Writing Prompts That Work Without You

Key Concepts

Prompt Anatomy - The three-part structure every effective prompt follows: context (background knowledge the LLM needs before acting), instruction (the specific task to perform), and output (the format and structure of the result). This anatomy applies at two scales: within a single prompt, and across an entire agent system.

Cognitive Load (in prompts) - When a single prompt is asked to do too many jobs at once (detect, analyze, format), quality degrades. The fix is decomposition: break the task into specialized prompts that each do one job well.

The 3 AM Test - A readiness check for production prompts. Ask: "If this runs tonight with no one watching, do I trust the output?" If your answer includes "probably," "depends," or "I'd need to check," the prompt has gaps.

Quick Reference

Three levels of prompt complexity:

1. Level 1 - A one-sentence ask. Fine for chat, too fragile for automation.
2. Level 2 - Applies the full anatomy: context (audience, brand, standards), instruction (specific focus areas), output (format, length, structure). Good enough for many tasks.
3. Level 3 - Breaks a complex task into a chain of specialized prompts, each with its own anatomy. Use this when a single prompt is doing multiple jobs.

Tips for agent prompts:

• Add constraints - Define the job by what the LLM should not do. "Do not interpret. Do not summarize what stayed the same." This keeps outputs tight for downstream nodes.
• Handle the null case - Tell the LLM what to output when nothing meaningful is found: "If no significant changes, output exactly: No changes detected." Without this, the model will surface noise to avoid silence.
• Grant permission for uncertainty - "If strategic intent cannot be determined, output exactly: Ambiguous." A labeled unknown beats a confident wrong answer.
• Never let the LLM rely on its own knowledge for facts - Always feed real data (scraped pages, raw numbers, actual reviews).

Using Claude Projects to write prompts:

1. Create a project and load in relevant resources: platform docs, API docs, example outputs, and any strong prompts from previous builds
2. Describe what the prompt needs to do, what data comes in, and what the output should look like
3. If building for a workflow tool, export the workflow as JSON and paste it in so Claude can see the full system
4. Review the draft using the anatomy and tips above
5. Test in your agent, paste errors or bad output back into Claude, apply the fix, and iterate