Model or dataset
obra/the-elements-of-style avatar
obra/the-elements-of-style

the-elements-of-style by obra: eleven generated manifests and no licence file

William Strunk Jr.'s Elements of Style (1918) in markdown format for AI agents

629 stars83 forksHTMLLicense varies

At a glance

What is it?
This repository republishes the 1918 Strunk text from Project Gutenberg as an agent skill, and ships eleven harness plugin directories generated from a single file by a tool that fails the build on drift. It is the cleanest generated-manifest arrangement in this batch, and it has one conspicuous hole.
Who is it for?
Use this repository if you want a writing skill for an agent and you value the generated-manifest discipline, because one source file drives eleven adapters, generated files carry a provenance header, a manifest lists them, and a validator fails on drift rather than letting a hand edit rot. Two things to settle before you depend on it.
Can I use it commercially?
Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
Is it still maintained?
Yes. The repository last received commits 95 days ago.
What is it written in?
Mainly HTML, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • obra/the-elements-of-style629 ★ · +59
Star history
Repository2026-09-112026-11-15
obra/the-elements-of-style570629

59 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The repository and homepage fields point at a repository name that is not this one

The package manifest declares its own source location, and it disagrees with where the package actually is.

The repository field and the homepage field both give a GitHub URL under the same owner with the name `elements-of-style`. The repository this manifest ships inside is `the-elements-of-style`.

One `the-` is missing from both fields, and they are the two fields other tooling reads to attribute the package. So anything that resolves the upstream from the manifest, whether that is a package page, a dependency bot, or an audit tool, is sent to a path that is not where the code lives.

That is the sort of thing that happens after a rename, and it is harmless if the old path redirects and a permanent failure if it does not. It is also self-inflicted in a specific way: this manifest declares no dependencies, no scripts and no build tooling at all, so nothing in the package's own toolchain could have caught the mismatch.

The rest of the manifest is small and unusual. The package is named after the book rather than after the repository, its type is a module, and it declares a main entry point that resolves inside a hidden directory belonging to one specific harness.

Eleven plugin directories, one skills directory, and eleven manifests generated from a single file

The root of this repository is almost entirely adapters.

There are hidden directories for a general agent format, and separately for Claude, Codex, Cursor, Devin, Hermes, Kimi, OpenCode and Pi. There is a plain uppercase Markdown file for one harness, a JSON descriptor for another, and a plugin descriptor for a third. There is a directory holding the generated-file manifest, a configuration file, and exactly one directory with content in it: `skills/`.

So eleven or twelve integration surfaces, and one payload.

The install table in the page lists eleven rows, each naming a harness or a class of client and pointing at its own install document. Three products are covered by a single row for a marketplace-style descriptor rather than getting their own.

The page states the relationship in one sentence: every supported harness reads the skill from the skills directory, and the per-harness manifests differ only in where they point.

That is the key architectural sentence. The differences between eleven integrations are paths, not behaviour. Which means the cost of supporting an eleventh harness is a path, and the cost of getting one wrong is a harness that silently reads nothing.

Generated files carry a provenance header, a manifest lists them, and a validator fails on drift

The manifests are generated by a separate tool, and the page is unusually firm about what that means for contributors.

The generator takes one configuration file plus the skill and produces every harness manifest, every bootstrap hook and every install document. The command is a single invocation through the package runner:

code
npx everyharness generate

Then the instruction: never edit a generated file by hand, because the validation command fails on drift and the next generate overwrites your change anyway.

Two supporting mechanisms make that enforceable rather than aspirational. Generated files carry a header saying they were generated and by what. And a manifest inside a hidden directory lists all of them, so the set of generated artefacts is itself a tracked file.

That combination is the difference between a convention and a guarantee. The header tells a reader not to trust the file as a source of truth. The manifest tells the tool which files to check. The validator turns an unedited hand change into a failing check instead of a diff someone notices in review.

The cost is on the other side. The whole multi-harness surface is now a function of a tool this repository does not contain, invoked at generation time through the package runner, and nothing in the package manifest pins that tool's version or lists it as a development dependency. Regenerating the same configuration on two machines a year apart is not obviously reproducible.

