Write Docs Agent Skill

Write Docs

A doc is a glossary: it names the moving parts, says why each exists, and

carries the principles a reader can't derive by grepping. The code is the source

of truth for *what exists right now* — the doc must never race it. Before writing

any line, ask the one question that governs this skill: **could the reader get

this faster and more reliably by reading the code?** If yes, point them at the

code instead of copying it in.

Keep vs shed

Keep — a reader cannot grep their way to these:

• The WHY: why a thing exists, the principle behind a split, the taxonomy file

names don't reveal.

• Non-obvious discoveries: platform quirks, "if you remove this, X breaks

because Y", a contract two files silently share.

• Pointers to where things live, and what KIND of thing lives there plus the

rule for what belongs.

Shed — the code or its git history already holds these, so a copy only rots:

• Exact rosters: scene names, scenario tables, helper lists, class ids, file

lists, command/flag matrices.

• Narrative and changelog: "we renamed X", "the old Y was removed as a dup",

what was tried and abandoned.

• Exact counts, tick values, current constants, line numbers — anything that

just restates the code.

Point, don't transcribe

Where you're tempted to list specifics, name the folder or the single file that

owns them and send the reader there — **prefer a folder over a file, a

source-of-truth registry over a transcribed copy.** Describe a helper module by

the *kind* of helper it holds and the rule for what belongs in it, not its

current roster. One concrete touchstone is fine to ground a principle; a full

inventory is the smell. "The matchups live in the table that defines them; the

ids key off the class registry" stays true after the next edit — reproducing

either does not.

One home per fact

Every fact has exactly one canonical home; every other doc links to it.

Repetition across docs is a maintenance bug — copies drift and the reader can't

tell which is current. A root/global doc gets a short section plus pointers to

the sub-docs. When two docs explain the same thing, pick the hub and cut the

other to a pointer.

Link the tree, downward

Docs form a tree reachable from one root/hub doc: the hub links to each sub-doc,

and a sub-doc links on to any module-level doc beneath it. Navigation flows

down — a reader starts at the root and follows links in, so the whole tree is

reachable from there. A doc links back up to its parent, or across to a sibling,

only to reuse a fact that already lives there — an app doc pointing at the

hub's deployment section instead of restating it, a boundary doc naming the

sibling it defers to. That is the one-home rule doing its job. What you do *not*

add is a rote "part of X" back-link that carries no information: it's noise, and

the tree is already navigable from the root without it.

Edit pass

When trimming an existing doc, delete on sight: enumerations of

code-discoverable items, changelog and narrative, and any sentence that restates

the code. Then read what survives as a stranger with no conversation history —

every remaining line should be a principle, a why, or a pointer. If a line would

be just as true and just as useful as a link, make it the link.