Skip to content
The AI Agent Factory

Exhibits

Approved by
Panaversity ·
30 min read

Recorded on November 8, 2026. Every exhibit shows what Nisha actually did in the practical. Nothing here comes from the 2.0 lab, which is your exercise, not hers.

All the facts in these exhibits come from the practical log, the six draft files and ChatGPT's review file.1

Exhibit 1: Nisha's method: six steps, with human checks in between

Nisha's method, in two rows. Row one: Step 1, Claude writes a first draft. Human check 1, Nisha reads it, cuts the scope and fixes mistakes. Step 2, Claude grades its own work out of 10. Human check 2, Nisha sets rules before Claude works alone. Step 3, Claude improves by itself, round after round. Row two: Step 4, ChatGPT reviews and grades, with exact fixes. Human check 3, Nisha sorts the points into accept, change and not now. Step 5, Claude uses the review to improve again. Step 6, a person checks the facts and decides when the work is done. The three human checks are shown in gold, and step 6 has a red border as the final human decision. A band explains the colors. A callout says the six steps come from the method taught in class, and the three checks were added by Nisha.

StageWhoWhat happensWhy it changes the decision
Step 1ClaudeWrites a first draft.Gives Nisha something concrete to react to. Her decisions on scope and voice came from reading it.
Human check 1NishaReads the draft, answers Claude's questions, adds missing context, cuts the scope and fixes mistakes before any grading.Sets the scope before anyone grades, so later grades judge the book she wants.
Step 2ClaudeGrades the draft out of 10 in three areas: accuracy, quality, and how easy it is for a beginner to understand.Shows which area is weakest, so she knows what to fix first.
Human check 2NishaSets clear rules before Claude works alone.Turns her decisions into written rules, so Claude's solo rounds stay inside them.
Step 3ClaudeImproves the draft, aiming for 9.5 in each area. It improves, grades again and repeats the rounds by itself, without being asked for each one.Fixes many problems without her, and shows where self-review stops.
Step 4ChatGPTReviews and grades the result and its sources, with specific fixes.A reviewer that did not write the draft finds different problems, so she does not rely on one judgment.
Human check 3NishaSorts ChatGPT's points into accept, change and "not now".Keeps the decision with her, so the review does not choose the changes or grow the work.
Step 5ClaudeUses ChatGPT's review to improve the work again.Turns accepted advice into a new version, and lets Claude challenge points it cannot confirm.
Step 6NishaA person checks the facts against their sources, reviews any problems still open, and decides when the work is ready.Replaces judgments with evidence. Only official sources show which disputed claims are right, and only she decides when to stop.

Steps 1 to 6 come from the method taught in class. The three extra human checks were Nisha's own addition.

Exhibit 2: Nisha's key prompts

The prompts below are copied exactly as they were sent.

a) First prompt to Claude (step 1)

I am a chartered accountant from Pakistan (ICAP) and a gold medalist. My name is Nisha. I have identified a problem: there are two main sets of accounting rules in the world. The first is IFRS, which is used internationally. The second is US GAAP, which is followed in America. IFRS is principle-based, while US GAAP is rule-based. Chartered accountants around the world are skilled in IFRS, but American companies follow US GAAP.

I want to write a book that helps accountants on both sides of the fence understand the other set of rules and make the move from one to the other.

My readers are:
- IFRS-trained accountants, like chartered accountants in Pakistan, who want to work for US companies or Big 4 firms
- US GAAP-trained accountants who need to work with IFRS

Please draft:
1. A basic concept or description of the book
2. An outline of the chapters

b) Grading prompt to Claude (step 2, sent in the same message as Nisha's answer about the gold medal line)

Now please review the current version of the document and grade it out of 10 in three areas:
1. Accuracy: Are the facts, dates, rules and standard numbers, correct? Check each standard number in brackets, and tell me if you are unsure about any of them.
2. Quality: Is it well organized, complete and professional? Would a publisher or a Big 4 hiring manager take it seriously?
3. Ease of understanding: Is it easy to follow for my readers? They know basic accounting but are new to the other framework, and many speak English as a second language.

For each area, give:
- The score out of 10
- The reasons for the score
- The specific problems, and where they are (for example, "Preface, paragraph 2" or "Chapter 12")
- What would need to change to reach 9.5

Be strict. Do not give a high score unless the work earns it.

