TL;DR — A prompt gives you one answer. A loop runs on a trigger, does a job, checks the output, and retries what fails. When the check is automatic, it corrects itself, no babysitting. This week you build one.
In this issue
Two kinds of loop (and which one self-corrects)
Build it — 6 steps
Where it breaks
This week in AI
1 · Two kinds of loop
A prompt gives you one output. A loop is a small system: input comes in, the model does a job, an action happens, and a check decides whether to accept the output or send it back.
Every loop has four parts:
Trigger — what starts it (a new record, lead, or reply)
Job — the one thing the model does (extract, structure, classify)
Action — what happens with a good output (save, route, send)
Check — the test that accepts the output or sends it back to retry
Here is where most “AI loop” advice gets fuzzy. Whether the loop corrects itself comes down to one question: can a machine tell right from wrong?
Testable output (a format, a value from a fixed list, a total that reconciles, code that passes) → the check runs automatically and the loop fixes its own misses. No human in the loop.
Taste output (a good opening line, an on-brand caption) → there is no automatic test, so you are the check. Still useful, but that is a human-in-the-loop, not a self-correcting one.
The check is the loop. No automatic test, no self-correction.
This week we build the first kind: a loop that grades and corrects itself.
2 · Build it — 6 steps
Tool: Claude, or any LLM API · Time: one afternoon · Result: a loop that cleans messy records and fixes its own mistakes.
The task: turn a messy list of records into clean, structured data, where every row must match a schema.
1. Pick a task a machine can grade. This is the whole trick. A loop self-corrects only when “correct” is testable. If the only judge is taste, you are the checker. Pick a checkable task first.
2. Write the schema. This is your checker. Define a valid row: required fields, allowed values, formats. Company size is one of 1-10 / 11-50 / 51-200. Email matches an email pattern. Industry comes from your list. You write this once.
3. Lock the job and the output shape. Tell the model: take one messy record, return these fields in this exact structure. A loose output cannot be checked, so it cannot self-correct.
4. Wire the check and the retry. Run each output through the schema. Valid rows pass. Failed rows go back to the model automatically with the exact error (”company size was ‘medium’, must be one of...”), and it retries. Cap it at three tries so a bad row cannot loop forever.
5. Run the batch, watch the failures. Run fifty records. The loop fixes most misfires itself. Your only job is to read what still fails after three tries. Those are gaps in your schema, not the model’s fault.
6. Close the gaps, then step out. Add the rules the failures exposed. Now the loop enforces its own standard and corrects its own mistakes without you. That is the line between a prompt you babysit and a loop that runs.
The loop in one glance Trigger: a messy record → Job: structure it → Check: schema passes or retries → Action: clean row saved. You write the schema once. The loop corrects itself after.
3 · Where it breaks
No automatic check → not a self-correcting loop. Just a task with a human checker. Know which one you are building.
No retry cap → one bad row loops forever and burns tokens.
Vague schema → the loop passes garbage or rejects everything. The rules are the product.
Biggest model for a small job → burned money. A mid-tier model cleans rows fine. Which brings me to the news.
4 · This week in AI
The model treadmill sped up. New in the last few weeks: OpenAI’s GPT-5.6 “Sol” preview, Google’s Gemini 3.5 Pro (2M-token context) nearing release, Anthropic’s Claude Opus 4.8, and open-weight GLM-5.2. The read: you almost never need the top model. For the loop above, a mid-tier model clears the job for a fraction of the cost.
Buyers are comparing, not defaulting. Sensor Tower’s 2026 State of AI report puts ChatGPT below half the assistant market for the first time (46.4%), Gemini at 27.7%, Claude at 10.3%. The read: “we use AI” is no longer a differentiator. What you build with it is.
The plumbing is standardizing. Google, Microsoft, GitHub, NVIDIA and Salesforce shipped an open standard, Agentic Resource Discovery, so agents can find tools at runtime. The read: connecting your loops to other systems is about to get easier. Build now, wire later.
Your move
Build the self-correcting data loop this week, or adapt the pattern to any task a machine can grade. Same four parts: trigger, job, action, check.
Reply and tell me what you automated. I read every one.
Arslan
