Make it a daily habit

You've installed it and gotten one win. This hour turns that into a habit you reach for without thinking. Still no coding. By the end: saved working rules that help Claude understand your work, a plan-before-it-acts habit, control over long sessions, and one repeatable workflow built end to end.

The real skill: managing Claude's working memory

The master key, from Anthropic's best-practices doc, verbatim: "Claude's context window fills up fast, and performance degrades as it fills." Translate that: Claude has a working memory for the current task. Fill it with the right stuff (relevant files, your rules, a clear brief) and it's sharp. Fill it with junk (three unrelated tasks, a giant log it didn't need) and it gets sloppy. The skill is deciding what goes into the working memory, and clearing it when you switch tasks.

1Save your working rules · 15 min

Every session, Claude starts fresh. You can give each project a short text file of standing instructions. The filename is CLAUDE.md, and Claude reads it at the start of every conversation in that folder. Write your context once; it's there every time. In a project folder, run claude, then:

/init

That scans the folder and drafts a starter file. Make it yours: short and specific to your work.

# Who I am
Revenue leader. I run comp, territory, headcount. Not an engineer.
Explain trade-offs in plain business terms, not code.

# My data
- Exports come from Salesforce; "Amount" is GBP, often formatted as text (£, commas).
- Fiscal year is Feb–Jan. "Q1" = Feb–Apr.
- Reconcile any revenue total back to the £35M plan before you trust it.

# How I want output
- Numbers in £M to one decimal. Tables in markdown.
- Drafts: blunt, operator-to-operator, no corporate filler.

# Always
- Show your working on any calculation so I can check it.

The rule that keeps it useful: for each line ask "would removing this cause Claude to make a mistake?" If not, cut it. A bloated CLAUDE.md makes Claude ignore your rules — the important ones get lost in the noise.

Where it lives: ./CLAUDE.md in the project folder. There's also ~/.claude/CLAUDE.md that applies everywhere — good for "I'm not an engineer, explain in business terms." It's instructions, not data — don't paste your whole price book in; point Claude at the file instead.

2Plan mode — think before you touch · 12 min

The default mistake is letting Claude charge into a multi-step task and solve the wrong problem. The fix is plan mode: read-only. Claude explores and writes a plan, changes nothing, until you approve. Press Shift+Tab to cycle permission modes until it says plan mode. Then:

read every file in this folder. I want to rebuild our Q3 territory carve so quota
matches winnable capacity per region. don't change anything yet — tell me your plan
and what data you're missing.

It comes back with a plan. Read it like you'd read an analyst's approach before they spend two days on it. Edit it, push back, then Shift+Tab back to normal mode and say "go ahead." The rule of thumb: plan when the task spans multiple files, you're unsure of the approach, or it's unfamiliar. Skip the plan when you could describe the change in one sentence.

One underused trick: tell Claude to ask you questions first. You'd never let a good analyst run off on a fuzzy brief — so demand the clarifying questions.

before you plan, ask me up to 5 questions you'd need answered to get this right.

It surfaces the assumptions you didn't know you were making, and the back-and-forth beats any one-shot prompt.

A close cousin, straight from how Cal Rueb works: ask for options, not an answer — "report back with two or three different options; don't start working yet." You pick the approach before a minute's spent down the wrong one.

give me 2-3 ways to restructure this comp plan so it rewards retention more.
lay out the trade-offs. don't change the file yet.

3Session control — why long chats rot · 10 min

• /clear — wipes the working memory. Use it every time you switch to an unrelated task. This one habit fixes half of all "Claude got dumber" complaints.
• /compact <focus> — long task not done? Summarises the history and keeps going. /compact keep the final quota numbers and the reconciliation tells it what to protect.
• Esc — stop it mid-action the second it goes sideways. Context is kept; retype your correction. Don't watch it do the wrong thing.

4Build one real workflow, end to end · 15 min

Make something you'll reuse weekly: raw input → structured output, fully in Claude Code. Drop your messy notes from this week's deal reviews in a folder, cd in, run claude:

read all the notes in this folder. produce one markdown doc with:
1. a 3-bullet summary per rep,
2. a single table of every commitment made, who owns it, and the due date,
3. the top 3 risks across the whole set, ranked.
save it as week-readout.md. show me the table first so I can sanity-check before you write the file.

Notice what you just used: a folder of real inputs, a specific multi-part brief, a built-in check ("show me the table first"), and a saved output. That's the whole game. The reuse trick: the wording that worked is your asset. Keep a note of your best briefs — next week, same brief, new folder, 90 seconds instead of an afternoon. Anthropic's engineers call this compounding: each good brief folds into the next, so the work gets sharper over time instead of messier. That habit, more than any command, separates people three months in.

The power move: build one to throw away

A trick the founders of Claude Code use that almost nobody teaches — because it only makes sense once work is this cheap. When you're not sure how to brief a task, don't try to get the brief right first. Let Claude build a rough version off a loose ask, watch what it gets wrong, then /clear and write the real brief using what you just learned. You're not after the first output — you're using it to discover what you actually wanted.

A muddled "build me a deal-health score from this pipeline export" tells you, in its mistakes, exactly which factors and thresholds you care about. Then you clear, and the second brief is sharp because the first one taught you. Throwing the first one away was always the plan.