Do not change the document yet. Only grade it.

c) The self-improvement instruction to Claude (end of the step 3 prompt)

Then improve the document by yourself: grade it again in the same three areas and repeat until every area reaches 9.5. Stay strict. Do not raise a score unless the problem is really fixed.

After each round, show a short table with the round number, the three scores and the main changes. Save the final result as version 3.

d) Review prompt to ChatGPT (step 4, shortened)

I am Nisha, a chartered accountant from Pakistan (ICAP). I am writing a book called "The Bilingual Accountant: Moving Between IFRS and US GAAP". [...] I have attached the preface and table of contents (version 3). Please review it as a strict expert who knows both IFRS and US GAAP.

Decisions I have already made. Please don't reverse them: [four decisions listed]

Please grade it out of 10 in three areas: accuracy, quality, ease of understanding. Open each source link and check that it supports what the document says.
For each area, give the score and why, each problem and where it is, and the exact fix.
At the end, list any facts or sources you could not check, and anything you are unsure about.
Be strict. Do not rewrite the whole document. Only list the problems and fixes.

The full prompts are in the practical log.

Exhibit 3: Every grade, from first draft to final version

Every grade, two reviewers. A line chart of grades out of 10, with a dashed line at the 9.5 target. Claude's own grades for accuracy, quality and ease start at 7, 6.5 and 7 for version 2. They rise to 8.5, 8.0 and 7.5 in round 1, 9.5, 9.2 and 9.0 in round 2, and 9.5, 9.5 and 9.3 in round 3. A shaded column shows ChatGPT's grades for the same version 3 as hollow marks: accuracy 8, quality 7 and ease 8. After it, Claude's grades for version 4 are 9.3, 8.8 and 9.0, and for version 5 they are 9.2, 9.0 and 9.3. Ease never reaches 9.5 after round 3. A note says all grades are the AI tools' judgments against Nisha's prompt, not probabilities of correctness or test results, that the two reviewers saw different information, and that version 6 was not graded. A label points out that accuracy and quality are both 9.5 at round 3, so one marker sits on top of the other.

VersionGraded byAccuracyQualityEaseWhat was happening
Version 2Claude76.57Step 2: Claude grades its own work
Version 3, round 1Claude8.58.07.5Step 3: Nisha's decisions applied
Version 3, round 2Claude9.59.29.0Step 3: fixes and fact checks
Version 3, round 3Claude9.59.59.3Step 3: final polish, and Claude stops
Version 3ChatGPT878Step 4: independent second opinion
Version 4Claude9.38.89.0Step 5: after using ChatGPT's review
Version 5Claude9.29.09.3Step 6: Claude checks Nisha's edits (last grade)

All grades were given by the AI tools against Nisha's prompt and her three areas. They are judgments, not probabilities of correctness and not test results. A 9.5 does not mean the draft is 95% right, and a lower grade does not make a reviewer more reliable. The reviewers also saw different information, so their grades are not a fair comparison. The target was 9.5 in every area. Version 6 (the final version) was not graded, on Nisha's instruction.

A score is a reviewer's judgment. Evidence is what makes the judgment defensible. When reviewers disagree, Nisha must identify their assumptions, inspect their reasoning and verify the disputed claims.

Exhibit 4: How the draft changed

The draft grew back. A bar chart of words in the file for each version, with a line for the number of sources. Version 1 by Claude: 2,470 words and 4 sources. Version 2, after human check 1: 1,531 words and 4 sources. Version 3, after steps 2 and 3: 2,247 words and 9 sources. Version 4, after step 5: 2,606 words and 15 sources. Version 5, by Nisha: 2,611 words and 15 sources. Version 6, final: 2,599 words and 15 sources. The bars for versions 2 and 5 are gold, because they came after Nisha decided. The final version is dark. A note says words are counted for the whole file, including the list of sources.

VersionMade byWhat changedWordsSources
1ClaudeFull book concept: market, readers, 25 detailed chapters, open decisions2,4704
2Claude, after human check 1Cut to preface and table of contents, in Nisha's voice and plain English1,5314
3Claude, after steps 2 and 3Key to short forms, mistakes fixed, plainer titles2,2479
4Claude, after step 523 of ChatGPT's 27 points used, most of them rewritten2,60615
5Nisha6 groups of small edits (8 places): shorter lines, fewer hard terms2,61115
6Claude, on Nisha's instructionOne unconfirmed note removed: final version2,59915

Words are counted for the whole file, including the list of sources.

