stylewright

Writing skills for coding agents, distilled from named standards.

Install them into Claude Code, Claude Cowork, and Codex, at user scope or project scope.

This document is written under the rules of one of the skills it ships. A continuous-integration job checks it with stylewright lint.

Why this exists

An agent writes documentation all day. It has no house style unless you give it one. Most guidance you can give it is invented on the spot, and it changes each time you ask.

stylewright gives an agent writing rules that come from somewhere. Each rule in the standards/ tier traces to a numbered rule in a published standard, and anything that is not such a rule says so rather than borrowing the standard's authority. You can check the trace.

The two tiers

Tier	What it means
standards/	Every skill is distilled from a named standard that you can read at a public URL. Every statement is accounted for: it cites a numbered rule, or it is marked as our own guidance, or it is marked as narrative.
craft/	Every skill encodes discipline that has no external standard behind it. The authority is ours, and the skill says so. A craft skill may govern how an agent works as well as how it writes, which ADR-0005 decided.

Install one tier, the other, or both.

Skills

Skill	Tier	Source	License of the source
simplified-technical-english	standards	ASD-STE100 Issue 9 (2025)	(c) ASD 2025. All rights reserved. See source/standards/simplified-technical-english.md.
plain-language	standards	Federal Plain Language Guidelines	CC0 1.0 Universal. Public domain.
compressed-deliberation	craft	None. Anthropic documents the defaults, and our own baseline traces the inflation to the instruction stack above them.	No source wording is reproduced, so no reproduction right is relied on. See source/craft/compressed-deliberation.md.
navigable-references	craft	None. The rule is ours, and no measurement stands behind it yet. See source/craft/navigable-references.md.	Not applicable. Nothing is reproduced.
proportionate-execution	craft	None. Anthropic documents the behaviours, and every rule answering them is ours. See source/craft/proportionate-execution.md.	No source wording is reproduced, so no reproduction right is relied on. See the same record.
de-slop	craft	None. Every rule is ours, and no measurement stands behind it yet. See source/craft/de-slop.md.	Not applicable. Nothing is reproduced.
subagent-returns	craft	None. The skill came out of an operator request, and every rule is ours. See source/craft/subagent-returns.md.	Not applicable. Nothing is reproduced.

A craft skill has no standard behind it, so measurement is the only evidence its rules can have. bench/ holds that protocol, and bench/README.md states the one rule that matters most: run the no-guidance control first. Ours came back clean, which is how we learned the skill we were about to write would have been aimed at the wrong thing.

That is the evidence a craft rule can have, and not the evidence each one already has. navigable-references, de-slop and subagent-returns ship with none, and each source record says so in those words. Read all three as discipline that we assert.

de-slop gives a shape to write toward, and names the departures from it. It governs rhetorical moves and never words. No skill here ships a list of words gathered from what one setting overuses, because such a list teaches an agent to swap each word for its nearest neighbour. Counting those words belongs in bench/score.mjs, which does not install. ADR-0021 records that decision, and it leaves a published standard's own vocabulary under its existing gate.

proportionate-execution ships with none either, and for a harder reason. Its rules govern a session of many steps, and the bench runner drives one prompt. ADR-0005 accepted that gap when it admitted operating discipline to this tier, so no rule in that skill claims a measured effect until the runner can drive a session.

subagent-returns sits behind the same wall. It governs the one message a subagent returns to the agent that spawned it, so an arm has to drive a delegation and keep what came back. The runner cannot do that either.

More skills are in progress. See CONTRIBUTING.md to add one.

The resident fragment

A skill loads when its trigger matches. Some rules have no trigger, because they apply to every sentence. navigable-references is one of them, and its worst moment is the moment you did not notice that it applied.

So that rule also ships as a resident fragment. The fragment is a file your instruction file imports, so the rule is always in context.

npx stylewright install --skill stylewright-resident --platform claude

The command copies one file and prints one line. Paste that line into your own instruction file:

@skills/stylewright-resident/navigable-references.md

This tool never writes to your instruction file. It prints the line and stops. Run stylewright doctor afterwards, and it looks in the instruction files your agent reads for that exact line. A tool that wrote the line could only claim the rule is active. The check looks.

