Writing Skills Inventory
Fourteen Claude Code skills that write, shape, or edit text. This page says what each skill does, which skill to use for which job, and where two skills do similar work. Inventory taken 2026-09-17 from ~/.claude/skills/.
1. Pick a skill: a one-table guide
Find the row that matches what you want to do. The right column is the skill to type or ask for.
| I want to… | Use |
|---|---|
| Write a nonfiction book (how-to, self-help, business), even with no topic or notes yet | book-writing |
| Find a book topic, or record what I know, when I have never written anything | author-interview (book-writing calls it for you) |
| Publish a finished book on Amazon KDP | kdp-publishing |
| Write a whole graduate textbook in maths, science, or engineering | stem-textbook |
| Write one rigorous lesson, tutorial, or textbook chapter (usually HTML with maths) | rigorous-explainer |
| Check a technical chapter: are the claims proved, the terms defined, the numbers right? | science-editor |
| Make a draft sound less like AI and more like me, or check whether it reads as AI | no-ai-slop |
| Make every reply short, active, and literal, like a maintenance manual | asd-ste100 |
| Collect raw ideas for an article, with no structure yet | /writing-fragments |
| Turn a pile of notes into an article, one paragraph at a time | /writing-shape |
| Turn a pile of notes into an article by choosing the next step at each point | /writing-beats |
Write or edit a skill, CLAUDE.md, or AGENTS.md | writing-for-agents |
| Document a codebase: README, onboarding, API and architecture docs | technical-docs-writer |
| Record how one build session went, including the failures | dev-journey |
2. The five families
The skills fall into five groups by the kind of text they produce. Skills in one group often call each other. Skills in different groups rarely do.
| Family | Skills | Output |
|---|---|---|
| Nonfiction book | book-writing, author-interview, kdp-publishing | A general-audience book, from topic to Amazon listing |
| STEM teaching | stem-textbook, rigorous-explainer, science-editor | Graduate-level chapters and books with proofs, figures, and exercises |
| Article craft | writing-fragments, writing-shape, writing-beats | One essay or article, built with the author in conversation |
| Prose style | no-ai-slop, asd-ste100 | No new content. They change how existing text or replies read |
| Technical documents | technical-docs-writer, writing-for-agents, dev-journey | Docs about software, documents that agents read, and build records |
3. Three pipelines: skills that hand work to each other
Nonfiction book
book-writing runs the whole book. When the author has no topic, no notes, no writing sample, or no story for a section, it calls author-interview (at its §0, Phase B, §5.2, or an [AUTHOR EXPERIENCE NEEDED HERE] marker). When the manuscript passes its Definition of Done (§22), §23 hands off to kdp-publishing.
STEM textbook
stem-textbook plans the whole book and keeps its state. For the craft of each chapter it reads rigorous-explainer (template, check scripts, teaching rules). For review of each drafted chapter it invokes science-editor. Its own table names the split: plan and state here, chapter craft there, rigor review in the editor.
Article
The fragments skill is the explore step: it collects raw pieces into one markdown file and imposes no order. The shape and beats skills are two exploit steps: each reads that file as a fixed pile and builds a separate article from it. Pick one of the two, not both, for one article.
The style skills sit outside all three pipelines. no-ai-slop can edit the output of any pipeline. asd-ste100 changes Claude's own replies, not your manuscript.
4. The 14 skills, one card each
Line counts are for SKILL.md only. Tags: type /name means Claude cannot start the skill by itself; you must type its name.
book-writing
Job: co-author a nonfiction book with a human author. The AI plans and drafts; the author keeps control of the thesis, sources, structure, voice, examples, and final judgment.
How: §0 starts from nothing (plain words, a book/ folder, an intake interview). Then: market check and positioning (Phase 0), book specification (A), sources (B), author persona (C), concept system (D), architecture (E), chapters (F), detailed outline (G), section-by-section drafting (H), reviews (I, J), 13 editorial passes (K), and optional production and publishing (§23). Author approval gates stop the work between phases.
Rule it will not break: it never invents author experience, sources, quotes, or reviews.
Not for: fiction, or a graduate STEM textbook.
author-interview
Job: interview a would-be author one plain question at a time and save their words, unedited, as book material in book/interviews/YYYY-MM-DD-<mode>.md.
Four modes: topic (find 2–3 topics the author can write honestly), knowledge (source notes in the author's words), story (facts for one personal story, then a draft the author approves), voice (a writing sample for an author who has none).
Limits: 6–8 questions per session. The AI's summary is kept apart from the author's answers.
kdp-publishing
Job: walk a first-time author through Amazon KDP in 8 steps: account and tax, the AI-content question, ISBN, interior format, cover, book details, price, preview and publish.