Exhibit 5: Who caught what

Who caught what. A colored grid of the 12 problems in Nisha's private notes, with one column for Claude's self-review and one for ChatGPT's review. Gold means caught, light gold means partly or caught but not fixed, red means missed, gray means already fixed before ChatGPT saw it, and a hatched box means ChatGPT did not see that version. Claude caught problems 1, 2, 3 and 5, caught but did not fix 4, partly caught 7 and 9, and missed 6, 8, 10, 11 and 12. It added problem 8 itself. ChatGPT found 1, 2, 3 and 5 already fixed, caught 4, 6, 7 and 11, partly caught 8, missed 9 and 10 and made both worse, and did not see version 4, where problem 12 was found. A note says notes 1 to 6 came from version 2, notes 7 to 11 from version 3, and note 12 from step 6, and that this was not a controlled test.

Warning

Not a controlled test. Claude and ChatGPT saw different versions and different information. Claude reviewed its own drafts inside the project where it wrote them. ChatGPT saw only version 3 and its sources, and never saw version 4. Read the counts below as a record of one evening, not as proof that one reviewer is better.

Nisha's private notes, compared with the two AI reviews.

#Problem Nisha notedClaude's self-review (steps 2 and 3)ChatGPT's review (step 4)
1The preface said foreign companies in the US can use IFRS. This is too broad, because only some can.CaughtAlready fixed
2One source was from 2018.CaughtAlready fixed
3Chapter 22 pointed to the wrong US standard.CaughtAlready fixed
4Chapter 12 missed the US standard on equity investments.Caught, not fixedCaught
5The preface said "most books only list differences" without proof.CaughtAlready fixed
6Chapter 13 covers debt but missed the main US standard on debt.MissedCaught
7Chapter 5's brackets listed IAS 1 but not IFRS 18 (the chapter line itself mentioned IFRS 18).PartlyCaught
8Chapter 25 contained news that would soon be out of date.Missed (Claude added it)Partly
9Several chapter lines were much longer than one line.PartlyMissed, and its fixes made lines longer
10The preface paragraph on foreign companies was too technical for beginners.MissedMissed, and its fix made it more technical
11Chapter 18 said "some subsidiaries" without saying which ones.MissedCaught
12Chapter 22 (version 4) used "foreign private issuer", a term the book never explains. Found in step 6.MissedDid not see version 4

"Already fixed" means Claude had fixed the problem before ChatGPT saw the draft. For note 7, the practical log shows that Claude added a line about IFRS 18 to Chapter 5 in step 3 but did not add it to the chapter's list of standards, so it is marked "Partly". Notes 1 to 6 were made while reading version 2, notes 7 to 11 while reading version 3, and note 12 during step 6. This was not a controlled test, because the two reviewers saw different versions. The notes are recorded in Appendix A of the practical log.

Exhibit 6: Nisha's fact check (step 6)

#Fact in the bookChecked againstResult
1More than 140 countries and regions (jurisdictions) require IFRSIFRS Foundation, March 2025 roadmap (page 18)Confirmed
2US companies listed in the US use US GAAP, and some foreign companies can use IFRSSEC concept release 33-11376 (page 10), and the SEC 2007 ruleConfirmed
3IFRS 18 replaces IAS 1 from January 1, 2027, and early use is allowedIFRS Foundation, IFRS 18 pageConfirmed
4Start dates of the new US expense disclosure rule (ASU 2024-03)FASB, ASU 2025-01Confirmed
5A company that loses foreign private issuer status must switch to US GAAPSEC Financial Reporting Manual, Topic 6 (6120.2, 6120.4)Confirmed
6Which subsidiaries can use IFRS 19IFRS 19, paragraph 7Confirmed
7IFRS 1 covers adopting IFRS for the first timeIFRS Foundation, IFRS 1 pageConfirmed
8What the PCAOB overseesPCAOB, Oversight pageConfirmed
9IFRS S1 and S2 were issued in June 2023 and are separate from the accounting standardsIFRS Foundation, ISSB introductionConfirmed
10The SEC paused its climate rules and proposed to withdraw them on May 29, 2026SEC proposal page and press release 2026-49Confirmed
11US GAAP topic numbers, including ASC 320, 321, 835-20, 360 and 105-10-05-1FASB Accounting Standards CodificationConfirmed
12The plain-English lines still say the right thingNisha's own judgment as an accountant (not an official source)Accepted after review

Recorded item by item in Appendix B of the practical log. Nisha's final prompt sums up the result as "All confirmed".

