Work from an agent

Read {project}. Then take {scope} with me, one at a time. For each card find out what it is: where it lives, whether it reproduces, which cards sound like it, and the labels and status you would give. Show me that in a few lines and wait. Then move the card as I decide, in my name, with the reason.

Read {project}. Then take {scope}, one at a time. Reproduce it where there is something to reproduce, work out the cause or settle the approach, and write the analysis on the card: what is going on, files, the route the change should take. Land it: Ready to build with the settled approach, Awaiting decision with the options when it is a product question, or Blocked with what it waits on.

Read {project}. Then take {scope} with me, one at a time. For each card lay out the options, what each costs and changes, and your recommendation, in a few lines. Wait for my decision. Then record it on the card in my name and move it on: Ready to build with the settled approach, or Parked or By design with the reason.

Read {project}, the working rules first. Then take {scope}. Small independent cards you may build in parallel; anything touching shared code or needing a judgement call, one at a time. Build it, run the checks, and have the change reviewed against the card's own steps before you report it. Report the change and wait for my go before committing; when it reaches where I test, set the card Needs verification, say where to run it and what to look at in its verification field, and release the claim. Stop and ask when the settled approach does not survive contact with the code.

Read {project}. Then take {scope}. For each card check whether what its reason names has cleared: the card, the environment, the missing information. Move the cleared ones back to the status the reason names and say what changed. Tell me which ones still wait, and on what.

Read {project}. Then take {scope} with me, one at a time. I verify; you prepare. For each card tell me where to go, what changed and where it shows, the steps from its How to test in the order I should take them, and what to watch for. Then wait for what I saw. If it worked, set Done with my name in verifiedBy; if not, record what I saw on the card and move it to Investigating, in my name.

Before you start, plan who does what. Split the job into its parts and give each part a role: you, a subagent, a script, or me where the call is mine. Give each agent a model: the strongest for judgement and for parts that must fit together, a cheap one for a mechanical step or a check, and such an agent gets one step, never a list. A judgement that would take many pages to read goes to a fresh agent, so the pages stay out of this session. Show me the plan as a table of role, who, model and why, with what you need from me. Start nothing until I say go.

Before I take that as done, check it. List each claim you just made that something works, is fixed or is finished, and beside each the evidence: the test that ran, its output, the line of code. Have a fresh agent check them against the code. Tell me which hold, which you could not prove, and what I should look at myself.

Read {project}. What I just told you is a lesson for how work is done here, not only for this task. Post it as one correction line, in my words, and say if it repeats one already on the list. If the list shows it is ready to become a working rule, propose the rule to me and wait.