How I Prompt

Who Should Read

A backend developer who gives coding agents short instructions, gets off-target results, and wants better prompting habits.

Core Mechanics

• One question drives everything
  • When writing a prompt, ask how the model is supposed to know what you mean.
  • Information comes only from training data or the context window (everything sent to the model).
  • The context window holds the system prompt, tool definitions, conversation, tool results and your prompt.
• A model won't ask what you mean
  • With just "fix the upload bug", the model won't ask which bug.
  • It says "you're absolutely right" and fixes whatever it assumes the bug is.
  • Picture a senior engineer kidnapped and handed a single note.
• Big features got all the information
  • The daemon mode prompt pointed to three folders, three files and docs.
  • It also stated constraints like no checkouts, no work trees and managing process lifetime.
  • This information lives in neither the training data nor the codebase, so it went into the prompt.
• Changes point at sources
  • Two weeks later the change prompt told the agent to read the relevant files and docs.

Community Reactions

• Screenshots help unevenly depending on the work
  • One commenter doing autonomous-vehicle road geometry says the AI can't see disconnected or self-intersecting polylines in screenshots.
  • It had to dump coordinates to believe the bug existed and often claimed fixes that weren't.
  • They say GPT-6 Astra is much better but still occasionally struggles.
• The asset search relies on codebase structure
  • A commenter argues the same search could burn tokens in a codebase with hashed filenames and thousands of images.
  • They say the author leans on familiarity with his org and that token costs can vary widely.
• Concrete counterexamples make rules stick
  • One commenter found that attaching a counterexample makes the agent follow a rule literally, while a bare principle is followed loosely.
  • "Be honest" failed, while "compare on canonical form because a single newline byte changed the hash before" worked.
• Newer models changed prompting
  • One commenter said older models required step-by-step manual orchestration.

How to Apply

• Put three kinds of sources in every prompt
  • At the top of today's prompt, list at least three relevant folders, files or docs.
  • Below, list constraints that aren't in the code, one per line.
  • Check each item against "where would the model get this?" before sending.
• Split find and ask into two prompts
  • First, ask the agent to find the relevant files.
  • Once they're in context, send the change request.
  • Use this only when you don't know where the target lives.
• Fix the broken side against a good reference
  • If one implementation works, open with "this is the standard, investigate its mechanism".
  • Then ask to make the broken one behave the same.
  • Point at files for what good looks like instead of just saying it's broken.
• Add a copy-debug-context button
  • Add a debug view that copies IDs, log links and data dumps as markdown.

Try It

# Paste into a coding agent like Claude Code or Amp (the talk's method turned into a prompt)

I want to {{one-line task description}}. Before changing code, check the information below first.

[Sources]
- Relevant folders: {{folder1}}, {{folder2}}
- Relevant files: {{file1}}, {{file2}}
- Relevant docs: {{doc path}}
- Reference implementation that works well: {{e.g., CLI queuing logic file}}

[Constraints]
- {{constraint not in code 1}}
- {{constraint not in code 2}}

[Desired behavior]
{{behavior description from my head}}

Instructions:
1. Before changing code, read the sources above and summarize the current behavior you understand and any missing information.
2. If something I described can't be known from training data or the codebase, ask instead of guessing.
3. Then implement.

Next step: if the result is wrong, find which information was missing and add it to the next prompt.

Terminology

Related Papers

• AI is a tool and using it well takes skill

  A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome

• Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting

  Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.

• GPT‑6 and Intelligent UI for everyone

  GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.

• Base Models Can Reason By Taking a Cue From Training Data

  Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.

• Labels Override Definitions in Jev-Style Typed Decision Models

  Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.

• How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs

  Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output

Related Resources

• https://thorstenball.com/talks/how-i-prompt/
• https://news.ycombinator.com/item?id=49334444