Stop Telling Claude Code What To Do

Detailed Summary

Intro to Meta-Prompting (0:00 - 0:56)

The video introduces meta-prompting as a method to significantly enhance the effectiveness of Claude Code usage. It contrasts this with "childlike" AI coding approaches that often lead to broken implementations or require numerous follow-up prompts for a single task. The presenter emphasizes that for complex tasks, a more structured approach is needed.

The 'Create Prompt' Prompt (0:56 - 13:36)

• The core of meta-prompting is a custom /create-prompt command, which functions as a prompt engineer.
• Instead of directly stating a coding request, the user first invokes /create-prompt followed by their high-level goal (e.g., "I want to create a dashboard").
• Claude, acting as the prompt engineer, analyzes the request, performs a clarity check, and asks clarifying questions if the initial request lacks sufficient context.
• It determines if the task is simple or complex, whether it requires one or multiple prompts, and if multiple, whether they can be parallelized or must be sequential.
• The prompt engineer also assesses the reasoning depth required, potential file references, bash commands, and the need for built-in error checking or validation.
• If clarification is needed, Claude will ask specific questions to resolve ambiguities before proceeding.
• Once clarity is established, Claude confirms its understanding and, upon user approval, generates the optimal prompt, saving it to a designated 'prompts' folder (creating it if necessary) with descriptive naming conventions.
• The generated prompts adhere to Anthropic's recommended XML tag structure, including sections for objective, context, requirements, constraints, output, success criteria, and verification steps.
• Conditional inclusions, such as phrases like "thoroughly analyze" for extended reasoning or "go beyond basics" for creative output, are added as needed.
• The prompt engineer also incorporates intelligence rules, prioritizing clarity, emphasizing context (why the task matters, who it's for), being explicit about output formats, assessing scope, and defaulting to precision over brevity.
• After saving the prompt, Claude presents a decision tree, allowing the user to run the prompt immediately, review/edit it, or save it for later.

The 'Run Prompt' Prompt (13:36 - 14:23)

• The /run-prompt command is designed to execute one or more generated prompts as delegated subtasks.
• This is crucial because it spawns self-contained, fresh context sub-agents, preventing the main chat window's context from being polluted by the prompt generation process.
• It ensures that the actual task execution starts with a clean slate, using only the context provided within the generated prompt.

Using Meta-Prompting in Practice (14:23 - 29:09)

• The presenter demonstrates meta-prompting by building a generative music CLI tool, comparing it with a single-shot approach.
• Without Meta-Prompting: A direct prompt to create the tool immediately leads Claude to generate code without asking clarifying questions. The resulting code is monolithic (a 286-line index.js) and produces buggy, unpleasing music and visuals.
• With Meta-Prompting: The /create-prompt command is used. Claude asks several clarifying questions regarding audio libraries, visual rendering (e.g., using Blessed, rainbow-like aesthetics), note sequencing (randomness with weighting), runtime, and npm package installation.
• After answering these questions, Claude generates a detailed prompt with a clear objective, context, specific audio/visual/runtime requirements, a technical approach (analyzing library selection, timing, harmonic weighting), what to avoid, success criteria, and verification steps.
• Executing this generated prompt via /run-prompt (which spawns a sub-agent) results in a modular codebase (63-line index.js referencing separate audio.js and visuals.js files) and a significantly higher quality, pleasing musical output with trippy visuals.
• A second enhancement task (adding fractal visuals, sine wave feedback, and real-time scale selection) is then performed using both methods.
• Again, the meta-prompted approach leads to Claude asking detailed clarifying questions (e.g., type of fractal, feedback style, visual indicator for scales), which helps refine the user's vision.
• The generated prompt for the enhancement is highly detailed, outlining objective, context (existing application, files to examine), visual/audio enhancement requirements, interactive scale selection, technical approach (researching fractal generation, FM system), what to avoid, success criteria, and verification.
• The meta-prompted enhancement, though taking about 8 minutes, produces a much more robust and functional result compared to the one-shot attempt, which still exhibits issues despite some visual improvements.

Outro (29:09 - 30:14)

• The presenter concludes by highlighting that while meta-prompting might seem long-winded for simple tasks, it is invaluable for complex projects requiring substantial changes, refactoring, or multi-file features.
• The key benefit is that it forces Claude to think about how it should be asked to do something, leading to more thoughtful, structured, and higher-quality outcomes.
• The code for the /create-prompt and /run-prompt commands will be made available on GitHub.

Notable Quotes

"The goal is not to tell Claude what you want it to do. The goal is to tell Claude what you would want it to do and therefore how would it like best to be asked to achieve that instead of explaining what you want and getting it to do it." — Tash

"forcing Claude to think about how it should be asked to do something produces more thoughtful, structured results for complex tasks." — Tash

"This is 100% overkill for simple changes but I've found it be essential for any substantial work I do like big refactoring tasks, database migrations, or multi-file features that require incredibly specific instructions." — Tash

"The prompt in and of itself is a self-contained solution to the problem." — Tash

"I think what's just really great about this approach is not just does it help you understand what you want before you ask for it, but it just makes the entire process of actually prompting so much easier because you don't have to think about the things that the LLM will need as a human being like okay well what does Claude need instead you ask Claude what do you need" — Tash