AI_Workflow

Command Catalog — Priming

Commands that load context into Claude’s working memory. These are the first commands you run in every session.

`/prep` — the universal primer

Starter kit: yes (one of the 10) Scope: personal (global) File: ../starter-kit/.claude/commands/prep.md

The entire command is 12 lines. That’s not a typo — it’s the point.

What it does

Reads CLAUDE.md at the repo root
Outputs exactly one line: Read project context. Ready.

That’s it. No summary, no explanation, no “here’s what I learned.” Just the read and the confirmation.

Why it’s terse

Most users’ first instinct is to have a priming command that reads the file AND summarizes it AND highlights anything important. This is wrong. Here’s why:

The summary wastes tokens. Every word Claude generates to summarize CLAUDE.md is a word that could have been spent on real work. A 400-token summary of a 200-line file is 400 tokens of garbage in your context budget.
You already read CLAUDE.md. You wrote it. You don’t need Claude to explain it to you.
Summaries lie. A summary is always lossy. If Claude summarizes “use pnpm not npm” as “uses pnpm”, Claude just lost the prohibition. Later, when you ask “add a dependency”, Claude says npm install.
Verification is free. The one-line confirmation Read project context. Ready. tells you the read happened. You don’t need more.

Rule: Priming commands should confirm the action, not perform the action. The action is Claude holding the file in context. The confirmation is the output.

When to use

Start of every session. Every single one.
After /clear. Always re-prime after clearing.
Never after /compact. (You shouldn’t be using /compact anyway.)

When NOT to use

Mid-task. If you’re in the middle of editing a file, don’t randomly re-prime. Wait until you /clear.
Before a one-off question. If you’re asking “what does this error mean” and you don’t need repo-specific context, skip /prep. Not every conversation needs the full context load.

`/prime [role]` — role-based priming (Chapter 05 territory)

Starter kit: YES (ships with the kit, but only useful with Chapter 05 infrastructure) Scope: personal (global)

/prime is /prep on steroids — it loads CLAUDE.md PLUS a role-specific file PLUS role-specific inventory files. Only useful when you’re running multiple parallel terminals with assigned roles.

The role list from TurtleWolfe’s workflow:

operator, stw-liaison, cto, architect, coordinator, security, toolsmith,
devops, product-owner, planner, wireframe-generator, preview-host,
wireframe-qa, validator, inspector, author, test-engineer, developer,
auditor, qa-lead, tech-writer, docker-captain

Each role maps to a file like .claude/roles/toolsmith.md which tells Claude what that role is responsible for and what inventory it needs loaded.

Do not use /prime on Day 1. Use /prep.

Do use /prime in Chapter 05 when you’re running the 27-terminal tmux assembly line and need each terminal to know its specific job.

Full coverage in Chapter 05 — Advanced Orchestration.

`/prime_repositories` — multi-repo priming

Starter kit: YES Scope: personal (global)

A specialized variant for workflows spanning multiple related repos (e.g., a plugin + its host application). Hard-codes the paths of related READMEs, architecture docs, and key source files, and reads them all at once.

Use case: You’re working on an integration between two repos and need both primed simultaneously. Write a custom /prime_<project> for each such workflow.

Day 1: skip.

The rule of priming

Every session starts with /prep. Every context refill uses /clear + /prep. /compact is the last resort, not the first.

Follow this and your context never gets polluted.

02-give-it-context/priming-patterns.md — the longer explanation with the war story
../starter-kit/.claude/commands/prep.md — the command file itself
principles/context-hygiene.md — the underlying principle

This site is open source. Improve this page.