SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsdocwriter-style-generator

docwriter-style-generator

Agents & models

View repository

Updated 10-15 21:11

About this skill

Build a writing style skill

This skill is a generator. It reads your writing, measures word-level habits with the bundled analyzer, runs a multi-agent pass, checks each habit with you, and writes a separate my-writing-style skill that Claude Code and Codex pick up automatically.

Write the same files to both homes:

~/.claude/skills/my-writing-style/   ← Claude Code
~/.agents/skills/my-writing-style/   ← Codex

If either home already has SKILL.md, ask:

AskUserQuestion:
  header: "Style skill"
  question: "You already have a writing style skill. What would you like to do?"
  options:
    - label: "Add a source"
      description: "Keep the current habits and samples; add another piece of writing"
    - label: "Update propositions"
      description: "Re-read the existing sources and refresh the habits"
    - label: "Rebuild from scratch"
      description: "Start over with new writing samples"
    - label: "Keep it"
      description: "Nothing to do — your style skill is already active"

If they choose "Keep it", stop.

If they choose Add a source, skip to §1 but only collect the new piece. Write it into sources/ next to the ones already there. Then do §2–4. Keep every already-confirmed proposition. Only calibrate habits that are new or that the new source changed.

If they choose Update propositions, skip §1. Re-read sources/ and references/propositions.json, then do §2–4. Keep a confirmed proposition when it still has three verbatim examples. Calibrate only new or changed habits.

If they choose Rebuild from scratch, continue from §1 and replace sources/.

1. Gather sources

Ask where their writing is:

AskUserQuestion:
  header: "Sources"
  question: "Where is your writing? I need 3–5 pieces (1000+ words each)."
  options:
    - label: "Local files"
      description: "I'll give you file paths or globs"
    - label: "URLs"
      description: "Blog posts, articles, docs — I'll give you links"
    - label: "I'll paste it"
      description: "I'll paste text directly"
    - label: "Mix"
      description: "Some files, some URLs, some pasted"

Based on their answer, ask for the paths/URLs/text. Label each source by what the user calls it.

Sources must be the author's words, verbatim. Do not paraphrase, summarize, tidy, or rewrite. The later analysis quotes these files word for word — a summary is the wrong sample.

Do not use WebFetch to produce a stored source. WebFetch often returns a digest. This skill ships scripts/save-source.mjs next to this file. Run it once per piece so the write is a copy, not a rewrite:

# local file
node scripts/save-source.mjs --label essay --file /path/to/essay.txt

# pasted text
node scripts/save-source.mjs --label talk --stdin
# then pipe or type the paste; do not reword it

# URL — fetches the page and strips nav/chrome only
node scripts/save-source.mjs --label post --url https://example.com/essay

The script writes the same file to both homes:

~/.claude/skills/my-writing-style/sources/<label>.txt
~/.agents/skills/my-writing-style/sources/<label>.txt

After each save, show the user the script's First: / Last: lines and word count, and ask whether that is their writing. If they say no, or if a URL extract errors as too short, ask for a local file or a paste of the full piece. Do not invent the missing text.

On a rebuild, replace sources/ so leftover pieces from the last run do not stay in the sample. When adding a source, leave the existing files alone.

After gathering, confirm:

AskUserQuestion:
  header: "Sources"
  question: "I have N pieces. Ready to analyze, or add more?"
  options:
    - label: "Analyze these"
      description: "Start the style analysis"
    - label: "Add more"
      description: "I have more writing to add"

2. Analyze

Do not jump straight to impressions of word choice. Measure first, then read.

2a. Word-level measurements

This skill ships scripts/analyze-style.mjs next to this file. Run it on every file in the style skill sources/ folder in one invocation, and write the report to references/metrics.json in both homes:

node scripts/analyze-style.mjs \
  --input ~/.claude/skills/my-writing-style/sources/essay.txt \
  --input ~/.claude/skills/my-writing-style/sources/talk.txt \
  --output ~/.claude/skills/my-writing-style/references/metrics.json \
  --words \
  --measured

Copy that metrics.json to ~/.agents/skills/my-writing-style/references/ as well.

--words keeps word-level scores. --measured drops zeros. A zero is not evidence of a habit — do not invent absences from missing metrics.

The report is a hint for where to look: plain vs complex words, formal vs casual, concrete vs abstract, contractions, hedges, signature phrases. Never copy a rate or score into a proposition.

Tell the user you finished the word-level measurements, then start the multi-agent pass.

