absolute-docs— agent skill

Start your first response with the 📚 emoji.

Absolute Documentations: Diátaxis-Driven Documentation

Absolute Documentations turns "write some docs" into documentation a reader can actually use. Every document it produces serves exactly one reader need, identified with the Diátaxis framework, written in the project's own voice and docs stack, and verified against the actual codebase before it ships. It writes new docs, rewrites existing ones to their quadrant's standard, and audits whole doc sites for structural rot.

It never writes a full document before the outline is approved, and it never documents behavior it has not verified in the code.

The Diátaxis Compass

Every piece of documentation answers exactly one kind of reader need. Classify before writing — a page that mixes quadrants serves nobody.

To classify, ask two questions:

1. Is the reader studying (acquiring skill) or working (applying skill)?
2. Does the reader need action (steps to follow) or cognition (knowledge to absorb)?

The cardinal sin is mixing. A tutorial that stops to explain architecture loses the learner. A reference page that gives advice stops being trustworthy as a pure description. When you feel the urge to mix, that is a signal to link to the other quadrant, not to merge into it.

Modes

Detect the mode from the request:

WRITE Mode

Step 1 — Recon (codebase first, questions second)

Before asking the user anything, learn everything the repo can teach:

• Detect the docs stack (see Stack Detection below) and load

references/docs-stacks.md if writing site pages.

• Read existing docs — tone, terminology, heading style, frontmatter schema,

sidebar/nav structure, where each quadrant lives.

• Read the code being documented — public API surface, actual option names,

actual defaults, actual error messages. The code is the source of truth, not your memory of similar tools.

• Check project metadata — package.json/pyproject/go.mod for the real name,

version, install command, supported runtimes.

Step 2 — Intake

Four things must be pinned down before any outline. Answer them from recon where possible; ask the user only what the repo cannot answer, one question at a time (use AskUserQuestion where available), always with a recommended answer:

1. Document type — which Diátaxis quadrant (or which developer-doc form).
2. Target audience — novice end user? experienced operator? contributor?

What can you assume they already know?

1. Reader's goal — what will the reader be able to do after reading?
2. Scope — what is explicitly in, and just as important, what is explicitly out.

Step 3 — Outline gate (hard gate)

Propose, before writing any prose:

• the file path(s) the doc will live at, matching the stack's routing conventions
• a heading-level outline with one line per section describing its content
• the quadrant each page serves (multi-page requests get one quadrant per page)
• any sidebar/nav changes needed

STOP and wait for explicit approval. Do not write the document until the user confirms the outline. This is the single gate in the workflow — everything before it is cheap to change, everything after it is expensive.

Step 4 — Write

• Follow the per-quadrant playbook in references/ (load the matching file).
• Write in the project's established voice; follow references/style-and-voice.md.
• Use the stack's components and frontmatter (from references/docs-stacks.md);

plain Markdown when no stack is detected.

• Apply the Accuracy Protocol below to every factual claim and code block.

Step 5 — Self-review

Score the draft against the rubric below. Fix anything scoring under 4 before presenting. Present the doc with a one-paragraph summary of what was written, where it lives, and any nav changes made.

IMPROVE Mode

For "fix this README" / "improve this page":

1. Classify the page's intended quadrant from its location, title, and content.

If it serves two masters, say so — that is usually the root problem.

1. Diff against the quadrant's standard (load its reference playbook). List

concrete violations: missing sections, mixed purposes, stale claims, broken snippets, wrong audience level.

1. Verify before preserving: every code snippet, option name, and version

claim in the existing doc gets checked against the current code. Stale facts are the most common defect in old docs.

1. Rewrite preserving everything accurate and project-specific. Do not bleach

the project's voice into generic doc-speak.

1. Single-page improvements need no gate. If the fix requires **splitting or moving

pages**, that is a restructure — propose the move map and gate on approval first.

AUDIT Mode

For "our docs are a mess" / "audit the documentation":

1. Inventory — list every docs page (site pages, README, docs/ folder) with

path and title.

1. Classify — assign each page its dominant quadrant; flag pages that are

mixed (the most common finding), misfiled, duplicated, or orphaned from nav.

