CLAUDE CODE

A starter workflow for this week

Define done, let Claude interview you, approve a plan, build one thin slice, and keep the change only if a fresh review agrees.

Figure 1: The week starts with a definition of done, a plan you edit, and one thin slice that works.

1. One real task, not a tour of the product

Pick a job you already understand well enough to reject a wrong answer. A broken test. A button. A script you can run. Do not pick a rewrite of the platform. The point of the week is to practice the loop on something you can check.

Write a few sentences, including what done looks like and how you will check it. If you cannot check it, you are not ready to delegate it.

2. The sequence

Have Claude interview you, one question at a time. Then have it write a plan file with a must-have tier, a nice-to-have tier, and a stretch. Edit the file. Approve it before source edits. Check that CLAUDE.md names the build, the test, and the gotchas. Build one thin path and commit as you go. Expand only after that path works. If you expand in parallel, use separate branches or worktrees and name the owner of each.

Test in a fresh session. Ask for ranked findings with evidence. If you changed a prompt, a skill, or a model-driven feature, run the small eval set. Polish with a checklist after the core works. Hand off with sample data and a fallback. Spend ten minutes writing the lesson back into the handbook or the command you will reuse.

Layman rule: Hand off only what you can explain. A demo you cannot recover is not a demo.

3. The traps that waste the week

Vague prompts produce generic output. Too many skills hide the one rule that mattered. A skill that never triggers needs a better description, not a longer body. A skill you have not read will misbehave, and you will blame the model. Parallel agents you cannot review will disagree with each other. A long messy session should be reset, with the plan file as the source of truth.

An unvalidated model judge will praise the wrong thing. A production path with no trace cannot be fixed. A third-party server you did not read has real access. A tool chosen from a ranking, and never tried on your task, is a rumor. Watching a tutorial is not the same as using the technique the same day.

4. Leave the learning where it belongs

If this week’s goal is your own judgment, ask Claude to teach and to quiz you. Do the hard part yourself. Over-delegating the learning is how a team becomes dependent on a transcript nobody can reconstruct.

Commands and shortcuts in this series match widely used Claude Code practice, and they do change. If a slash command here does not exist in your session, the current documentation is the list. Re-test the tool on your own tasks every few months.

5. The ten pages

The enterprise series starts at a different altitude: a model is not a system.