The important habits

Level 100 fixes the leaks. Level 200 is the discipline that separates “occasional Claude Code user” from “this is part of how I work.” None of these are dramatic — they’re the quiet habits that compound over months.

Background reading. Plan mode, subagents, and skills hygiene are the architectural patterns covered in Part 3: Scaling the Discipline. Model switching is the central lever in Part 2. The habits below are the practitioner version of those essays.

1. Use plan mode for anything multi-step

The score this fixes: plan mode usage.

The situation: you ask Claude to “add a search feature to the dashboard.” It starts writing code. Five minutes in, you realise it’s putting state in the wrong place and using a library you stopped using last month. You back out, but you’ve burned ten minutes and a few thousand tokens.

Why plan mode helps: in plan mode, Claude reads the codebase, drafts a plan, and asks you to approve it before touching anything. The cost of a wrong plan is one read pass. The cost of a wrong implementation is the read pass plus generating code you have to throw away.

What to do: start any task with three or more steps in plan mode. The rule of thumb: if you can’t describe the change in a single sentence, plan it.

# in Claude Code
/plan

Then describe the goal. Read the plan it produces. Push back on the bits that look wrong. Approve when it’s right.

How to verify: plan mode usage on the scorecard. Aim for 20%+ of sessions. If you’ve used Claude Code for a month and still have 0% plan mode, you’re spending more on rework than you realise.

2. Send subagents for parallel work

The score this fixes: tool efficiency, indirectly cache efficiency.

The situation: you need to understand five different parts of a codebase before you can change anything. You ask Claude one question, wait for the answer, ask the next, wait, repeat.

What to do: dispatch subagents in parallel. Tell Claude something like “explore A, B, C, and D in parallel — three separate Explore agents — and summarise findings.” Each agent works in its own context, results come back together, and your main session stays uncluttered.

The superpowers:dispatching-parallel-agents skill teaches the pattern. It’s the difference between five sequential conversations and one parallel one.

How to verify: session length drops (more done per turn) and tool efficiency rises.

3. Audit and prune skills monthly

The score this fixes: skill bloat.

The situation: you installed a plugin three months ago to try it. You’ve never used it since. Its fifteen skills load metadata into context on every session start. You don’t even remember it’s there.

What to do: once a month, run:

npx tokensquirrel claude skills --audit --fix

The --fix flag walks you through stale plugins interactively. Press y to uninstall, n to keep, s to skip. It takes two minutes and reclaims thousands of context tokens.

If you want to be more conservative, start with --unused (only flags skills never invoked) before going to the full audit.

How to verify: skill bloat score on the scorecard. If you’re at C or below, this is the easiest grade win available.

4. Switch models deliberately

The score this fixes: model switching.

The situation: you have Opus running for everything. Including grepping for a string. Including renaming a variable.

What to do: match the model to the task.

• Haiku — quick lookups, single-file edits, formatting, anything you’d type in two seconds yourself.
• Sonnet — most day-to-day coding. Good default.
• Opus — genuine reasoning: design decisions, gnarly debugging, architecture trade-offs, multi-file refactors where the choices matter.

In Claude Code, switch with /model. The cost ratio between Haiku and Opus is large enough that habit alone — using the right model — saves real money over a month.

How to verify: the model switching score (looks for ≥3 models in active rotation) and a noticeably lower daily cost figure on the headline.

5. Trim MCP servers to what you actually use

The score this fixes: MCP/plugin hygiene, MCP tool-schema weight.

The situation: you connected a database MCP server, a Slack MCP server, a Linear MCP server, and a Google Drive MCP server “just in case.” You use one of them weekly and the rest never. Each one ships its full tool schema in every request.

What to do:

1. List what’s connected: /mcp in Claude Code.
2. For each server, ask: did I use this in the last two weeks?
3. If no, disconnect it. You can always reconnect when you need it.

Aim for ≤3 MCP servers attached. The tool-schema-weight scorer measures exactly this.

How to verify: MCP tool-schema weight score and MCP/plugin hygiene score both rise.

What’s next

Level 300 is for the customisation-curious — automating audits, writing custom skills, tuning hooks, and shipping audit data to a dashboard or Slack channel.