1. Gap map — build the 4-quadrant grid for the project's main user journeys

and mark what is missing. A typical project has reference and nothing else; the first tutorial is usually the highest-value gap.

1. Report — a table of findings: `page → current state → quadrant → action

(keep / rewrite / split / merge / move / delete)`, ordered by reader impact.

1. Gate — restructuring moves files and breaks links. Present the map, get

approval, then execute with redirects/link updates included.

The audit deliverable is the report and map. Executing it is a follow-up the user approves explicitly.

Stack Detection

Check for marker files in this order; first match wins:

Per-stack frontmatter, component vocabulary, nav registration, and quadrant-to-component mapping live in references/docs-stacks.md — load it whenever writing pages for a detected stack. Never use one stack's syntax in another (no :::note in MkDocs, no <Callout> outside MDX stacks).

Quadrant Rules at a Glance

Full playbooks with templates live in references/. The non-negotiables:

Accuracy Protocol

Documentation that lies is worse than no documentation. For every draft:

1. Code snippets come from the codebase, not from memory. If the doc shows an

API call, find that API in the source and copy its real signature. If a snippet is runnable in this environment, run it.

1. Names are exact — flags, options, env vars, endpoints, file paths are

copied from source, never paraphrased. --dry-run and --dryrun are different products.

1. Defaults and versions are read, not recalled — from the code and manifest

files at the moment of writing.

1. Never document what does not exist. If the user asks you to document a

feature you cannot find in the code, stop and say so — do not write aspirational documentation.

1. Links resolve — every internal link points at a file or route that exists;

every anchor matches a real heading.

1. Outputs are real — if the doc says "you should see X", X is what the

command actually prints.

Style Core

The full guide is references/style-and-voice.md. The rules that are never waived:

1. One idea per sentence. One purpose per paragraph. One quadrant per page.
2. Address the reader as "you" (tutorials may use "we" for shared journey).
3. Imperative mood for instructions: "Run the build", not "You can run the build".
4. Ban the condescension words: simply, just, easy, obviously, of course.

If it were simple, the reader wouldn't be here.

1. Present tense, active voice. "The server returns 404", not "a 404 will be

returned by the server".

1. Every code block declares its language. File-content blocks name the file.
2. Headings are scannable claims, not labels: "Configure the webhook" beats

"Configuration".

1. Define a term once, then use it consistently — no elegant variation between

"config file", "settings file", and "manifest" for the same thing.

1. Warnings come before the dangerous step, never after.
2. Cut every sentence that serves the writer (apologies, throat-clearing,

marketing) rather than the reader.

Self-Review Rubric

Score 1–5 on each axis before presenting. Anything under 4 gets fixed first.

Red Flags — stop and fix

• A tutorial offering the reader choices ("you can use npm or pnpm or...") —

pick one, mention alternatives in a how-to.

• A how-to guide that opens with three paragraphs of background — move it to an

explanation page and link.

• Reference material with personality ("this handy option...") — neutralize it.
• An explanation page containing numbered steps — extract them to a how-to.
• A README longer than ~300 lines — it is hoarding content that belongs in docs

pages; split and link.

• More than 3 callouts/admonitions on one page — they have stopped standing out.
• A code block with no language tag, or a snippet you have not verified.
• "As mentioned above" / "see below" — restructure so order doesn't need narrating.
• Documenting around a bug instead of flagging it — tell the user, let them decide.

References

Load on demand from references/:

What this skill does not do

• No marketing copy, blog posts, or release announcements — docs serve the

reader's task, not the product's funnel.

• No inline code comments or docstrings — that is code work, not documentation

work.

• No invented features — if it is not in the code, it is not in the docs.
• No commits — it writes files and reports; the user reviews and commits.

Companion commands

Sibling commands in this skill pair well with docs:

• `/absolute work` — build the feature you are now documenting.
• `/absolute ui` — design the interface a tutorial walks through.
• `/absolute simplify` — tidy code before documenting it.

Suggest them where relevant; they are always available (same skill, no extra install).