Implement a change end to end with your AI agent

Give the agent a request and it works through it in fixed steps, then tells you what it checked:

/work "add a CSV export to the orders list"
/implement-ticket PROJ-123

What it does

/work takes a free-form request through refine → plan → implement → test → verify → report. /implement-ticket runs the same loop for a ticket and refines the ticket first. Both:

ask one question when the request is ambiguous, instead of guessing;
read the code they are about to change — callers, tests, data — before editing;
write the test first where a test makes sense;
run something that exercises the change before saying done, and say so when nothing could be run;
stop at a report. They never commit, push or open a pull request on their own.

When to use it

Any change you would otherwise describe to the agent in a chat message: a feature, a bug fix with a known symptom, a refactor with a clear goal. For bigger work that spans days or several pull requests, plan it as a roadmap first.

Example

/work "orders list: add a CSV export button that downloads the filtered rows"

A good report names the files changed, the tests added, the command that was run to check the export and its result, and anything left open — for example that the export was not tried with ten thousand rows.

What it needs

Any supported AI coding tool. No API key beyond the one your tool already uses.
For /implement-ticket, a ticket the agent can read — pasted into the chat, or through your tool’s Jira or Linear connection.

Limits

It changes how the agent works, not how capable the model is. On a strong model the measured gain was a null; on a weaker one it was significant. Benchmark & cost.
It spends more tokens than a plain prompt, because checking is work.
Outside Claude Code the steps are instructions the agent is asked to follow, not checks that can stop it. What each tool can block.