NewAI Content Creation is now live in Early Access
Turning Point Training Instituteby Training Center
Skip to content
0%
Module 4 · Build, review and repairLesson 2 of 2
Video lesson

L08 · Independent review and repair

IntermediateVideo lesson
7 min readIntermediate
Downloads & resources
Download this lesson's materials, ready to use.

Skill packs and prompts are provided in English so they work best with the AI tools. Use them as-is.

Independent review and repair

Learning time: about 44.3–71.9 active minutes. Choose ONE presentation: the text (about 6.9 minutes, estimated) OR the lesson video (4:20, measured). Then 40–65 minutes of practice and checking (estimated). Do not count both presentations. Setup delays and optional extra study are separate; learner timing has not yet been piloted.

Example prompt

Where you are: Module 4, lesson 8. Your builder has completed one bounded task. You will give the second agent enough evidence to examine it independently.

What you will do

Prepare a review handoff, investigate a finding, make a controlled repair and recheck the changed result. Save a review record that distinguishes a reported problem from a confirmed problem. This continues outcome O4 and introduces O7, evidence-based checking.

Teaching

A reviewer needs the objective and actual files, not just the builder's summary. If Claude Code built the result, Codex can review it, or you can reverse their roles. Provide the same facts and acceptance checks. Ask the reviewer to explain each suspected failure with a file location, reproduction step and consequence.

Use an exact reference for what is reviewed: a commit, a selected change or a preserved file version. If the builder keeps editing while the reviewer reads, the two may discuss different text. Pause building during this first review. A current review cannot be inherited by later edits without examining the affected changes.

Request read-only review and check the actual tool permissions. Codex documentation describes local review without changing the working tree; a normal chat request still needs a clear boundary. A review comment, a GitHub review submission and a merge are separate actions. This lesson authorizes only local analysis and a saved local note.

Two models can miss the same problem. They may repeat the same assumption, trust a misleading test or agree that an incorrect answer sounds reasonable. You still run the checks and inspect the result. Independent review here means separate examination of the evidence; it is not a guarantee of independent errors or professional assurance.

Worked example

This example is author-created and simulated. Suppose a draft budget note says the Workshop Launch Kit breaks even at eight attendees. The preserved facts are fixed cost USD 180, USD 8 variable cost, USD 30 price and capacity 12. A reviewer recomputes eight attendees: revenue USD 240, total cost USD 244, loss USD 4. That concrete mismatch confirms the finding.

Use a complete instruction such as:

text
Review my fictional Workshop Launch Kit calculation note without editing
the project. Identify the folder, current revision and any uncommitted
changes first. Read the project instructions, brief and
build-notes/budget-spec.md. Do not rely on the builder's summary.

Required facts: fixed cost USD180, variable cost USD8 per attendee,
price USD30, capacity 12. Check formulas, whole-person break-even,
8- and 12-attendee results, and input limits. Interest is fictional,
not bookings. Check whether the task changed files outside its scope.

For each finding give the exact file and passage, expected behavior,
observed behavior, a reproducible calculation or check, and consequence.
Distinguish confirmed defects, questions and optional improvements.
List checks not performed. Return the review in this chat. Do not edit,
push, post comments, merge, send, purchase or deploy anything.

After confirming the defect, the owner asks the builder to correct the break-even result only. The owner checks the repaired note and the related examples. A fresh review then records that nine is the first whole attendee count to cover costs. It also states that the calculator itself was outside this review.

Do it yourself

Review this separate fictional draft without looking at the worked-example calculation: fixed cost USD 120, variable cost USD 4, price USD 16, capacity 12. The draft claims “Break-even is 10; at 10 attendees surplus is USD 16; at capacity surplus is USD 24.” Its completion note says “All calculations are correct.”

Write a finding record for each genuine defect. Give the precise statement, your arithmetic, consequence and repair scope. Then write a repair instruction that preserves the supplied facts. Treat stylistic preferences as optional. If using a second live agent, provide this full case and compare its findings with your arithmetic before accepting any edit.

Feedback and retry

Pause here before checking the feedback.

Break-even 10 is correct: contribution is USD 12 and fixed cost is USD 120. At 10, revenue and cost both equal USD 160, so surplus is zero. At 12, revenue is USD 192, cost USD 168 and surplus USD 24. Only the ten-attendee surplus claim needs numerical repair. The broad completion claim is also unsupported by the faulty result.

A suitable repair instruction identifies that sentence and asks the builder to recheck all three figures without changing inputs. Do not “fix” the correct break-even number just because another sentence was wrong.

Retry: a new draft uses fixed cost USD 90, variable cost USD 5, price USD 20 and capacity 8. It states “At five attendees surplus is zero; break-even is six; at eight surplus is thirty.” Find the defect and specify a focused recheck. Five attendees produce a USD 15 loss; the other two claims are correct. Evidence should preserve the correct statements and repair only the wrong one.

Check your understanding

Choose an answer and write one sentence explaining why before you read the answers.

  1. An agent finds no defects, but it could not open the spreadsheet supplying the figures. Which review conclusion fits?
    • A. Approved because no issues were found.
    • B. Source figures verified.
    • C. Limited review; input verification remains open.

Answers

  1. C. The reader may have checked prose or formulas, but it cannot certify source values it never read. Record the missing access and complete that check before claiming it passed.

Build the reviewed specification before continuing

The first build in L07 produced a specification, not a working calculator. After reviewing and correcting that specification, authorize this separate implementation task before L09–L11. Start from the same confirmed project folder and inspect its state first:

text
Implement the reviewed budget specification in calculator.html in this confirmed teaching workspace. Read PROJECT-BRIEF.md, budget-spec.md and the supplied TEST-CASES.md first; if names differ, identify the actual files before editing. Create a single local HTML file with no external libraries, network calls, tracking, payments or backend. Use labelled inputs and an explicit USD result. Enforce whole-person capacity/attendance and reject blanks, negatives, attendance above capacity, unsupported money precision and ranges. Use integer cents for supported two-decimal money. Support at most USD1,000,000 per money input and capacity100,000. Handle price at or below variable cost explicitly. Do not alter inputs to make the decision positive. Edit only calculator.html and a new IMPLEMENTATION-TEST.md. Run the supplied normal/invalid/cents tests if the environment permits; state exactly which were executed and which were only inspected. Do not commit, publish, send or spend. Stop with the files, actual results and remaining limitations for my review.

Open the actual file locally yourself. Confirm the default result and an invalid-input response. Give the second agent the implementation and test log for a read-only review using this lesson's evidence method. Repair confirmed defects and repeat the affected checks. Preserve the accepted checkpoint. If the implementation or review is incomplete, continue reading but do not claim the later parallel or delivery exercise has a working baseline. L13 will deepen these checks; it is not the first time you open the tool.

Summary

Keep the reviewed reference, confirmed findings, repairs and recheck evidence. Your final decision should say what passed, what remains unchecked and which version the decision covers.

Sources