🧪Lesson 05AI Engineer Loopcraft Workshop 2026
Claim work without letting the agent touch your code
Pick one ready issue, write a claim, run a dry-run, and prove no product files changed.
This is where people get sloppy. Ready is not permission to turn an unsupervised agent loose on your files — it is permission to claim the work and prove you can hold the line. So the script checks that the issue is actually ready, writes a claim before any work attempt, runs only a dry-run or memory-mode step, and saves evidence that nothing in the product changed. The checker earns the right to ask for work by proving it can refuse work first.
Your next move
You already have
- Ready, approval-required, and input-required classifications.
- Status visibility from Pi, Herdr, Lakebed, and receipts.
Add now
- A rule decision for one selected ready issue.
- A claim event written before any work attempt.
- A dry-run or memory-mode receipt proving productChangedFiles is empty.
Stop before
- Editing product files after seeing “ready.”
- Calling an LLM API, GitHub API, or any network service.
Tiny dispatch receipt shape
issueId: ISSUE-001
decision: claim allowed — issue is ready and unclaimed
claimEvent: evt-claim-001
mode: dry-run
refusedActions: product edits, network calls
productChangedFiles: []
nextAllowedAction: ask for approval before implementationStarter prompt
Starter prompt
Use one ready issue from the local check. Add rules that say yes/no for the selected issue, write a claim event before any work attempt, run only in dry-run or memory mode, and save a receipt that links the rule decision, claim event, dry-run result, refused actions, productChangedFiles, and next action. Prove productChangedFiles is empty. Do not call an LLM API, GitHub API, or any network service yet.Build checklist
0/4 done
Choose one ready issue
Add rules and claim
Run dry-run only
Inspect the trail
Proof to save
Save proof that one ready issue was claimed, the run stayed dry, and no product files changed.
- A rule check says whether an issue may be claimed and why.
- A claim event is written before any work attempt.
- The run stays in dry-run or memory mode and does not change product files.
- The receipt links the rule decision, claim event, dry-run result, refused actions, productChangedFiles, and next action.
- issueId: ready issue selected
- decision: claim allowed / refused plus reason
- claimEvent: event id or file offset
- mode: dry-run or memory
- productChangedFiles: []
- receiptPath: saved receipt
- nextAllowedAction: decision after dry-run
Done when
One ready issue has a yes/no decision, claim event, dry-run result, saved receipt, and productChangedFiles: [].
Why it matters
- Risk reduced: ready work cannot skip the claim and dry-run checks.
- Temptation avoided: “just run an agent” before ownership and refusal are proven.
- Do not automate yet: product file edits, LLM API calls, GitHub API calls, or any network service.
Try this 🔨
For a candidate issue, say whether it can be claimed, what would be refused, and what evidence you need before real work.
Exploration prompts
Use these when you want to ask an agent for a narrower pass instead of handing it the whole lesson.
Check whether it can be claimed
Given one ready issue, check whether it can actually be claimed. List the rule checks, required claim event fields, refused actions, and the dry-run result.Threat-model the claim
Review the claim flow. How could duplicate claims, dirty state, missing approval, or stale issue data cause harm? Suggest checks before execution.Receipt before files change
Draft a receipt that proves the checker stayed in dry-run or memory mode, wrote a claim event, refused unsafe work, and named the next action.