Prompting That Works: Tell Claude What It Can't See

Most bad sessions have the same few causes

When a session goes sideways, it's almost always one of these:

1. Claude is missing information: something that isn't in the repo, or is slow to rediscover.
2. Claude has no check to run, so it can't tell success from failure.
3. A rule needs enforcement, not a request. (That's a job for hooks, Unit 7.)

This lesson tackles the first two, which you control in every prompt you write.

The "what can't it see?" test

Before you hit Enter, ask: what do I know that Claude can't find by reading the repo?

Things Claude typically can't see: the bug report, the customer's exact words, decisions made in meetings, which approach the team already rejected, production-only behavior and deadlines.

Give it a check

The single highest-leverage habit: tell Claude how to verify success.

• A test command: "run pnpm test -- billing after each change"
• A build or typecheck: "cargo build must succeed with no warnings"
• A reproduction: "this curl should return 201"
• A visual target: paste a screenshot of the design and ask Claude to compare

implement the empty-state design in the attached screenshot. then run the dev server,
screenshot the page and compare it to mine. list any differences and fix them.

Feed it context directly

• Files: @src/billing/invoice.ts pulls a file in.
• Images: paste with Ctrl+V or drag in screenshots of bugs, designs and diagrams.
• Pipes: cat error.log | claude -p "explain the root cause"
• URLs: "read https://… and apply its migration guide"
• CLIs: "use gh issue view 312 to read the bug report"

Be specific about scope

Claude can be ambitious, which is great when you want it and costly when you don't. Say which you want:

minimal fix only: don't refactor surrounding code, don't rename anything.

be ambitious: this is a greenfield prototype, propose a clean architecture.

Effort, not magic words

Rather than sprinkling "think hard" into prompts, control reasoning directly with effort levels. /effort sets how much reasoning Claude puts in: medium is the default, high for subtle work, low for fast iteration. Including ultrathink in a prompt raises effort for that one turn.

Check your understanding