The licence is asserted in three places and the licence file is missing

For a project whose entire content is somebody else's text, the licence status is the thing to check first, and here it is asserted three times and filed zero times.

The project metadata records no licence at all. The package manifest declares the licence as a two-word phrase meaning public domain, which is not one of the identifiers the npm ecosystem recognises, so tooling that validates that field strictly will reject it or warn. The page states that the text is public domain and gives the source.

The source attribution is the strongest part of the whole licence story. The page names the author, the publication year, and the exact Project Gutenberg item number, and links the specific HTML file rather than the catalogue page. The converted source is also committed alongside the conversion, at the repository root as an HTML file, so anyone can check the transformation rather than trust it.

And yet the root listing has no licence file.

For a 1918 book that is almost certainly correct in substance, since the work is long out of copyright and the text says so. The gap is procedural. A repository whose only asset is a republication should carry a file stating the terms, for the same reason a dependency should: so that the next person who finds it does not have to reconstruct the argument.

The reference is about 12,000 tokens, and the skill is built around not loading it

The design decision here is a context budget, and it is stated in the open.

The skill contains a list of every rule so the agent can see them at a glance, and a warning that the full reference costs roughly 12,000 tokens. The guidance is to hand the draft to a subagent when context is running short, and to open the full text only while writing or editing rather than carrying it the whole time.

That is progressive disclosure applied to prose, and it is the right shape for a reference of this size. A book-length style guide pinned into every request would cost the same tokens on every turn regardless of whether the agent was writing anything.

The reference itself is the 1918 text in three parts: a short set of elementary rules of grammar and punctuation, a longer set of elementary principles of composition, and an alphabetical guide to words and expressions commonly misused.

Two things are worth flagging about the numbers. The token cost is given as an approximation twice, once as roughly and once as about, and no measurement is published. Token counts for a given text depend on the tokenizer, so a figure for one vendor's models will not match another's, and the page does not say whose it is.

And the third part, the alphabetical misuse guide, is the part most likely to age badly. Rules about which words are commonly misused encode the usage of a specific publishing culture in a specific year, and applying them through an agent to text written now is a stronger claim than applying them to 1918 prose.

Eleven install documents exist to encode eleven path strings

Look at what the install table's rows actually resolve to and the shape of the documentation becomes clear.

Each row is a harness name and a pointer to its own file under an install directory. Eleven rows, eleven documents. And the page has already told you what those documents contain, since the manifests differ only in where they point.

So the eleven documents exist to record eleven path strings, formatted as per-harness instructions so a reader following one does not have to work out which of the other ten applies to them.

That is defensible and it is also the obvious place where this design will get expensive. The moment one harness needs a genuinely different instruction, the assumption that the difference is only a path stops holding, and there is no shared layer underneath to put it in. Eleven near-identical documents diverge into eleven unrelated ones unless someone maintains a common template.

The generated approach helps here, because the documents come from the same configuration file as the manifests. If the shared shape is expressed once in that file, divergence is a per-harness override rather than a rewrite. The page does not describe the shape of that configuration file, which is the one thing a contributor would need before attempting it.

The install table is also delimited by sentinel comments, so the table region of the readme is machine-managed while the prose around it is not.

The package declares an entry point inside a hidden directory, with no scripts and no dependencies

The manifest is short enough to read in full, and three of its fields are unusual.

The declared entry point is a file inside a hidden directory belonging to one harness, ending in a JavaScript extension. A second block declares, for a different harness, an extension written in TypeScript plus the skills directory. So the two integration mechanisms in the package point at two different languages and two different hidden directories.

There are no scripts at all. No build, no test, no lint, no prepare hook. For a repository with generated manifests and a validator command, the absence of a test script is conspicuous: the drift check the page warns about is not wired into anything this manifest can run.

There are no dependencies either, not even development ones. The generator is invoked through the package runner and therefore resolved at generation time rather than pinned by this project.

