+++
date = '2026-04-06T13:15:06Z'
draft = true
title = 'Agentic Programming Rules of Thumb'
tags = ['ai', 'llm', 'agentic', 'programming', 'workflow', 'AI-reviewed']
+++

## How to work with AI

Like everybody else, I am trying to figure out how to use AI effectively in my software job and my computer hobbies.

## Rules of thumb

- **Keep emotions out of it.** Every irrelevant token in a prompt may reduce output quality. Don't insult, greet, or praise the AI.
- **Use repository-level instructions.** I currently use `AGENTS.md` for this. I've seen other naming suggestions. The industry hasn't settled on a standard like `~/.agentsrc` — that would be nice.
- **Think about the problem. A lot.** The better you partition a problem into well-defined steps, the better the output.
    - If you don't know the steps, the first step can be asking the AI to help break down your vague problem.
    - Err on the side of small steps.
- **Restart often.** Every bit of irrelevant context reduces output quality. Once you finish a task like "write a pytest `tests/test_some_feature.py::test_some_bug` asserting X and Y and verifying Z" — kill the session, start a new one for the next step.
- **Word your prompt as precisely as possible.**
    - The AI may understand "add support in the function for dictionaries in addition to lists." It will understand much better: "In function `foo(l: list)`, refactor `l` to accept either a list or a dictionary. If the user supplies a list, keep the current behavior. If the user supplies a dictionary: ignore the keys, extract the values into a list (or iterable), and process that collection the same way `foo` used to process the list."
    - Precise prompts reduce the chance of a fruitless back-and-forth.
    - If a session got polluted with arguing — the AI ignoring your new instructions — you likely now have a better understanding of the problem. Use that failure to write a sharper prompt and start over.
- **Use files for output.** Instruct the AI to write to `plan.md` or `tests/test_some_bug.py`. Easier to track and review.

## Patterns

### Work on the problem, then work on the solution

**Step 1: define the problem.**

Loop:

1. Write everything you know about the problem in a document (`problem.md`).
    - This is meant to be rapid. Don't focus on grammar or style. Focus on writing down everything relevant.
2. Ask the AI to review the document and flag anything unclear, imprecise, or redundant.
3. Ask the AI to list additional details it needs about your problem. Instruct it to ignore document structure and grammar unless they cause ambiguity.

When the AI starts being pedantic or irrelevant, address it in your instructions. For example: "There are no absolute benchmarks for query performance on this system. I will evaluate performance manually. We still state that evaluating performance is an objective."

**Step 2: ask the AI to edit your problem document.**

The objective: a plan you understand, written in a logical flow matching what you will try to accomplish. Prompt something like:

> Format document `plan.md` into an operational, chunked plan for my problem. Use simple steps. Include all details relevant to each step and nothing else. Keep language dry.

Best practices for this step are TBD and likely depend on the problem and the AI engine used.

**Step 3: solve each step independently.**

For each step, review the plan carefully. Modify the current step based on findings from previous steps if needed. Then instruct the AI to solve that step in a new session, with only the context relevant to it.

## Links

Practical, opinionated guides on AI-assisted software development workflows. Not prompt-engineering tutorials — these are about how to organize your work. I found the first link myself; the rest were surfaced by a GitHub Copilot agent session (Claude Opus 4.6).

- [Simon Willison — Agentic Engineering Patterns](https://simonwillison.net/guides/agentic-engineering-patterns/) — comprehensive guide covering principles, testing (red/green TDD with agents), subagents, git workflows, and anti-patterns. The most structured resource on this list.
- [Anthropic — Claude Code Best Practices](https://code.claude.com/docs/en/best-practices) — official guide from Anthropic. Key ideas: give the AI verification criteria (tests, screenshots), explore first then plan then code, manage context aggressively (`/clear` between tasks), use subagents for investigation to keep main context clean. Also covers CLAUDE.md authoring, hooks, skills, and scaling to parallel sessions.
- [Harper Reed — My LLM codegen workflow atm](https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/) — end-to-end workflow for both greenfield and brownfield. Greenfield: brainstorm spec → plan a plan with a reasoning model → execute step-by-step. Brownfield: use repomix to dump context, generate missing tests / code review / issues via LLM, then fix them one at a time. Practical and opinionated.
- [David Crawshaw — How I program with LLMs](https://crawshaw.io/blog/programming-with-llms) — thoughtful practitioner perspective. Key insight: treat LLM tasks like exam questions — give a specific objective and all background material, ask for work that is easy to verify. Notes that extra code structure (smaller packages, more tests) is now much cheaper, shifting old tradeoffs. Builder of sketch.dev.
- [Thorsten Ball — How I Use AI](https://registerspill.thorstenball.com/p/how-i-use-ai) — diary-style log of two days of actual AI usage (Zed inline assist, ChatGPT, Copilot). Valuable for seeing the mundane reality: translation, type conversions, TDD test scaffolding, debugging Unix process groups. Shows when AI helps and when it doesn't.
- [Steve Yegge — The death of the junior developer](https://sourcegraph.com/blog/the-death-of-the-junior-developer) — coins "Chat-Oriented Programming" (CHOP). Argues chat-first coding is the new default, with hand-writing as fallback. Key observation: chat is safer for senior devs than junior ones, because you need to detect when the AI gives specious but technically-correct advice.