Readme Writing

README Writing

The README is the front door of a project. Most readers arrive with one of three jobs: decide if this is worth their time, get it running, or find one specific answer. A good README serves all three in the order they appear — orientation first, action second, depth on demand. This skill is the deep reference for writing one: the required sections, the writing moves that make it scannable, the runnable-example discipline, and the failure modes that quietly drive readers away. Heavy detail lives in references/; copy-paste material in examples/; a runnable linter in scripts/.

Mental model

A README is read in three passes, and you write top-to-bottom to match them:

Optimize ruthlessly for the first 60 seconds. A reader who can't tell what the project does, or who hits a broken install command, leaves and does not come back. Everything else is recoverable; the first screen is not.

The required sections (in order)

A complete README has these, roughly in this sequence. Not every project needs every one, but you must make a deliberate choice to omit, never an accidental one.

1. Title + one-line description — what it is and who it's for, no preamble, no "Welcome to".
2. Badges — build status, version, license, coverage. Signal of health, scanned in a glance.
3. Why / value — the problem it solves, optionally a screenshot, GIF, or short result example.
4. Install — the minimal commands to get the dependency in place, copy-pasteable and verified.
5. Quickstart / Usage — the smallest end-to-end example that produces a visible result.
6. Configuration — options, env vars, defaults — as a table, not prose.
7. Contributing — how to set up a dev environment and the contribution flow (link to CONTRIBUTING.md).
8. License — the SPDX name and a link to the LICENSE file.

The full anatomy, with what to include and skip per project type (library vs CLI vs service vs framework), is in references/anatomy.md.

Lead with what and why

The opening is the most-read and most-botched part. State the what in one line, then the why. Do not bury it under a logo, a table of contents, or a wall of badges.

# Forge

Forge builds and maintains AI agents and skills through a self-improving maker/checker loop —
so your automation gets better on every run instead of rotting.

[![build](https://img.shields.io/badge/build-passing-brightgreen)]()
[![npm](https://img.shields.io/npm/v/forge)]()
[![license](https://img.shields.io/badge/license-MIT-blue)]()

Within two sentences the reader knows the category (agent tooling), the mechanism (maker/checker loop), and the payoff (self-improving, no rot). Compare the anti-pattern: "Welcome to Forge! Forge is a powerful, flexible, modern, enterprise-grade platform for..." — three adjectives and zero information. See references/writing-style.md for the full list of filler phrases to cut.

The 60-second quickstart

The quickstart is a promise: paste these commands and see it work. It must be copy-pasteable, self-contained, and actually run — test it in a clean checkout before shipping.

## Quick start

\`\`\`bash
git clone https://github.com/acme/forge && cd forge
npm install
npx forge audit ./my-project    # prints a scorecard
\`\`\`

You should see a table of findings within a few seconds. Next, try `forge fix` to apply them.

Rules that make or break it:

• Show the expected output, or at least describe it ("you should see a table"). A command with no visible result leaves the reader unsure it worked.
• One happy path only. Do not branch into "if you use yarn / pnpm / bun" in the quickstart — pick one, link the rest. Optionality kills momentum.
• No undeclared prerequisites. If it needs Node 20+, a running Postgres, or an API key, say so before the commands, not in a stack trace the reader hits later.

Show, don't tell

Prose describing behavior ages badly and is hard to trust. A runnable example is self-verifying — either it works or the reader sees it break. Prefer concrete examples over adjectives at every turn.

## Usage

\`\`\`js
import { Forge } from 'forge';

const forge = new Forge({ target: './my-project' });
const report = await forge.audit();

console.log(report.score);        // 87
console.log(report.findings[0]);  // { rule: 'no-secrets', severity: 'high', file: '.env' }
\`\`\`

The inline comments showing return values turn a snippet into documentation. This is why examples/ ships a good-readme-example.md and a README.template.md you can adapt — see Files in this package.

Make it scannable

Readers skim, they do not read. Structure for the eye:

• Headings every few paragraphs so the table of contents (and the reader's scroll) has anchors.
• Tables for anything with structure — config options, CLI flags, comparison matrices. A 5-row table beats two paragraphs of "the --depth flag controls...".
• Code blocks for anything runnable, always fenced with a language for syntax highlighting.
• Short paragraphs. One idea each. Walls of text are scrolled past, not read.
• Link out for depth. The README is a launchpad, not the manual. Link to docs/, the wiki, or a hosted site rather than inlining a 200-line configuration reference.

Audience calibration matters too: a library README assumes a developer who will read code; a CLI README assumes someone who wants commands; an end-user app README assumes neither. Details in references/anatomy.md.

Common pitfalls (failure modes)

• The broken quickstart. Commands that don't run in a clean checkout — stale flags, missing install step, undeclared env var. The single most damaging README bug. Test it cold.
• Burying the lede. A logo, badge wall, or 30-line table of contents before the one-line description. The reader scrolls looking for "what is this" and gives up.
• Adjective soup. "Powerful, flexible, blazing-fast, enterprise-grade" — these are unfalsifiable and information-free. Replace with a concrete capability or benchmark.
• Telling instead of showing. Paragraphs describing the API instead of a code block using it.
• The novel. Inlining the entire configuration reference, changelog, and architecture doc. The README becomes unmaintainable and unscannable. Link out.
• Drift. Examples that no longer match the current API because nobody re-ran them. Treat README snippets as testable artifacts (see scripts/lint-readme.mjs and references/maintenance.md).
• No license. A repo with no LICENSE is, legally, all-rights-reserved — nobody can safely use it. Always state the license.

When NOT to write a heavy README / trade-offs

A README is not the right home for everything, and more is not better:

• Deep, multi-page docs belong in a docs/ site (Docusaurus, mdBook, ReadTheDocs), not the README. A README that scrolls for ten screens has failed at being a front door. Link instead.
• API reference that's mechanically derivable (TypeDoc, rustdoc, Sphinx) should be generated, not hand-written into the README where it will drift.
• Internal-only / throwaway repos may need only a title and a one-line run command — don't gold-plate a README nobody outside the team will read.
• Monorepos often want a thin root README that routes to per-package READMEs, rather than one giant document trying to cover every package.

The trade-off is always completeness vs. scannability. When they conflict, scannability wins in the README and completeness moves to linked docs.

Files in this package

• references/anatomy.md — full section-by-section anatomy + what to include per project type
• references/writing-style.md — voice, filler phrases to cut, scannability techniques, badges
• references/maintenance.md — keeping a README honest: testing snippets, drift, automation in CI
• examples/README.template.md — a fill-in-the-blanks starter README with every required section
• examples/good-readme-example.md — a complete, worked example README for a sample CLI tool
• scripts/lint-readme.mjs — runnable Node check that a README has the required sections + a quickstart code block

Pairs with the documentation-structure skill (for organizing the wider docs/ tree the README links into) and the technical-writer agent (for tightening prose and voice). Also see the changelog-writing skill for the companion CHANGELOG.md.