It looks for a line of text, and it reads no Markdown. So a copy of the line inside a code fence, or in a sentence saying you took it out, still counts as an import. A file it cannot read counts as no import, and so does one above a megabyte. So does an import you spelled another way, such as an absolute path, a ~ path, or a ./ prefix. The check compares the exact line it printed, and not every path that resolves to the same file. Read the warning as a prompt to look, and not as a verdict. ADR-0022 states both directions and why they are the ones to accept.

The fragment is generated from the skill, so the two cannot drift. A check in the pipeline fails when they do.

The fragment is the recommended delivery for this rule, and the navigable-references skill stays as the alternate. Pick one. doctor warns when you have installed the fragment, imported it, and installed the skill as well, because one rule delivered twice costs context and says nothing new.

No tier selects the fragment. --tier all installs every skill and not the fragment, so a plain install cannot deliver one rule twice. Ask for it by name, as the command above does. npx stylewright list names it beside the skills and says the same thing.

It installs for claude and cowork only. @path is a Claude Code feature, and we have verified no import form for Codex. Issue #24 holds that question open, and an import line that silently fails is worse than no fragment at all.

The standards disagree, and that is the point

plain-language tells you to use contractions. simplified-technical-english tells you not to. Neither is wrong. One writes for a general public reader who must act, and the other writes for a technician who follows a procedure.

Pick the standard that matches your reader, and follow one standard in one document. The grounding matrix shows you which rule came from where, so you can see exactly where two standards part company.

Install

With the installer

Run one command and answer four questions. You do not need to know the flags, the directory layout, or the skill names.

npx stylewright install

The dialogue asks:

1. Which skills do you want. Every skill starts selected. Remove the ones you do not want.
2. Which platforms. The installer looks in your home directory first and pre-selects the agents that it finds.
3. Which scope, user or project. Each choice shows the directory that it resolves to.
4. Whether to go ahead. The summary shows every file destination before anything is written.

Answer nothing and press Enter four times to install everything to the agents that you already use.

Skills already installed are marked in the list, so a replacement never surprises you.

With flags, for a repeatable command

Any selecting flag turns the dialogue off. Use this in a script or a dotfiles repository.

npx stylewright install --tier standards --platform claude,codex --scope user
npx stylewright install --skill simplified-technical-english --platform claude

Repeat --skill to select more than one. Run npx stylewright list for the names.

To run the code on the default branch instead of the last release, use npx github:rookslog/stylewright in place of npx stylewright.

The installer copies files. It does not create symbolic links. A symbolic link breaks when the clone moves.

With Codex, and no installer

Codex ships a skill installer. Give it a path in this repository:

scripts/install-skill-from-github.py \
  --repo rookslog/stylewright \
  --path skills/standards/simplified-technical-english

By hand

Copy the directory. This always works.

git clone https://github.com/rookslog/stylewright
cp -R stylewright/skills/standards/simplified-technical-english ~/.claude/skills/

Where files go

Platform	Scope	Path
claude	user	~/.claude/skills
claude	project	./.claude/skills
cowork	user	~/.claude/skills
codex	user	~/.codex/skills
codex	project	./.codex/skills
agents	user	~/.agents/skills

Cowork reads the Claude directory. The two names resolve to one path.

The resident fragment goes under the same path, in a directory of its own: <skills>/stylewright-resident/. It is a recorded file like any other, so update refreshes it and uninstall removes it exactly.

Commands

Command	What it does
install	Copy skills to the platforms and scope that you select.
update	Copy new versions. Stop when a file has local edits.
uninstall	Remove only the files that the installer wrote.
list	Show the skills in this repository.
doctor	Report problems, such as one skill installed in two places, or a resident fragment that nothing imports.
lint	Check a Markdown file against the mechanical rules below.
ground	Check that each grounding matrix still matches its skill.
new-skill	Scaffold a new skill that passes both checks from the start.

