AI_Workflow

Priming Patterns

How to load and reload Claude’s context within a session. The commands you use to say “read my rules, then shut up and wait for work.”

The commands

Command Lives in What it does When to use
/prep .claude/commands/prep.md Reads CLAUDE.md, outputs Read project context. Ready. Start of every session
/prime [role] .claude/commands/prime.md Reads CLAUDE.md + role-specific file + inventory Advanced: when you have role-based terminals (Chapter 05)
/clear Built-in Wipes conversation, keeps repo context When conversation is getting long
/compact Built-in Summarizes and keeps summary Almost never — see below

The /prep command — the 12-line gold standard

Here’s the entire /prep command:

---
description: Prepare to discuss this repository (non-verbose)
scope: personal
---

Read and internalize CLAUDE.md without summarizing.

## Instructions

1. Read `CLAUDE.md`
2. Do NOT summarize or explain

## Output

> "Read project context. Ready."

That’s it. Twelve lines. No fluff. And it’s exactly right.

Why it’s exactly right:

  1. “Do NOT summarize or explain” — This is the critical line. Without it, Claude reads your 200-line CLAUDE.md and produces a 400-token summary (“I see this is a Next.js project using…”). That summary burns context AND tells you nothing you didn’t already know. The output rule prevents the waste.

  2. A single confirmation string — "Read project context. Ready." is 4 words. You know the read happened. Claude is now armed. Move on.

  3. No error handling — If CLAUDE.md is missing, Claude will say so. You don’t need the command to handle edge cases.

  4. Scope: personal — This is a global command, not per-repo. You want it in every project without copying.

The flow: /prep every session

Here’s the complete opening sequence for any work session on a repo:

cd ~/repos/my-project
docker compose up -d        # start dev env in background
claude                       # launch Claude Code

At the Claude prompt:

/prep

Wait for Read project context. Ready. Then ask your real question. Every session. Don’t skip it. The 2 seconds of /prep prevents 30 minutes of Claude getting things wrong for lack of context.

Mid-session: when context fills up

Claude Code shows a context-usage meter (usually bottom-right or in the status line). When it gets to ~70-80%, you have two choices.

The right choice: /clear + /prep

/clear

(conversation wiped)

/prep

(CLAUDE.md re-read, ready)

Now continue working. You’ve lost the conversation history but kept the repo context. The chat starts empty but Claude still knows your project rules.

The trap: /compact

/compact takes your conversation so far and replaces it with a summary. Sounds efficient. It’s a trap. Three reasons:

1. Compacted summaries preserve mistakes. If Claude misread a file 40 messages ago and later corrected itself, the compacted summary often keeps the misreading as “a fact Claude learned about this project.” You now have a broken assumption baked into the context.

2. Compacted summaries lose file state. Claude knew file X had contents Y five messages ago. After compact, Claude knows “file X is related to the auth flow” and has to re-read it. The re-read blows up the context just as bad as /clear would have — except now you also have lossy half-memories polluting the rest of the session.

3. Compacted context is sticky. Once you compact, the lossy summary stays for the rest of the session. A bad compact at message 50 can ruin messages 51-200.

The war story: TurtleWolfe debugged a compacted session where Claude kept patching a test file at src/components/Button/Button.test.tsx. The file had moved to src/__tests__/Button.test.tsx weeks earlier. The compacted summary from an older session still said “tests live alongside components” — Claude kept looking in the wrong place, getting it wrong, apologizing, and looking in the wrong place again. Three cycles. /clear + /prep would have solved it in one message.

Rule: /clear + /prep is almost always the right move. /compact is for very specific situations (long refactor discussions where you genuinely need the history, and you accept the pollution risk). If you’re not sure, /clear.

Advanced: role-based priming (Chapter 05 territory)

When you start running parallel terminals with assigned roles (Chapter 05), /prep isn’t granular enough. You use /prime [role] instead:

/prime developer
/prime architect
/prime security

Each loads CLAUDE.md plus a role-specific file plus any inventory files the role needs. The full story is in Chapter 05 — Advanced Orchestration.

For Day 1: ignore /prime. Use only /prep.

The golden rule

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

Live by this and you will save dozens of hours per month.

This site is open source. Improve this page.