered Drift Test; a historical event and the current configuration are not necessarily the same fact.

• Keep — unique, essential, frequently referenced

• Condense — valuable but verbose

• Delete — duplicate, one-time, self-evident, outdated

• Archive — a whole document whose premise is dead (a cancelled system's design notes, a retired service's postmortem). It is not stale current-state to be corrected and not duplicate content to be dropped. Without this category, "superseded" and "it is a record" pull the same document in opposite directions and whichever you read last wins.

  A premise-dead section inside a live document is the same problem one level down, and Archive does not accept it — it is defined over whole documents. Treat it as a Delete whose reason is the dead premise, and name it in the Phase 3 plan rather than deciding it yourself: "the Kafka migration section, for a migration that was cancelled" is exactly the kind of call the owner may answer with "actually we still reference that".

  An Archive verdict removes that document from this consolidation. It does not go on to Phase 3 or Phase 4 — there is nothing to merge and nothing to condense. Handle it with Mode 1's archive procedure — the paragraph under Mode 1 — Post-change governance that begins "Archive" means: out of the reader's path, still in history — say in your report that you did, and carry on consolidating the rest. Do not put it in the Phase 3 line-count plan either: it is leaving intact, so counting its lines as "reduction" would claim a saving you did not make.

Run the Drift Test over every row as you go, and let it override the categories in both directions:

• A section that only restates values computable from elsewhere is a Delete regardless of how well written it is (question 1).
• Historical sections still receive questions 1 and 2 per claim. Their locally derivable aggregates and copied restatements can be removed; independent atomic observations may remain under question 3. Name any proposed removal of a unique historical atom in the value analysis and respect existing authorization; age or "one-time record" alone is not a deletion reason.

Detailed criteria: references/value_analysis_template.md.

Phase 3: Consolidation plan

Propose the target structure with real numbers, and present it for agreement — this is the plan Phase 4 asks you to get signed off before anything is deleted:

Before: 726 lines (3 files, high redundancy)
After:  ~100 lines (1 file + a pointer from the project's main doc)
Value preserved: pending — established by the preservation check after execution

The before/after line measurements describe the proposed structure. "Value preserved" is not: at plan time it would be the deleting party asserting its own deletion was harmless, which is the self-certifying shape this skill exists to remove.

It is filled in after Phase 4 step 7, and it is not a percentage — a percentage of what would have to be invented. It is a count against a stated denominator: "all 34 items on the step-1 inventory survived", or "31 of 34" naming the missing items and where each went.

The denominator is the Keep and Condense items only. The step-1 inventory deliberately also covers Delete sections, and those items are supposed to be absent from the consolidated file — counting them here would make "all survived" unreachable by construction, and a target nobody can hit is one people stop aiming at. They are accounted for on their own line: "and 6 items from Delete sections, each named with where its copy survives or why it is going". Two numbers, two questions, neither one hiding in the other. Write the denominator, not just the verdict — "every item I checked survived" is true of one item as easily as thirty-four, so it reports your diligence rather than the document's fate. Anything short of all is a finding to report, never a number to round up, and never a reason to undo your own condensation so the report reads better.

Phase 4: Execution

Deleting someone's documentation is irreversible from the reader's side even when git can undo it. Present the Phase 3 plan and get agreement before step 3. Phase 3 says "propose" for a reason; the Core Principle governs how carefully you analyze, not whether you are authorized. If you are running unattended and cannot ask, do steps 1–2, stop, and report what you would delete.

1. Extract the preservation evidence first — before you write a single line of the consolidated document. The order is what makes the check mean anything: a claim list written after the rewrite is a list of what you happened to carry over, so it always matches. Written before, it is a specification the rewrite has to satisfy.

   Be honest about what enforces that: nothing in the finished artifact distinguishes a list written before from one derived after, so this is a convention, not a gate. Give it the only teeth available — emit the list before you start step 2, into the report or a scratch file. A list that exists in writing before the rewrite can be checked against it; one that appears afterwards cannot be told from a transcription of the result.