2b. Multi-agent pass

Read all the sources. Find habits at three levels by running three parallel Agent forks:

  • Words and phrases — what words does this person reach for? Plain or complex, concrete or abstract, formal or casual. Contractions, hedges, signature phrases. Use the measurements from 2a here.
  • Sentences — how are sentences built? Length variation, openers, coordination vs. subordination, punctuation rhythm, lists, parentheticals.
  • Passages — how does the writing hang together? Paragraph transitions, parallelism, reader address, first person, stance, how other voices (quotes, citations) are handled.

Then merge: drop duplicates, drop vague advice any writer follows, drop anything that leans on boilerplate rather than prose. Reword anything that sounds like a linguistics lecture into plain advice. Prefer fewer sharp habits over many soft ones.

Tell the user how many habits you found before moving to calibration.

What a proposition looks like

{
  "statement": "the habit, one sentence",
  "instruction": "what to do when writing (imperative)",
  "examples": ["verbatim passage 1", "verbatim passage 2", "verbatim passage 3"],
  "focus": ["the key sentence in passage 1", "...", "..."],
  "contrast": {
    "passage": "20-60 word verbatim excerpt that shows the habit",
    "rewritten": "same passage rewritten without the habit"
  }
}

Rules that matter:

  • Examples must be verbatim. Copy word for word from the sources/ files. Each example must appear as a contiguous substring of a saved source. No invention, paraphrase, or tidy-up. If you cannot find three real examples, drop the proposition.
  • Examples are passages, not sentences. Quote 3-4 consecutive sentences so the reader sees the habit in context. Set the focus to the one sentence the habit is actually about.
  • Contrasts are short. 20-60 words. The writer will see both versions side by side and pick which sounds like them. The rewrite must be competent prose — the question is preference, not quality.
  • One habit per sentence. No two propositions should claim the same focus sentence or contrast passage. If they do, merge them.
  • Plain language. No "register," "discourse markers," "rhetorical moves," "modality." Say what to do in the sentence.

3. Calibrate

For each proposition, show the writer its contrast pair. Use AskUserQuestion with previews. Randomize which slot (A or B) gets the original vs. the rewrite. Batch 3–4 propositions per question to keep it moving.

AskUserQuestion:
  header: "Voice check"
  question: "Which sounds more like your writing?"
  options:
    - label: "A"
      preview: <candidate A text>
    - label: "B"
      preview: <candidate B text>
    - label: "Neither"
      description: "Both miss the mark — skip this one"

Writer picks original → keep. Writer picks rewrite or neither → drop.

After all calibrations, report how many survived and ask to confirm:

AskUserQuestion:
  header: "Results"
  question: "N out of M habits confirmed. Save as your writing style?"
  options:
    - label: "Save it"
      description: "Write the skill for Claude Code and Codex"
    - label: "Start over"
      description: "Discard and try with different samples"

4. Write the my-writing-style skill

Write a complete skill to both homes. Create the directories if they do not exist. The two copies must match.

SKILL.md

---
name: my-writing-style
description: >-
  Apply the learned writing style when drafting or revising prose. Follow these
  habits where they fit; ignore any that would make the sentence worse.
---

# Learned author style

These are tendencies learned from a few pieces of writing, not rules. Follow
them where they fit; ignore any that would make the sentence worse. A draft
that mechanically hits every instruction reads like an imitation.

Do not copy subject matter, facts, or turns of phrase from the examples —
only how the sentences are built.

The user's own rules and requests always win over these instructions.

## <level heading>

* <instruction>

  > <example passage — **bold** the focus sentence>

  > <another example>

  > <another example>

(repeat for each proposition, grouped by level: words, sentences, passages)

Group propositions by level. Three longest examples per proposition, longest first. Bold the focus sentence in each.

references/propositions.json

The full proposition objects as a JSON array.

references/examples.md

All examples grouped by proposition statement, focus sentences bolded.

references/source-manifest.json

[{ "label": "...", "wordCount": 123 }] for each source used.

references/metrics.json

The analyzer report from step 2a.

scripts/

Copy this skill's scripts/ folder into the generated skill so later passes can re-run the word-level measurements:

scripts/analyze-style.mjs
scripts/save-source.mjs
scripts/style-metrics.mjs
scripts/style-metric-registry.mjs
scripts/style-data.json

After writing, tell the user their style skill is active and will load automatically in Claude Code and Codex. They can run this generator again to add a source or update the habits.