A Practical LLM Coding Workflow

A coding agent earns its keep when you give it a job with a clear finish line. These are the prompts I use, followed by the discipline that keeps the output reviewable.

Learn the repository first

Before changing anything, make the agent describe what it sees:

what does this project do?
which technologies does this project use?
where is the main entry point?
explain the folder structure
what can Claude Code do?
how do I create custom skills in Claude Code?
can Claude Code work with Docker?

The answers are cheap, quick and easy to check.

Make one change at a time

State the change and the file, and nothing else:

add a hello world function to the @main.py file

Small requests keep the diff readable and the review simple.

Let the agent drive Git

what files have I changed?
commit my changes with a descriptive message
create a new branch called feature/quickstart
show me the last 5 commits
help me resolve merge conflicts

Read every commit message before you accept it. The agent will describe what it thinks it did, which is not always what it did.

Bugs, features and cleanup

Examples:

add input validation to the user registration form
there is a bug where users can submit empty forms, fix it
refactor the authentication module to use async/await instead of callbacks
write unit tests for the calculator functions
update the README with installation instructions
review my changes and suggest improvements

The discipline behind it

Addy Osmani’s 2026 workflow supplies the rules that make the prompts pay off:

Plan first. Write a specification and a plan with the model before it writes code.
Work in small chunks. Give it one step from the plan at a time.
Provide rich context. Feed it the code, docs and constraints it needs up front.
Use the right tool. Switch models when one struggles instead of arguing with it.
Stay in the loop. Review, test and debug everything, and keep accountability for the result.
Automate the checks. Linters, tests and CI catch what a review misses.
Commit often. Version control is the rollback plan.
Customise behaviour. Rules files and style guides cut the generic output.
Spend the agent on boilerplate. Keep the design decisions for yourself.

Compare harnesses

Run the same prompts in a different agent and compare the diffs. The prompts stay constant, so the differences you see belong to the harness and the model rather than to your instructions.