Guide

Run skills in loops

A loop is an agent repeating one job until a condition is met: act, check the result, go again. The agent does the repeating. Your job moves to choosing the goal, deciding what counts as proof, and improving the system when it gets something wrong.

What makes a loop safe to leave running

A check the agent can't argue with. A test run or a scripted drive of the app, rather than its own opinion that the work looks done.
A record. Something you can read later to see what it decided and why.
A stop condition. A clear definition of done, and a limit for when it isn't getting there.

Four levels of hands-off

Move down a level only once you trust the checks at your current one.

Level	Your job	Skills
Prompt and watch	Read every step as it happens.	/pocock:grill-me /pstack:how
Approve at checkpoints	Agree the plan, then review each finished slice.	/pocock:to-tickets /pocock:implement /addy:incremental-implementation
Run several at once	Set the task, then judge the combined result.	/pstack:swarm /pstack:arena
Step away	Review the evidence and the decision log afterwards.	/pstack:figure-it-out /pstack:show-me-your-work

Build the check first

A loop is only as good as its check, so this is the first thing to set up.

/pstack:create-verification-skill generates a project skill that drives your app the way a user does. /pstack:maintain-verification-skill keeps it honest as the code changes.
/pocock:tdd gives each pass through the loop a test to satisfy.
/addy:browser-testing-with-devtools checks real browser behavior for UI work.
/addy:doubt-driven-development sends non-trivial decisions to a fresh-context reviewer.

Keep context across sessions

Long work outlasts a single conversation, so plan for the gaps.

/pocock:wayfinder maps work too big for one session as decision tickets on your tracker.
/pocock:handoff compacts a conversation into a document another agent can pick up.
/pstack:recall rebuilds your recent context when you come back.
/addy:context-engineering helps when output quality drops or you switch tasks.

Improve the system, not just the output

When a loop gets something wrong, fix the cause so the next run doesn't repeat it.

/pstack:correct turns a repeated mistake into a check that fails on it.
/pstack:reflect reviews a session and routes what it learned into skill edits.
/pocock:retro runs a retrospective on a coding session.
/personal:skill-writer captures a workflow you keep re-explaining as a new skill.

Running one on a schedule

In Claude Code, /loop repeats a prompt or slash command on an interval, and /schedule runs agents on a cron schedule. Any skill here can be the command. For example:

/loop 30m /pstack:maintain-verification-skill

Treat that interval as an example and pick one that matches how fast the thing changes.