My Sample Workflow

My Sample Workflow

Three modes, depending on the task.

1. Light Tasks

Single prompt, no back-and-forth.

  • Fix a Python graph so the y-axis range starts at 0
  • Remove all LaTeX figure titles from a .tex file
  • Delete all empty lines in a file

Characteristics: lightweight · well-defined · specific

2. Structured but Multi-Step Tasks

For tasks that need planning before any code is written — e.g. building a data cleaning pipeline.

Step 1 — Dump all requirements to the agent using the Plan skill. Or draft the plan in a chat interface first, then hand it off to the agent.

The goal is a plan.md the agent can execute locally:

You are helping me build a data cleaning pipeline in Python for a panel dataset.
Here is what I need: [list your requirements].
Do not write any code yet. Produce a structured plan as a markdown file called
plan.md. Break the work into numbered steps, flag decisions I need to make, and
note any assumptions.

Step 2 — Review the plan, resolve any open decisions, then tell the agent to execute.

3. Unstructured Idea Bouncing

For half-formed hypotheses you want to think through before committing to a structure.

Example: “My hypothesis is X. I want to test it with Y. What do you think?”

  • Set the persona first. One line: who should the AI be? (e.g., “You are an empirical economist specializing in labor markets”). For a local agent, add it to CLAUDE.md. Always include “Do not execute anything” to keep it in discussion mode.
  • Iterate. Poke at assumptions, refine the approach. Once things are concrete, you have a plan — go to Section 2.

A Note on Auditing

Be specific. Be suspicious. Push back.

I code badly, and sometimes I cannot audit all the code AI generates. What I do instead:

  • Run tests to verify the output behaves as intended
  • Ask for the same thing in different wording and compare results
  • Paste the code into a separate session — or a different AI — and ask whether it does what you want

The best way to critique AI output is with another AI.