Launch deal: $99 lifetime for your first always-on AI system.
Build Lean SaaS cube logoBuild Lean SaaS
Back to Always-On Agents
Course lesson 11checklistintermediate

The First Closed Loop

Run one tiny real task through the full AI brain loop: input, scratchpad, GitHub Issue, implementation, verification, and status report.

Austin Witherow
5 min read
Use the attached checklist

The brain only matters if it moves work to completion.

A beautiful vault, clean scratchpad, and perfect GitHub Issue do not matter until they help you ship something verified. This lesson is the smallest full loop: one real input becomes one tiny finished result with receipts.

Outcome for this lesson

By the end, you should have completed one tiny task through this path:

Pick a task small enough to finish in one sitting.

Choose a tiny task

Good first closed-loop tasks:

  • fix one typo on a public page;
  • add one missing checklist item;
  • update one stale link;
  • add one short template section;
  • document one repeated workflow;
  • create one small content improvement;
  • add one focused test for existing behavior.

Bad first closed-loop tasks:

  • redesign the whole site;
  • rebuild your vault;
  • automate every inbox;
  • create a giant backlog;
  • migrate all project management;
  • connect production credentials before the manual loop works.

The goal is not to impress yourself. The goal is to prove the loop.

Step 1: capture the input

Start with the original input exactly enough that the source is recoverable.

Then normalize it:

Step 2: update the scratchpad

Create or update the scratchpad with current state:

This is what lets the system recover if the chat dies.

Step 3: create the GitHub Issue

Promote only the smallest executable piece.

Your issue should include:

  • goal;
  • context;
  • scope;
  • out of scope;
  • acceptance criteria;
  • verification plan;
  • reporting destination.

If the issue feels too large, split it before execution starts.

Step 4: execute one narrow slice

Work in the smallest safe implementation path:

  • one branch;
  • one focused set of files;
  • one PR when code/content is involved;
  • no opportunistic redesign;
  • no unrelated cleanup;
  • no hidden side effects.

For non-code artifacts, the same principle applies: one output, one destination, one verification path.

Step 5: verify before saying done

Use receipts that fit the task.

Do not claim done from intent. Claim done from proof.

Step 6: close every surface

A closed loop updates all places that need final state:

  1. GitHub Issue: comment or close with receipts.
  2. PR: merge or clearly mark blocked.
  3. Scratchpad: update Done, Pending, Blocked, Next.
  4. Original thread: post a concise status with links.
  5. Workflow/skill: update if the task revealed a reusable correction.

If the original thread still says “next action: create issue,” but the issue shipped yesterday, your brain is stale.

Copy this closed-loop checklist

Verification checkpoint

You are done when you can point to:

  • the original input;
  • the scratchpad;
  • the GitHub Issue;
  • the PR, artifact, or changed file;
  • the verification receipt;
  • the final status update.

If any link is missing, the loop is not closed yet.

Next lesson

Once you have closed one loop, the next improvement is self-healing: when the assistant gets corrected, the system changes so the same failure is less likely next time.

Next action

Keep this inside the course path

Continue the lesson sequence, install the skill when one exists, or use DevelopJoy when you want the workflow wired into your real workspace.