   The unit is one entry per load-bearing item, not per section and not per category. How finely to cut is a judgment call and two careful agents will land on different counts — that is tolerable. What is not tolerable is which clauses go missing when someone cuts coarsely, so two of them get their own entry no matter how you are counting:

   • The consequence clause. "Roll back before retrying" is an instruction; "retrying on a half-applied migration corrupts the events table and there is nothing that repairs it automatically" is the reason anyone obeys it. Coarse lists keep the instruction and drop the consequence, which is the half that makes it survive a skim.
   • The symptom → cause mapping. "A 401 from the deploy step means the token lapsed, not that the account is wrong" cannot be re-derived from "the token expires every 90 days". It is the troubleshooting content, and it is one line long, so it disappears first.

   Those two are a floor, not the boundary of what matters. They are named because they are the clauses coarse lists drop most reliably — not because a clause outside them is safe. Three that fall through and are worth checking for by hand: ordering constraints ("apply the migration before restarting the workers" — no consequence stated, so the first rule does not reach it), scope qualifiers ("this applies to the staging cluster only"), and thresholds buried in prose ("backs off up to 5 attempts"). Each is one clause long, each changes what the reader does, and none announces itself.

   Go through every section — including the ones marked Delete — and write down, from the source:

   • Keep → one distinctive verbatim line per item in it (each gotcha, each constraint, each snippet — a section holding three gotchas owes three lines, not one)
   • Condense → each load-bearing claim: every constraint, number, condition, warning
   • Delete-as-duplicate → the items in it, and for each, where the surviving copy is. This is the step most likely to be skipped and the one that catches the worst failure: Phase 1 maps overlap by heading, so a section can be marked duplicate on its title while holding one gotcha no other file has. Skip the inventory and that gotcha is invisible to every later check — it was never on the list, so step 7 reports "every item survived" and means it.
   • Delete for any other reason → the items in it and why each one is going. A deletion you can name is a decision; one you cannot is an accident you have not noticed yet.
   • All of the above → every URL, verbatim

   What this list can and cannot do. It catches recall failures — you meant to carry something over and did not. It cannot catch judgment failures: if you decided at Phase 2 that something did not matter, it never enters the list, and step 7 will happily pass without it. That gap is real and nothing in this procedure closes it. The Phase 3 plan goes to a human, but it carries target structure and line counts — not this list — so it cannot catch a misjudged item either; it is a check on the shape of the result, not on what went into it. Do not report step 7 passing as evidence that nothing valuable was lost — report it as what it is, a check that everything you identified was carried over. If you want the judgment reviewed too, the only thing that does it is showing the item list to someone who did not write it.

   One sanity check before continuing: does the item count look like the source material? A three-page runbook that yields four items means you skimmed it — and a thin list is exactly the list that will later pass.

2. Create the consolidated document, carrying every item on that list

3. Delete the redundant source files

4. Update every reference to them (CLAUDE.md / the project's main doc, README, imports)

5. Leave a pointer to the new location in the main doc or README — create one if none existed. Consolidation moves content to a place nobody has a path to yet; without this step the work is discoverable only by whoever did it.

6. Verify no broken links (see Updating references below — grepping the bare filename is not enough)

7. Before calling it finished, run the preservation check below, using the list from step 1. It is a gate on this phase, not an optional appendix.

   If you skipped step 1 and the sources were tracked in git, they are still recoverable — but mind which ref: before the deletion is committed the content is at git show HEAD:<path>; once it is committed that path no longer exists in HEAD and you need git show HEAD~1:<path> (or the commit before the deletion). The path is repo-root-relative, so from a subdirectory use git show HEAD:./<path>. Both wrong forms fail loudly with fatal: path ... does not exist, so you will know. What you must never do is take the expected strings from the consolidated file — that measures it against itself.