Split of work: the author does every step that needs their account, identity, or money. The AI explains, drafts text, and checks.
Caveat: its KDP facts were checked on the KDP Help Center on 2026-09-17 (for example, the 70% eBook royalty band is $2.99–$12.99 from 2026-07-07). KDP rules change; check again before relying on them.
stem-textbook
Job: co-author a graduate textbook in science, maths, or engineering, from book specification to final manuscript.
How: 8 phases with author gates A–G: specification; a graph of results plus a book-wide notation table; architecture; detailed outline; drafting by section; chapter review by science-editor; book-level passes; final integrity. State lives in BOOK-STATE.yaml so a new session can resume.
Reader it assumes: a graduate fluent in linear algebra, multivariable calculus, and probability, with close to zero exposure to the subfield.
rigorous-explainer
Job: write one self-contained technical lesson, usually an HTML chapter with MathJax.
How: start from a problem that matters, define every term and symbol before use, derive each claim under stated assumptions, interpret it with a worked example, and keep theory apart from empirical or implementation evidence. The Python scripts check links, maths rendering, proofs, prose, figures, and code. It can also publish to GitHub Pages.
science-editor
Job: act as a senior Springer textbook editor. It judges whether a graduate reader new to the subfield could learn from the pages.
The five-part test for every claim: a precise statement; every term defined; a proof in the smallest case that shows the mechanism; the rescue or limit case; a tie to something concrete (code, a figure, a measured number, a page of a paper).
Output: a verdict, blocking defects located as file:line with the full replacement text, line-level rewrites, structural notes, and what already works. It first loads a domain brief for the manuscript.
no-ai-slop
Job: two modes. Edit (default): make the smallest edit that removes AI patterns and keeps the writer's voice, then list what changed. Detect: name each pattern found, quote the line, and give a short fix, with no rewrite and no guess about whether AI wrote it.
Catches: a banned-word list (delve, leverage, robust…), binary contrasts ("not X, it's Y"), throat-clearing openers, colon reveals, importance puffery, weasel attribution, fake-profound endings, em-dash clusters, and more.
asd-ste100
Job: make every reply follow ASD-STE100 Simplified Technical English for the rest of the session: at most 20 words in an instruction, 25 in a description; active voice; one word, one meaning; no idioms; numbered procedures; a warning before its step.
Rule it will not break: it copies code, paths, file.py:42 references, numbers, and error text exactly. It stops only on "stop STE" or "normal mode".
writing-fragments
Job: the explore step. Interview you hard about a subject and append every usable piece (a sharp sentence, a claim, a story, a half-thought, a quote) to one markdown file, separated by ---. No outline, no headings.
Most valued piece: a leading word, one coined term that names the whole idea.
writing-shape
Job: an exploit step. Read the pile, agree what the reader already knows, offer 2–3 openings with different theses, then grow the article paragraph by paragraph. It argues about the form of each block (prose, list, table, callout) and never edits the pile.
writing-beats
Job: an exploit step, choose-your-own-adventure style. At each point it offers 2–3 candidate next beats (one move each: a scene, a point, a question, a twist), you pick one, and it writes only that beat.
Shared rule with writing-shape: a concept must be grounded (known before the start, or set up by an earlier beat) before a beat can use it.
writing-for-agents
Job: a reference for writing any document an agent reads: a skill, CLAUDE.md, AGENTS.md.
Key ideas: context pointers (a skill description is one), context load versus human cognitive load, what to keep in the file versus push to a linked file, clear completion criteria for each step, leading words, stating the wanted behaviour instead of a prohibition, and pruning stale lines.
technical-docs-writer
Job: write the full documentation set for a software project in 5 phases: audit, plan the structure, write, validate, keep the docs true. Its conventions are drawn from Google, Stripe, AWS, and Vercel docs.
Output: a docs/ tree (overview, quickstart, onboarding, concepts, guides, reference, architecture, troubleshooting, and a user guide when non-developers use the product). Every file complete: no stubs, no "TBD".
dev-journey
Job: write a full record of one build session, from the first prompt to the finished product: decisions, tools, agents, every error text verbatim, near-misses, costs and limits hit, and unfinished work.
Timing: write it at the end of the session, from the live transcript, before the context is summarized. Claude may offer it once at the end of a long build.
5. Overlaps and how to choose
Each pair below does related work. The question under each pair decides which skill to use.
book-writing vs stem-textbook
Both plan and draft a whole book with approval gates, and both came from the same source workflow. They differ in what the book is built from. book-writing uses an author persona, stories, word budgets, and market positioning. stem-textbook uses a graph of results, a notation table, proofs, and page or lecture-hour budgets. stem-textbook lists "nonfiction habits in a STEM book" as a common mistake, because persona and anecdote push out derivation steps.
