The must-do habits — Tutorials 100

Tutorials are use-case driven, not feature driven. Each one starts with a situation you’ll recognise, names the TokenSquirrel score it improves, and tells you exactly what to change.

Background reading. Every habit on this page maps to a lever in Part 2: The Practitioner’s Toolkit — /clear and /compact are context management, model switching is model selection, the CLAUDE.md trim is the skills-vs-CLAUDE.md architectural choice. Read it once for the reasoning; come back here for the action.

Level 100 covers the five habits that move the needle most. If you only ever do the things on this page, your overall grade will go up by a letter.

1. End every task with `/clear`

The score this fixes: /clear frequency, session hygiene, cache efficiency.

The situation: you finished refactoring the auth middleware. You’ve got a working diff, the tests pass, you’re done. You leave the session open and start asking about the unrelated bug a teammate just flagged.

Why this hurts: the auth context you just built up is now riding along on every prompt about the unrelated bug. Tokens you don’t need, paid for on every turn, and Claude’s attention split across two threads.

What to do: when a task is done, type /clear. Start the next task in a fresh session. This is the single highest-leverage change you can make — it improves three scores at once.

How to verify: run npx tokensquirrel claude --detail after a week. The /clear frequency score should be at least a B, and your average session length (under “Session hygiene”) should drop noticeably.

2. Write a focused first prompt

The score this fixes: session hygiene, tool efficiency.

The situation: “Hey, can you help me with this thing where the user can sometimes log in but other times the session expires weirdly and also I think the password reset is broken?”

Why this hurts: Claude responds with clarifying questions. You answer them. Twenty messages later you’re still scoping. The whole session is conversation, not work.

What to do: before pasting, write the prompt as a one-paragraph spec. State the goal, the file you think is involved, and the symptom. Example: “Investigate why /api/auth/refresh returns 401 for users whose session is older than 24h. Suspected cause: the JWT refresh logic in src/auth/refresh.ts. Reproduce with the test in auth.test.ts:142.”

How to verify: tool efficiency rises (more action, less chat) and session hygiene improves (fewer turns to closure).

3. Keep CLAUDE.md lean

The score this fixes: CLAUDE.md weight.

The situation: every project memory you ever wrote down has accumulated in CLAUDE.md. It’s now 3,000 words. It’s injected into every prompt, on every turn, in every session.

Why this hurts: the bigger the CLAUDE.md, the more context tokens go to instructions before any of your prompt is even read. You’re paying for it every turn.

What to do:

Run npx tokensquirrel claude --detail and check the CLAUDE.md weight score.
If it’s below B: open your project’s CLAUDE.md and your user-level ~/.claude/CLAUDE.md.
Delete anything that hasn’t been load-bearing in the last month. Anything that says “we used to…” or “in the old setup…” is dead weight.
Aim for under 1,200 words total across all CLAUDE.md files.

How to verify: the CLAUDE.md weight score moves up. The next time Claude is “weirdly cautious” about something, ask yourself if a stale rule in CLAUDE.md is steering it.

4. Allowlist the commands you run all the time

The score this fixes: indirectly improves tool efficiency and reduces friction (fewer permission prompts means more flow).

The situation: every time Claude wants to run npm test or git status, you get a permission prompt. You always say yes. You’ve said yes ten thousand times.

What to do: add the safe, read-only stuff to your settings allowlist. The fewer-permission-prompts skill scans your transcripts and proposes an allowlist for .claude/settings.json; or add them by hand:

{
  "permissions": {
    "allow": [
      "Bash(npm test:*)",
      "Bash(git status)",
      "Bash(git diff:*)",
      "Bash(git log:*)"
    ]
  }
}

How to verify: fewer interruptions in your next session. Sessions become shorter because the model isn’t blocked waiting on you to click yes.

5. Check your spend once a week

The score this fixes: /cost tracking, plus the meta-habit of actually paying attention.

The situation: you have no idea how much you spent on Claude this month until the bill arrives.

What to do: every Friday, run two commands:

npx tokensquirrel claude
npx tokensquirrel claude mot --format md > ~/notes/squirrel-$(date +%Y-%V).md

The first gives you the headline. The second is a permanent receipt — markdown you can read in a viewer, drop in a doc, or paste into Slack.

How to verify: after a month, you have four MOT reports. Compare them. The behaviours you’ve been working on should show up as better grades. If they don’t, that’s its own signal.

What’s next

Level 200 covers the important habits — the ones that don’t usually break things outright but quietly bleed tokens. Level 300 is the customisation rabbit hole: writing skills, tuning hooks, automating audits in CI.

1. End every task with /clear