Lesson 05: Catch and Correct

Lesson objectives:

• Send a specific correction back to the agent.
• Keep a mistake log with the mistake type, how you spotted it, and the corrected result.
• Record the first agent mistake you catch in your own agent chat.

Prerequisites: Lesson 04 | Previous << 04 Spot the mistakes | Next README >>

Finding the mistake is only half the job

Spotting a mistake is not the end of the task. It is the beginning of a loop. The agent can revise if you tell it exactly what to fix. A good correction is not "please do better." It is a pointed message that names the mistake, points to the evidence, and restates the acceptance criteria. This loop — spot, report, revise, verify — is what makes delegation reliable¹.

Explanation

A correction loop has four steps. First, you compare the output to the acceptance criteria you wrote earlier. Second, you describe the failure in specific terms. Third, the agent revises. Fourth, you verify the new output against the same criteria. This is the same feedback loop that Anthropic's agent-building guide recommends: human feedback at checkpoints keeps the agent from drifting off course¹.

This diagram shows the correction loop. The key is that the result does not go straight to use. It goes through the criteria, and if it fails, the loop sends it back for a specific fix.

A vague correction makes the loop longer. "Fix this" gives the agent room to guess again. A specific correction names the location, the error, and the standard. For example: "The output lists three action items, but the transcript has five. Add the two missing items: the budget review and the client check-in. Each must have an owner and a deadline."

A mistake log is a simple record of what you caught and how you fixed it. It can live in a note, a spreadsheet, or a document. The log turns one mistake into a reusable pattern. After a few entries, you will see which mistakes your agent makes most often and which acceptance criteria you keep forgetting to write.

A good mistake log entry has four parts:

1. The original ask — what you asked for.
2. The mistake type — hallucination, stale facts, scope shrink, or non-answer.
3. How you spotted it — the comparison you made against the source or the criteria.
4. The corrected result — what the output looked like after the correction, or what you changed in the ask to prevent it.

Worked example (follow along)

Maya asked her agent to turn a meeting transcript into a list of action items with owners and deadlines. Her acceptance criteria were: every item has a specific owner, every deadline is a calendar date, and every item is one line.

The agent returned:

This is the same example from Lesson 03, but now Maya sees it through the lens of the four mistake types. She labels it: scope shrink and non-answer. The agent changed the task from "specific owner" to "team owner" and from "calendar date" to vague phrases. It also did not answer the requirement that each deadline be a date.

She sends this correction:

"The output does not meet the acceptance criteria. First, 'Design team' is not a specific owner. Replace it with a person. Second, 'ASAP' and 'next sprint' are not calendar dates. Replace them with month/day/year dates. Third, '15th' has no month or year. Add the full date. Return the corrected table."

The agent returns:

Maya verifies the new table against the same criteria. It passes. She records the entry in her mistake log.

Your turn (faded example)

You asked an agent to summarize a contract in exactly three sentences, mentioning the price, duration, and cancellation clause. The agent returned:

"This contract is for software services. It runs for one year and can be cancelled. The price is competitive and listed in the attached document."

Fill in the correction message.

The output has three problems. First, it does not state the __________ price; it only says "competitive." Second, the cancellation clause is mentioned as __________ instead of quoting the clause itself. Third, the duration is given as __________, which is fine, but the price and cancellation need to be fixed. Please return a revised summary that states the exact price, quotes the cancellation clause, and keeps the duration.

Answer: First blank: exact. Second blank: "can be cancelled" or vague. Third blank: one year.

Summary + what's next

Delegation is not a single transaction. It is a loop. You write a clear ask, check the result against your criteria, name the mistake type, send a specific correction, and keep a log. The log is the real product of this course: it turns every caught mistake into a better ask next time. You now have the full cycle: clear ask → checked result → spotted mistake → corrected output.

Footnotes

1. Anthropic: Building Effective AI Agents — https://www.anthropic.com/engineering/building-effective-agents ↩ ↩²