The Playbook
94 principles for prompting Claude Code and other AI coding assistants — the small, reusable lessons that turn vague asks into prompts a senior engineer would compliment.
Clarity15
Saying exactly what you want, with no room for the AI to guess. Naming files, functions, and behavior up front.
Be Specific
Describe Expected Behavior
Describe the Visual Target
Describe What, Not How
Tell Claude what the function should do, not how to write it.
Invoke the Right Skill
Match your task to an existing skill or slash command and call it by name.
Master the Art
Name the Services
Name the service, the version, and the endpoint.
Name Things Precisely
Use exact names from the code so Claude knows what to change.
Name Your Outputs
Pick the Right Tool
Match the tool to the job: Figma, Stitch, or Penpot.
Pin the MCP Tool
Name the MCP tool by id and list its arguments — don't say 'check the system'.
Specify the Mechanism
Specify the Technology
Summarize the Change for Review
Tell Claude to write the PR for the reviewer: what changed, why, and how to test it.
Write a Trigger-Worthy Description
The description is the trigger. Use the user's words, and say when to skip.
Context20
Giving Claude the surrounding state it needs before asking — what code exists, what depends on what, what already works.
Declare the Tools You Need
Name the tool or MCP server by name, and say what to do if it isn't reachable.
Describe Before You Prescribe
Describe the Data Shape
Claude can't clean data it can't picture.
Explain the Integration Point
Keep Memory Reusable
Save patterns and preferences that survive across sessions; never save snapshots of finished work.
Map to Real Components
Use Code Connect so it reuses your components, not new markup.
Name the Framework
Name Your Files
Point It at the Project Rules
Name the CLAUDE.md or AGENTS.md file explicitly; ask the assistant to cite the rule it's applying.
Quote Current, Specify New
Reference All Affected Files
Respect the Calendar
Give the agent your timezone, hours, and constraints before it schedules.
Scope the Query
Filter the search; don't pull your whole inbox or database.
Scope the Selection
Aim the agent at one frame, not the whole file.
Set the Scene
Show Sample Data
Paste a few real rows: samples beat descriptions.
Specify Prop Contracts
Use the Real Design System
Pull the real tokens; don't let the model invent values.
Use What Exists
Write Durable Instructions
Put project-wide non-negotiables in CLAUDE.md; everything else belongs in prompts, memory, or nowhere.
Constraints20
Naming the lines you don't want crossed: don't change file X, keep the API the same, match the existing style.
Bound the Skill's Scope
Tell the skill what NOT to do — name the look-alike tasks it should refuse or escalate.
Define Component Boundaries
Define Inputs and Outputs
Define the Business Logic
Business jargon is ambiguous: spell out the exact rules.
Describe Edge Cases
Describe the Happy and Sad Paths
Describe Type Contracts
Draft, Don't Send
Produce a reviewable draft before anything goes out.
Explain the Bottleneck
Gate Destructive Actions
Require an explicit OK before delete, archive, decline, or send.
Name Your Stack
Say the stack up front: framework, styling, language.
Pick the Right Model
Match the model to the task's depth: fast models for narrow edits, deeper models for tangled work.
Pin Your CI Environment
Name the runner, versions, caching, and secrets so every CI run is reproducible.
Preserve Existing Behavior
Respect the Constraints
Scope Your Commits
One logical change per commit — name the files and say what changed and why.
Scope Your Request
Specify the Scope
Specify the Trigger
Name the event, the condition, and the action.
Spell Out the States
Name what a static frame can't show: hover, empty, loading, breakpoints.
Errors11
Putting failure modes into the prompt — what should happen when things break, what error types to throw, what to log.
Describe How to Reproduce
Describe the Sequence
Describe What You See
Diagnose Before You Prescribe
Include the Error Message
Quantify the Problem
Read the Error First
Show Expected vs Actual
Specify Error Handling for APIs
APIs fail: tell Claude what to do when they do.
Triage Like an Engineer
Verify the MCP Output
Ask for the raw response shape first; refuse to act on fields the agent can't quote.
Output Format5
Telling Claude the shape of the answer it should produce, not just the substance.
Close the Design Loop
Send the built UI back into the design tool for review.
Describe the Destination
Describe where data lands and what shape it needs.
Make the Output Findable
Write the result where you and the agent can retrieve it later.
Show Examples in the Skill Body
Drop in 2–3 concrete <example> blocks. Patterns travel; rules drift.
Specify Output Format
Multi-Step19
Breaking a complex ask into a sequence Claude can follow without losing the thread.
Architect Before You Ask
Ask for a Plan First
Request a plan before any multi-file change — approve the approach, then let it code.
Break the Workflow into Steps
Number each step so Claude builds pieces, not monoliths.
Combine Context with Action
Coordinate Across Files
Delegate to a Subagent
Spawn a subagent for research-heavy subtasks; brief it on scope, limits, and report shape.
Describe Environments and Stages
Describe Relationships
Describe the Architecture
Describe the Request Flow
Design Is a Conversation
Generate, read it back, correct — design isn't one-shot.
Identify the Pattern
Layer Your Instructions
Make the Skill Stepwise
Split the skill into ordered phases; finish one before starting the next.
Pass Context Between Tools
Hand one tool's output deliberately to the next.
Recognize Plan-Worthy Tasks
Plan when a wrong assumption would be expensive; skip it when the task is one file and obvious.
Sequence the Screens
Describe the whole flow so generated screens connect.
Specify Cache Strategy
Stay Organized Under Pressure
Anti-Patterns4
The shapes of prompts that don't work, and what to write instead.