Update is safe

Run update with no flags. It reads the manifests that the installer wrote, so it already knows which agents and scopes hold your skills.

The installer records a hash of each file that it writes. update compares the hash first. When you have edited an installed skill, update stops and tells you which files changed. Add --force to overwrite.

update also stops when a file sits at a path it is about to write and the manifest never recorded it. That file is yours, and the tool does not know what is in it.

A file that an earlier version installed, and that this version no longer ships, is removed. Otherwise it stays on disk with no owner, and uninstall cannot reach it.

An interrupted run leaves no file without an owner either. The installer records the paths it is about to write, and what it will write there, before it writes them. The next install, update, or uninstall clears what the interrupted run left, and names each file it clears.

The cleanup removes a file only when it holds exactly what the interrupted run was going to write there. A file you wrote at one of those paths does not match, so it stays. Each copy also lands whole or not at all, which is why the content is something the tool can check.

One command at a time in a directory

A command holds the directory it works in, and a second command in the same directory refuses rather than working from a picture the first has already changed.

A command that is killed leaves the directory held. Every later command then refuses and names the file to remove:

Another stylewright command is working in ~/.claude/skills. Run again when it
has finished, or remove ~/.claude/skills/.stylewright-lock if no other run is
active.

Remove that file when you are sure no other run is going, and the next command clears up after the killed one. A command that finds a held directory while it works out what to do names it and moves on, rather than reading a record that another run may be changing. Whether a run is still alive is the one thing this tool cannot check for you, and guessing it wrong deletes files that a live run is still writing. doctor reports a held directory.

Narrow it with --skill, --platform, or --scope. A skill that this repository no longer ships is reported and left alone. Uninstall it by name when you want it gone.

What no-flag update covers. User scope, plus this directory. A project install in another directory cannot be discovered from here, so run update in that project too.

Uninstall is exact

uninstall removes the recorded files and nothing else. A note that you added to a skill directory stays.

When the last skill goes, the manifest goes with it, and the empty skills directory goes too. The agent's own directory stays, because the agent owns it.

The lint

stylewright lint checks the part of ASD-STE100 that a program can check:

• Procedural sentences of no more than 20 words.
• Descriptive sentences of no more than 25 words.
• No semicolons.
• No contractions.
• Procedure steps that start with an imperative verb.

The lint skips code blocks, tables, link targets, and blockquotes. A blockquote holds quoted material, and a before-and-after guide quotes text that is wrong on purpose.

The lint does not check vocabulary. Vocabulary compliance needs the controlled dictionary, which this repository does not ship.

Grounding matrices

Each file a skill ships to a writer has a grounding matrix in grounding/. The matrix disposes of every unit of content in that file, and each row says what that unit claims.

One matrix disposes of one file. SKILL.md answers to grounding/<tier>/<skill>.md, and a file under references/ answers to a matrix that mirrors its path, such as grounding/standards/simplified-technical-english/references/examples.md. A row names a heading, and two files in one skill can carry the same heading, so a shared row space would let a row claim the wrong occurrence. A file with no matrix fails the check, and so does a file under grounding/ that grades no file any skill ships. ADR-0030 records the decision.

Rows come in three kinds:

• A G row traces to the standard. It names the rule, such as Rule 5.1.
• An E row is our own editorial guidance. Its rule cell is empty.
• An N row is narrative. It orients you and asserts no rule.

The distinction shows you which instructions carry the authority of the standard, and which carry only ours.

The check confirms that a G row names a rule. It cannot confirm that the rule says what the row says. A person read the source and wrote the row, and no program has checked that reading.

So a G row carries the rule's own words too, in a Source text cell beside the identifier. You compare two texts in one row, instead of comparing our sentence against a 400-page PDF you have to find first. The cell says unquoted until somebody quotes the rule, and a quotation is written in quotation marks, so you can always tell which words the row claims are the source's. No program checked that claim either. The marks say where a quotation starts and stops, and the source says whether it is accurate.