Ask: does the book have to prove or derive its claims? Yes → stem-textbook. No → book-writing.
stem-textbook vs rigorous-explainer
Both write rigorous chapters, and stem-textbook uses rigorous-explainer's template and scripts. rigorous-explainer does one lesson with no book-wide state. stem-textbook adds the whole-book layer: a shared notation table, results that depend on earlier chapters, BOOK-STATE.yaml, and gates.
Ask: is this one standalone piece, or chapter N of a book? One piece → rigorous-explainer.
science-editor vs no-ai-slop
Both edit a draft. science-editor checks substance: claims, proofs, definitions, numbers, citations. no-ai-slop changes only how the prose reads and adds no claims. science-editor's own description says to prefer it when claims, proofs, notation, or citations must be checked.
Ask: could a sentence be wrong, or only badly written? Wrong → science-editor. Badly written → no-ai-slop. For a technical chapter, run science-editor first, then no-ai-slop on the fixed text.
science-editor vs book-writing's editorial passes
book-writing runs its own 13 passes, including Pass 7 (fact-checking and citations). Those passes suit general nonfiction. They do not apply the five-part proof test.
Ask: is there mathematics to check? Yes → science-editor.
writing-shape vs writing-beats
Both read the same kind of pile, share the grounding rule, and write a separate article file. writing-shape commits to an opening and thesis, then fills the structure paragraph by paragraph. writing-beats keeps the route open and offers a fork at every step.
Ask: do you already know the argument? Yes → writing-shape. You want to find the path as you go → writing-beats.
author-interview vs writing-fragments
Both get material out of your head by questioning you and save it to a file. author-interview saves your words unedited, question by question, for a book, and never adds a fact. writing-fragments saves pieces from both sides of the conversation, in any shape, for an article.
Ask: book or article? Book → author-interview (book-writing calls it). Article → writing-fragments.
asd-ste100 vs no-ai-slop
Both change sentences, not content. asd-ste100 is a mode: it controls Claude's replies for the whole session. no-ai-slop is a one-time edit of a draft you give it. Their rules can conflict: STE removes idioms, metaphors, and rhythm; no-ai-slop keeps the writer's humour, digressions, and cadence.
Ask: are you changing Claude's replies or your own text? Replies → asd-ste100. Your text → no-ai-slop.
technical-docs-writer vs dev-journey vs writing-for-agents
All three write documents about software work, for different readers. technical-docs-writer writes lasting docs for people who use or build the code. dev-journey writes the story of one session. writing-for-agents covers documents an agent reads. The first two meet at one file: technical-docs-writer's history/development-journey.md is the place a dev-journey document belongs, and it links an existing one rather than recreating it.
Ask: who reads it? A developer or user → technical-docs-writer. A future you, auditing one session → dev-journey. An agent → writing-for-agents.
writing-for-agents vs skill-creator (outside this inventory)
writing-for-agents covers how to word a skill. The separate skill-creator skill (not in this repo) creates, tests, and benchmarks skills.
Ask: are you writing the words, or building and testing the skill? Words → writing-for-agents.
6. How each skill starts
Eleven skills are model-invoked: Claude reads their description and can start them when your request matches, and you can also type the name. Three are user-invoked: writing-fragments, writing-shape, and writing-beats set disable-model-invocation: true. Claude cannot start them and no other skill can call them. Type /writing-fragments, /writing-shape, or /writing-beats.
Three skills are persistent or proactive. asd-ste100 stays on until "stop STE" or "normal mode". dev-journey may offer itself once at the end of a long build. book-writing and stem-textbook keep state files so work resumes across sessions.
7. Origins and licenses
| Skills | Source |
|---|---|
| writing-fragments, writing-shape, writing-beats, writing-for-agents | mattpocock/skills, MIT license. Fragments, shape, and beats sit in its skills/in-progress/ folder. |
| no-ai-slop | petergyang/no-ai-slop, MIT license (copy in skills/no-ai-slop/LICENSE). |
| book-writing | Adapted from the book workflow of Sean Dollwet (How to Write a Book with AI in 2026, YouTube), with a beginner path and the helper skills added. |
| stem-textbook | Built 2026-09-17 from the same book workflow, adapted for STEM, plus rigorous-explainer and science-editor. |
| author-interview, kdp-publishing, rigorous-explainer, science-editor, asd-ste100, technical-docs-writer, dev-journey | author-interview and kdp-publishing were written with book-writing. The other five name no upstream source in their files. |
Install
Copy a skill folder from skills/ in this repo into ~/.claude/skills/. Keep the folder name. Copy the partner skills too when a skill calls them: book-writing needs author-interview and kdp-publishing; stem-textbook needs rigorous-explainer and science-editor.