The Free Claude and Codex Skills That Improve Writing

If your AI drafts read like AI, there is a small set of free tools that actually changes the output, and they sort cleanly by the stage of the draft they act on. Pre-draft briefing is well served by free agent skills. Post-draft checking is well served by four mature open-source prose linters. The middle — while the model is actually composing — is served almost entirely by static instruction files that nothing enforces.

The full census is one table below, and every tool in it is free: anything whose writing feature sits behind a paid tier was excluded outright. The collection date and inclusion rules are in the methodology block.

Three terms carry the table. An agent skill is an instruction package a coding agent loads on demand: a folder whose SKILL.md file pairs a short machine-readable header — a name and a description — with the Markdown instructions the model reads before doing the matching task. An instruction file such as CLAUDE.md or AGENTS.md is prose the agent reads at the start of a session and follows at its own discretion. A validator gate is different from both: a deterministic, rule-based pass/fail check run against finished text, which can block a commit or a publish the way a code linter blocks a merge.

Read down the mechanism column and the shape of the market is plain. Everything deterministic sits post-draft, and none of it was born in the agent ecosystem — all four linters predate it. Everything agent-native is either a brief or the model judging itself. Which tool you reach for is a separate question from which model you run it on; our companion census of how the writing leaderboards disagree covers that side.

A census is only as citable as its inclusion rules, so here they are in full — including what was excluded, what was fetched first-hand versus summarised from search, and the limits of the absence claims this page makes.

Anthropic’s public skills repository is the natural first stop, and it is genuinely substantial — 19 skills in the listing we fetched, most under Apache 2.0, and 172,870 GitHub stars at collection. We fetched three of those SKILL.md files — internal-comms, doc-coauthoring and brand-guidelines — and the pattern is consistent: they shape the brief, and none of them validates or blocks non-conforming prose.

internal-comms is the only official skill named for writing internal documents — updates, newsletters, FAQs, incident reports. Its own instructions are a router, not a style guide:

doc-coauthoring goes further than a brief: it enforces a three-stage collaborative process — context gathering, section-by-section refinement, then a reader test that opens a fresh context window to check the document stands on its own. It asks for feedback as numbered shorthand, nudges toward cutting content after three low-change iterations, and leaves final review explicitly to the human. It is the most process-shaped writing skill in the official set, and its quality bar is a question, not a rule.

The third card deserves the emphasis. A skill literally named brand-guidelines is the one a reasonable person would expect to enforce a brand’s writing voice, and it does not — it governs typography and color. Much of the secondary coverage of this repo lists skill names without reading what each one does, which is exactly how that assumption spreads. Nothing in the official set, as fetched for this census, gates prose style — the closest candidates are a router and a process.

The stage where style would matter most is the stage with the least tooling. While the model is composing, exactly one free mechanism carries your style rules to it: the instructions it is reading at that moment. That can be a skill body or an instruction file, and the best community example of doing it well is deliberately minimal. Meng To’s skill collection — MIT-licensed, 5,668 stars at collection, and actively maintained — embeds its house style as a single line inside the SKILL.md itself, pushing longer explanations to a separate file so the instruction the model actually reads stays terse.

That is the in-draft constraint pattern at its clearest: the rule travels inside the composing context, phrased short enough to survive it. What it is not, on any tool in this census, is enforced. Nothing free and purpose-built checks the emerging draft against the rule as it is written — compliance rests entirely on the model honouring the instruction.

Practitioner commentary — opinion, not a vendor specification — is blunt about where that leaves instruction files: they carry the preferences a team can live with the model occasionally ignoring, and nothing stronger. We have catalogued the failure patterns of these files separately in our guide to CLAUDE.md config smells; unenforced writing rules are one specific smell class from that family, and the census suggests it is the class with the least free tooling pointed at it.

The deterministic tooling exists. It is mature, permissively licensed, and entirely free — and all of it predates the agent-skill ecosystem. Vale describes itself as “a linter for prose”: it runs offline, ships built-in implementations of 16 publishers’ style guides, and wires into CI through a GitHub Action, pre-commit hooks and editor language servers. Vale’s own site cites GitLab’s documentation pipeline running it as a required check across roughly 2,800 pages with 82 rules in under 20 seconds — figures from Vale’s marketing copy, not re-verified against a GitLab primary in this pass, but the integration pattern they describe is the point: a rule-based pass/fail on prose, blocking a merge the way a code linter does.

Between them the four cover complementary ground: Vale enforces a named style guide, proselint packs 70+ opinionated checks, write-good targets the tics that make prose read machine-written — passive voice, weasel words, weakening adverbs — and textlint is the pluggable chassis you assemble rules onto. All of them are consumption-side: they check text that already exists, in editors, pre-commit hooks and CI. Which is exactly why the next section’s finding matters.

There is now a vendor-neutral way to package exactly this kind of tooling for agents. Agent Plugins 1.0 reached general availability on August 12, 2026, announced in GitHub’s own changelog. We checked that changelog and one search pass of the Awesome Copilot marketplace — 48-plus plugins at the time of the pass — for a named writing, editorial or style plugin, and found none in either. That is a narrow claim about two sources on one pass, not an audit of the whole ecosystem. But it matches the rest of the census: four mature deterministic linters on one side, a portable packaging standard on the other, and no free, purpose-built bridge between them that this pass could locate.

The same narrowness applies on the Codex side. Third-party guides document that Codex CLI reads the same SKILL.md format from its own skills directory, and that a symlink can share one skill across both clients — claims we did not verify against an OpenAI or Anthropic primary in this pass. What we can say from the repos actually checked is that every Codex-compatible writing skill we found is cross-compatible Claude-ecosystem content via the shared format, rather than anything Codex-exclusive.

So where should a team that ships prose actually spend effort?

The build path, for anyone taking the fourth option, starts with our step-by-step skill-building tutorial and — for what still breaks when a skill crosses harnesses — our review of what Agent Plugins 1.0 GA still does not port. And if the underlying problem is that your pipeline produces prose nobody is systematically checking, that is a content-system design question before it is a tooling one — the kind our content engine service exists to answer, with enforcement designed in rather than bolted on.