Make the Docs Trustworthy

• A second copy of a fact destroys that. The two disagree in the end, and neither one announces itself as the stale one.
• The weight sits in Search before you write. Everything after it is what to do once the search has answered.
• You opened this in the middle of something. This is how to do that work, not a replacement for it. Name what you were doing before you start, and return to it when this is done.

Search before you write

• You MUST look for the fact first. Where it already exists, edit it there.
• A second copy adds no source. It adds a sentence that will disagree with the first one.
• The reader cannot tell which of the two is current. That is the whole cost, and it lands on someone else.
• Where something already prints the fact, write the pointer and stop. A manifest, a config file, a generated block, a command's own help.

Write only what cannot be found by looking

• Keep the convention nobody wrote down.
• Keep the reason a thing is the way it is.
• Keep the trap that has caught people.
• Drop whatever one command or one file already answers. That is restatement, and restatement is stale the moment it is written.

Which rules to read

• One rule per row. The match sets where to start.
• The change class sets what is permitted. Name it before the edit.
• Run the search this gate demands before any rule. It can make the write unnecessary.
• Read both rows where a claim matches two. Reading is the cheap half. Writing the wrong thing is the expensive half.
• Supersede where a rule would delete a decision record.

rules/one-place-for-a-fact.md

rules/restatement-is-drift.md

rules/staleness-without-a-diff.md

rules/one-artifact-one-job.md

rules/unreviewed-prose.md

rules/record-what-code-cannot-show.md

rules/supersede-or-delete.md

rules/match-the-existing-shape.md

• One-place against restatement. Split by what the duplicate is. Another piece of prose, against an artifact that renders the fact already.
• Staleness against one-artifact. Split by symptom. A sentence that is untrue, against a file whose job cannot be stated.
• Record against supersede. Split by moment. The first time it is written, against every time after.

• Search for the fact before writing it, and edit it where it already lives.
• Write only what cannot be found by looking. Anything a command already prints becomes a pointer.
• Say what you chose not to write, so the omission is a decision rather than a gap.

Prose nobody has read is not a source

• Do one of two things before citing a document as your reason. Re-derive its claim from the system, or say plainly that you did not.
• You SHOULD re-derive rather than disclaim where the claim decides anything. Disclaiming is the honest option, not the cheap one.
• You MAY take either one. The silent third option is the one that costs.
• That silence turns one wrong sentence into a convention.
• Agent-written prose piles up fastest and gets trusted first. That is backwards.

Settle the claim against the running system

• This skill runs when nothing points at what changed. Someone says the docs are wrong, and there is no diff to anchor it to.
• Settle each claim yourself. Run the command the page describes. Open the path it names.
• Check the identifier is still exported. Follow the link and see where it lands.
• Report it confirmed or refuted, and name the thing that settled it.
• Judging docs inside a change is a review's job. There a diff supplies the anchor, and findings are reported rather than applied.

Name the change you are making

• Deleting a page whose subject was removed is correct. It is part of the removal.
• Deleting a record because its answer changed is not. It destroys the evidence that the old answer was once right.
• The next reader then argues it again from nothing.
• Propose any deletion, and any new file, with what depends on them named. Correcting and relocating need no proposal.

Keep the verbatim set through any rewrite

• A rewrite, a compression, or a move preserves these exactly. Code, paths, identifiers, commands, error strings, numbers, versions, and frontmatter.
• Frontmatter is the one that fails quietly. A model told to preserve it rewrites it anyway.
• Lift it out before the change and put it back after.
• Check the result against the original. Never trust that the instruction worked.

What to produce

Claim        "the setup guide is wrong"
Resolution   refuted for steps 1-3: ran them, they work.
             confirmed for step 4: it names a flag the command's own
             help no longer lists.
Changes      correct   <setup page>:22, replace the flag with a pointer
                       to the command's help
             delete    <old setup page>: its subject was removed here
Left alone   <troubleshooting page>: every command in it still runs
Not written  the retry default: the config command already prints it
Unreviewed   none

• One instance, not the set. What generalises is the shape.
• Every claim gets a verdict, and every change gets a class.
• What you chose not to write is said out loud.

Do not skip this when

• The fact seems new. That is what every duplicated fact seemed like once.
• The page is small. A small page is read whole and believed whole.
• You are only fixing a typo. That is when the surrounding claim goes unchecked.
• Nobody asked. Prose nobody checked is what the next session will cite.

Routing

• The table above selects the rule. Read it in full, and say which one you opened, in one line.
• Whether a decision is worth recording at all is settled by the change-review rules, not here. This skill owns writing it, placing it, and keeping it true.
• A direct instruction from the user outranks anything here.