Best AI Prompts for Claude Code (2026)

Prompts for Claude Code#

Read this repository and draft a CLAUDE.md of under 200 lines. Include only what you could not work out from the code: the exact build, lint and single-test commands, code style that differs from the defaults, branch and PR conventions, and required environment variables. Then ask me about anything you could not determine, and list each line you were unsure earns its place. My stack is [DESCRIBE YOUR STACK AND PACKAGE MANAGER].

Start in plan mode. Read [PASTE THE DIRECTORIES OR FILES TO CHANGE] and write a plan to refactor [DESCRIBE THE CHANGE], listing every file you would touch and what is out of scope. Wait for my approval. Then implement it in small steps, run [PASTE YOUR TEST AND TYPECHECK COMMANDS] after each step, fix any failure at its root cause, and show me the test output when you finish.

Look at the failing build log I pasted: [PASTE THE ERROR OUTPUT]. Find the root cause rather than suppressing the error, write a failing test that reproduces it, fix it, and confirm the build passes. Then create a branch named [PASTE YOUR BRANCH NAMING RULE], commit with a message in our convention [DESCRIBE YOUR COMMIT FORMAT], and open a pull request with gh that links the issue.

Write a shell script that loops over the files in files.txt and runs claude -p once per file to [DESCRIBE THE PER-FILE TASK], ending each answer with OK or FAIL. Pre-approve only the tools the task needs with --allowedTools, use --permission-mode dontAsk, run with --bare, and log each result as JSON. Run it on the first three files and show me the output before you run the rest.

Audit how Claude Code is configured on this [DESCRIBE THE MACHINE, E.G. SHARED CLOUD VM]. Propose a .claude/settings.json with deny rules for [LIST SECRETS FILES AND DANGEROUS COMMANDS, E.G. .env, curl, rm], an allowlist limited to the commands we actually use, and a PreToolUse hook for anything that must never run. Explain each rule in one line, and note what a Bash deny rule would not catch.

What to add to any Claude Code prompt#

• A check Claude can run. Anthropic's best-practices page says a test suite, build exit code or linter lets Claude iterate until it passes, and asks for the evidence (test output) rather than a claim of success.
• Files, symptoms and patterns by name. Reference files with @, describe the symptom and likely location, and point to an existing file to copy the pattern from.
• Scope and exclusions. Say which files are in play and what is out of scope. Anthropic's own spec advice is to name the files and interfaces involved and state what is out of scope.
• Why a rule exists. Anthropic's prompting guide says explaining the reason behind an instruction helps Claude deliver more targeted results.
• Plan mode for multi-file work. Anthropic suggests exploring and planning first when you are unsure of the approach or the change touches several files, and skipping the plan if you could describe the diff in one sentence.
• A fresh session after repeated corrections. After two failed corrections on the same issue, Anthropic recommends /clear and a better first prompt.

A worked example#

This fills in the second prompt for a hypothetical Next.js repository.

Start in plan mode. Read @src/lib/billing/ and @src/app/api/invoices/ and write a plan to refactor invoice totals so all rounding happens in one function, src/lib/billing/money.ts, instead of in three route handlers. List every file you would touch. Out of scope: the Stripe webhook code and any database schema change. Wait for my approval. Then implement it in small steps, run `npm run typecheck && npx vitest run src/lib/billing` after each step, fix any failure at its root cause, and show me the test output when you finish.

A good result for this task has these parts:

• A plan that names the files and states the out-of-scope items back to you before any edit.
• A characterization test or existing test that passes before the refactor and still passes after it.
• Test and typecheck output pasted into the reply, not a statement that it passed.
• A diff that touches only the listed files; anything outside the plan is flagged and asked about.
• No test edited to make it pass, and no error suppressed.

For unattended runs, review the allowlist before you start. --allowedTools pre-approves tools, and Anthropic notes a deny rule on one Bash form does not match the same program invoked another way, so it is not a security boundary on its own. Pair it with the sandbox for anything that must hold. Without --bare, a claude -p run loads the hooks and MCP servers in the project's own settings, even in a folder you have not trusted, so read those first. Keep API keys out of prompts and shell history.

The full prompt library is on our prompt library page.

Limits of this guide#

This page covers how to ask and configure, as Anthropic documents it on 2026-10-22. It does not show that any wording makes Claude Code more accurate, and we did not run these prompts. Claude Code ships changes often, and flags such as --bare, --permission-prompts and the auto permission mode depend on your installed version, so check claude --help first. We could not read OpenAI's prompt-engineering guide through the Internet Archive from our tooling, so the cross-vendor points come from OpenAI's live page and Google's guide.