Build a rhythm you can repeat.
Catch up, choose the right workflow, keep context useful, and learn from mistakes that actually recur.
The aim isn’t to spend the day managing agents. It’s to spend less time re-explaining the same context and more time making decisions that matter. A few repeatable habits help more than a giant prompt you paste into every session.
Get current before starting.
On a shared repository, begin with catch-up when you’ve been away or the branch has moved. It groups changes by feature and points out what affects your work.
It doesn’t fetch or pull. Remote mode uses the tracking refs already on your machine, so check their freshness before treating the briefing as current server state.
› /catch-upChoose the smallest workflow that fits.
Use implement-feature for a contained change. Use plan-change when you need to understand the impact before building. Use hungry-jack when the work spans systems and needs coordination. Its default balanced effort is a good starting point.
For a working but heavy diff, simplify-code can help. For evidence and review, use verify-work and review-changes. You don’t need every skill on every task.
› /hungry-jack --effort balanced add progress and safe retries to our import workflow. Reuse the existing upload and queue systems.Install the hungry-jack companion bundle from its skill page before using this prompt.
Get the companion bundleCarry decisions forward, not the whole conversation.
When a session gets unwieldy, a fresh one can be easier to reason about. Carry over the goal, constraints, affected files, decisions made, checks already run, and the exact next step. Include a base commit or branch where useful.
Keep durable project conventions in the repository’s instruction files. Keep temporary task details with the task. Don’t add every one-off preference to AGENTS.md.
› Prepare a handoff for a fresh agent session: goal, acceptance criteria, decisions, affected files, checks and results, unresolved issues, and the next concrete step. Do not change files.Fix recurring failure patterns deliberately.
If the same kind of bug keeps escaping, use reflect with the actual evidence. It checks whether the problem is systematic and proposes a prevention, preferring tests and tooling over another vague instruction.
Read the exact proposal before approving a persistent change. A clean, successful task doesn’t need a reflection ritual.
› /reflect the confirmed regression and failing test in this session. Check whether the pattern appears elsewhere and propose the smallest prevention. Show the exact change before applying it.End with a change you can explain.
Before moving on, be able to say what changed, why it changed, what proved it worked, and what remains uncertain. If that takes a long explanation, the task might be too broad or the diff might need another look.
Save the useful context in your normal issue or pull request workflow. A good handoff lets the next person review the work without reading your entire agent conversation.