skill

three-pass-writing

Claude Code skill: write your product's copy with a three-pass pipeline (Sonnet draft, Opus critique, Sonnet rewrite) grounded in a read-only audit of your own repository.

RAAAAAGEEEEE/three-pass-writingcommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/RAAAAAGEEEEE/three-pass-writing itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 22, 2026

Three-Pass Writing

A Claude Code skill that writes your product's copy the way a good team does it: someone drafts, someone else tears it apart, then it gets rewritten.

Three passes, three subagents, two models:

Pass	Model	Job
1	Sonnet	Draft
2	Opus	Critique, never rewrite
3	Sonnet	Apply the fixes, change nothing else

The point is not "more passes". It is that a model reviewing its own output defends it, and a model reviewing someone else's cuts it.

How it works (Comment ça marche)

Three steps for a beginner, and no pasting the repository into the conversation:

Install the skill once: git clone https://github.com/RAAAAAGEEEEE/three-pass-writing ~/.claude/skills/three-pass-writing (available in all your projects), or the same clone into .claude/skills/three-pass-writing at the root of a project (that project only). In Windows PowerShell, replace ~ with $env:USERPROFILE. The plugin route and details are in docs/INSTALLATION.md.
Ask for it: from your product's repository, just write "write a cold email to agency owners", or type /three-pass-writing followed by your request.
Follow along: the first run audits the repo read-only and builds .claude/writing-context.md with you; then Sonnet drafts, Opus critiques and Sonnet rewrites.

This is how every Claude Code skill works: a folder with a SKILL.md in ~/.claude/skills/<name>/ (personal) or .claude/skills/<name>/ (project); Claude loads it automatically when your request matches its description, and /<name> runs it by hand. [official: skills, checked 2026-10-05]

The problem it solves

Ask any model to write a cold email about your product and you get fluent copy that promises a feature you never shipped, quotes a number nobody measured, and opens with a line that would work for any product on earth.

This skill starts by reading your repository — what actually ships, what is still a TODO, which numbers have a source — then asks you the handful of questions the code cannot answer, and writes the result into a context file. Every piece of copy afterwards is grounded in that file. A claim with no source becomes [[to confirm: ...]] instead of a confident invention.

Who it is for

Solo founders and small teams who write their own outreach, posts, landing copy and release notes, and who already work in Claude Code.

Status

v1.5.1, beta. Manifests validated with claude plugin validate --strict, installation reproduced from a clean state (see docs/INSTALLATION.md), and the full pipeline run end to end on a live product — bootstrap, three passes, and the review loop stopping where it should. The prose quality depends on the context file the bootstrap builds with you: a thin bootstrap produces thin copy. Read docs/LIMITATIONS.md before relying on it.

Requirements

Claude Code v2.1.220 or later (checked with claude --version)
Access to both Sonnet and Opus on your plan — the pipeline is built on the split
A repository for the product you are writing about (or you get the brainstorm alone, without the audit)

Install

As a plugin, from this repo used as its own marketplace:

claude plugin marketplace add RAAAAAGEEEEE/three-pass-writing

claude plugin install three-pass-writing@three-pass-writing

Or as a personal skill — clone into your skills folder, where the bundled .claude-plugin/plugin.json makes Claude Code load its three subagents too:

git clone https://github.com/RAAAAAGEEEEE/three-pass-writing ~/.claude/skills/three-pass-writing

Both paths are detailed, with the Windows variant and the verification step, in docs/INSTALLATION.md.

Quickstart

From the repository of the product you want to write about:

claude

Then, in the session:

/three-pass-writing write a cold email to agency owners who just posted a hiring ad

First run in a project triggers the bootstrap: a read-only audit of the repo, a short brainstorm, and a .claude/writing-context.md file. Every later run reuses it.

What a run looks like

Audit — 11 files read
PRODUCT     Turns a repo into grounded marketing copy — README.md:3
SHIPPED     three-pass pipeline, repo audit, 7 format profiles — SKILL.md:78
NOT SHIPPED variant A/B tracking — docs/LIMITATIONS.md:24
PROOF       0 external calls, 3 subagents per piece — SKILL.md:96
GAPS        who exactly is the reader? what may be quoted publicly?

[brainstorm: 4 questions, defaults proposed]

Pass 1 (sonnet) → draft, 3 alternative openings, 6 facts sourced
Pass 2 (opus)   → VERDICT: REWRITE — 1 P0, 2 P1
                  P0 "cuts writing time by 60%" → no source anywhere
Pass 3 (sonnet) → final, 118 words

FINAL
<copy-ready text>

STILL OPEN
[[to confirm: the 60% figure — measured, or dropped?]]

Illustrative example, not a recorded transcript.

How the pipeline works

SKILL.md is the entry point and stays small. It routes to reference files that load only when needed: the audit procedure, the brainstorm protocol, the pipeline mechanics, the review rubric, the format profiles. The three passes run as subagents defined in agents/, each pinned to its model.

Details in docs/ARCHITECTURE.md.

Configuration

Everything project-specific lives in one file you own and can edit by hand: .claude/writing-context.md. The skill generates it, you correct it, your corrections win.

See docs/CONFIGURATION.md.

Security and privacy

The skill reads your repository and never opens .env, credential files, keys, or customer data. It makes no network call, sends no email, publishes no post, and commits nothing on its own. The context file it writes lands in your repo, so it is built to hold no secrets.

Full posture in docs/PRIVACY_AND_SECURITY.md.

Limits

It does not measure anything. No open rates, no A/B results, no analytics.
It cannot verify a claim that exists nowhere in your repo — it flags it and asks you.
Three subagent calls per piece, one on Opus. Not the tool for fifty subject-line variants.
Cold outreach compliance guidance is not legal advice.

The honest list is in docs/LIMITATIONS.md.

Roadmap

Non-contractual, in rough order: a /writing-context refresh path that diffs the repo since the last audit; per-format sample libraries derived from the user's own accepted pieces; running the eval set in evals/ and extending it to the review rubric.

Evals

Four cases in evals/ check the invariants that matter most: an unsourced number is flagged, no em dash or emoji in the copy, nothing is ever sent, and the bootstrap runs before any writing. They follow the claude plugin eval format, which is early access. They have not been run yet, see docs/LIMITATIONS.md.