What the manifest does carry is a keyword list aimed at one ecosystem, ending with a term naming that ecosystem's package format, and the licence phrase discussed earlier.

It is worth putting in context what this package is. It is not a library anyone imports for functionality. It is a manifest bundle: a set of descriptors that tell eleven different tools where to find a folder of Markdown. The lightness is appropriate. The missing pieces are the ones that would make the generated surface safe to regenerate.

This is the 1918 text, and the questions people ask about it are about later editions

The last thing to get right is which book this is.

The repository is the 1918 text by Strunk alone. It says so in the title, in the purpose section, in the manifest description and in the source attribution, and it credits one author and one year.

That is a different artefact from the book the name usually refers to. The later revisions added material, reorganised the rules and were attributed to two authors, and the book's current reputation rests largely on those revisions and on the twentieth-century writing advice they contain. Someone who asks for this title, or whether it is still relevant, or whether it is a grammar book, is almost certainly asking about the later work.

This repository does not contain it, does not mention it, and makes no claim about it. Its scope is stated precisely: the 1918 edition, converted from a specific Project Gutenberg file, packaged as an agent skill.

So there are two separate questions hiding under one name. Is the 1918 text worth reading on its own terms, and is it worth wiring into an agent that writes your documentation. The second question has an answer the page gives you indirectly, by shipping it as a skill with a token budget and a list of what it covers. The first is a matter of taste about a century-old text on mechanics and composition, and the repository does not pretend to have settled it.

Editorial conclusion

Use this repository if you want a writing skill for an agent and you value the generated-manifest discipline, because one source file drives eleven adapters, generated files carry a provenance header, a manifest lists them, and a validator fails on drift rather than letting a hand edit rot. Two things to settle before you depend on it. First, the licence situation: the text is public domain and the provenance names its Project Gutenberg source precisely, but there is no licence file in the repository and the package manifest uses a non-standard licence string, so add your own file if you intend to redistribute. Second, which book you actually want: this is the 1918 text by Strunk alone, not the later revision most people mean when they ask for this title.

Frequently asked questions

What is in the obra the-elements-of-style repository?

William Strunk Jr.'s 1918 text packaged as an agent skill, consisting of a skill file listing the rules and when to apply them, and a reference containing the complete 1918 text at roughly 12,000 tokens, converted to markdown from a Project Gutenberg source.

Which harnesses does the-elements-of-style install into?

The install table lists eleven, including Claude Code, Cursor, Codex, Devin CLI, Kimi Code, Gemini CLI, OpenCode, Pi, Hermes Agent, Agent Plugins 1.0 clients, and a marketplace-style descriptor covering three more products. Every harness reads the skill from the same skills directory, and the per-harness manifests differ only in where they point.

How are the-elements-of-style manifests maintained?

They are generated by a separate tool from a single configuration file, along with the bootstrap hooks and install documents, and the page instructs contributors never to edit a generated file by hand because a validation command fails on drift. Generated files carry a provenance header and a manifest lists them all.

Does this repository contain the Strunk and White edition?

No. It contains the 1918 edition by Strunk alone, credited to one author and one year, converted from a named Project Gutenberg item. The later revisions are not mentioned anywhere in the repository, which states its scope as the 1918 edition only.

What licence does the-elements-of-style use?

The text is stated to be public domain and attributed to a specific Project Gutenberg source, but the project metadata records no licence, the package manifest uses a non-standard licence phrase, and there is no licence file among the repository's root entries.

How does the-elements-of-style handle the reference's token cost?

The skill lists the rules at a glance and warns that the full reference costs roughly 12,000 tokens, suggesting a draft be handed to a subagent when context is short. The full text is opened only while writing or editing rather than being carried for the whole session.

Official sources

  1. Issues
  2. obra/the-elements-of-style on GitHub
  3. README
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/obra-the-elements-of-style.svg)](https://hysenlabs.com/projects/obra-the-elements-of-style)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/obra-the-elements-of-style.svg)](https://hysenlabs.com/projects/obra-the-elements-of-style)

Was this useful?