Every matrix here says unquoted in every row today. Each one declares whether it may quote its source at all, in a line you can read, and five of the six say it may not. stylewright ground --check prints how many rows quote their source, beside the audited count.

So each G row records its own audit. The Audited cell says unaudited, or it carries the date a person read that row against the source and a digest of the row they read. No run of the checker raises a row out of unaudited, and editing any other cell in the row voids the audit. stylewright ground --check prints how many G rows in each matrix carry a valid current audit, beside its verdict. A date that is stale, malformed, or later than the day of the check does not count toward it.

An audit answers to one reading of the source, so each matrix names that reading above its table, as **Source version:** and a pin. The pin is the edition, the commit, or the day somebody read the source, and it says unread while nobody has read it. It joins the digest, so moving the source on voids every audit in the file at once. A rule number does not change when a standard reaches its next issue, and without the pin every audit read as current over an issue nobody had opened.

A clean check means the matrix matches the skill. It has never meant that a person confirmed a citation. Read the printed count rather than the verdict, because the count is the only line that reports the second thing.

Matrices ship at the root of the npm package, so the published ground command has them to read. No install pathway copies them with a skill. A matrix is an audit record for a reader, and an agent does not need it while it writes.

A source record works the same way. It sits in source/<tier>/<skill>.md, beside the matrix and outside every skill directory. ground --check refuses a file in a skill directory that nothing governs, so a record put back beside a skill fails the gate. ADR-0025 records that decision.

stylewright ground --check --all fails when a skill changes and its matrix does not. Every heading, paragraph, list item, table and code block counts, including the ones before the first heading. Front matter in SKILL.md does not, because the harness reads it as metadata rather than as instruction for a reader. No harness reads a reference file, so a front matter block in one is refused instead.

The check reads Markdown a line at a time, and it models no container. So it states the forms it reads: a blank line, any construct at column 0, a line that continues the paragraph above it, and an indented code block that stands on its own. It refuses every other line and names it, rather than reading a nested construct as the wrong unit.

A blockquote is one block, named by a digest of what it holds, as a table and a fenced block are. The quote runs from its first marker to the first line without one. Leave a blank line under it. A line directly below a quote is refused, because a reader may keep that line inside the quote and the check holds no state to say whether they do. ADR-0031 records the decision.

A continuation line states what it may begin with. It carries a letter, a digit, or ordinary sentence punctuation. It also carries a backtick or a tilde, where the line opens no fenced block. Any other lead is refused, because a lead character alone cannot say whether the line opens a container. Write such a line so it begins with a word, or write the construct at column 0.

Two constructs are refused at column 0 as well, because the check reads neither of them:

• A heading with no text, such as #, which opens no section.
• A list item with no content, such as -, which opens no item.

ADR-0016 records that decision, and ADR-0029 records the continuation form.

Licensing

Sources impose different licenses, so one license for the whole repository cannot work.

• The engine, the tests, and the tooling are MIT. See LICENSE.
• Each skill in standards/ carries the license of its source. See the LICENSE file in the skill directory.
• Each skill in craft/ is MIT.
• The resident fragment is MIT. It is generated from a craft/ skill, and it reproduces nothing else.

Notices

No skill in this repository is affiliated with, endorsed by, or approved by the owner of the standard that it cites.

ASD-STE100 and Simplified Technical English are European Union registered trademarks owned by the Aerospace, Security and Defence Industries Association of Europe, number 017966390.

A skill states the rules in our own words, quotes the source where the exact wording is what you need to check, and links to the official document. No skill carries enough of a standard to replace it.

Contributing

Read CONTRIBUTING.md first. It covers how to choose a source, how to write a grounding matrix, and what a pull request needs.

Do not create skill files by hand. The scaffold puts them in the right places and starts green on both checks.

node bin/stylewright.mjs new-skill plain-language --tier standards \
  --source "Federal Plain Language Guidelines" \
  --url "https://digital.gov/guides/plain-language" \
  --license "CC0 1.0"

Run every check before you open a pull request:

npm run check

This project follows the Contributor Covenant. Report a vulnerability privately, as described in SECURITY.md.