Exhibit 7: The final result: table of contents (version 6)

The Bilingual Accountant: Moving Between IFRS and US GAAP, by Nisha

Part I: Background on the two frameworks

  1. The two frameworks and who uses them
  2. Principles-based and rules-based: what the words really mean
  3. How to find answers: the standards and the vocabulary
  4. US financial reporting: the SEC, auditors and controls

Part II: The main differences, topic by topic

  1. Presentation of financial statements and cash flows
  2. Revenue
  3. Inventory
  4. Property, plant and equipment, and investment property
  5. Intangible assets and research and development
  6. Impairment, assets held for sale and discontinued operations
  7. Leases
  8. Financial instruments I: classification, measurement and credit losses
  9. Financial instruments II: derivatives, hedging, debt and equity
  10. Provisions, contingencies and events after the reporting period
  11. Income taxes
  12. Employee benefits and share-based payment
  13. Foreign currency and hyperinflation
  14. Business combinations and consolidation
  15. Earnings per share, segments, interim reporting and related parties
  16. Rules for special industries

Part III: Moving from one framework to the other

  1. Converting financial statements between the frameworks
  2. When a company changes its framework
  3. Career guide A: from IFRS to US GAAP
  4. Career guide B: from US GAAP to IFRS
  5. What is changing

Appendices

  • A. Master differences and reference tables
  • B. Terminology dictionary
  • C. Journal-entry quick reference
  • D. Self-test and interview questions
  • E. Reading list
  • F. Excel conversion toolkit

The full version 6 file also has a short preface, a one-line description and the main standards for each chapter, a key to the short forms, and 15 official sources.

Exhibit 8: Time and cost

1 h 40 min69About Rs 60,000
From the first draft (22:09) to the final version (23:49)Versions of the documentPrompts: 7 to Claude, 2 to ChatGPTA month for both subscriptions (about Rs 30,000 each, made up for the case)
Saved atWhat was finished
22:09Version 1: Claude's first draft (step 1)
22:27Version 2: after Nisha's first check
22:50Version 3: after Claude's grading and three improvement rounds (steps 2 and 3)
23:11ChatGPT's review (step 4)
23:29Version 4: after Claude used ChatGPT's review (step 5)
23:36Version 5: Nisha's own edits (step 6)
23:49Version 6: final version (step 6)

Times are when each file was saved on November 8, 2026. The time spent writing the first prompt, before version 1, is not included.

Exhibit 9: Words to know

AI Worker. An AI tool given a job, a brief and limits, as described in The New Way of Working. In this case, Claude is the writer and ChatGPT is the reviewer.

Chartered accountant (CA) and ICAP. A CA is a highly trained, qualified accountant. In Pakistan, CAs qualify through the Institute of Chartered Accountants of Pakistan (ICAP).

Big 4. The four largest accounting firms in the world: Deloitte, EY, KPMG and PwC.

IFRS. International Financial Reporting Standards, the accounting rules used in most of the world. They are written by the IASB (International Accounting Standards Board).

US GAAP. US Generally Accepted Accounting Principles, the accounting rules used in the United States. They are written by the FASB (Financial Accounting Standards Board).

The Codification (ASC). The official database of US GAAP rules, organized by topic number. For example, ASC 330 is the topic on inventory.

SEC. The US Securities and Exchange Commission, which regulates companies listed on US stock markets.

Foreign private issuer. A non-US company listed in the US that meets the SEC's conditions. It may prepare its accounts under IFRS (as issued by the IASB) instead of US GAAP.

PCAOB. The Public Company Accounting Oversight Board, the US body that oversees the auditors of companies listed in the US.

Standard. One official rule document on one topic, such as IAS 2 (inventories under IFRS).

IAS 1 and IFRS 18. IAS 1 is the IFRS standard on how to set out the main financial statements. IFRS 18 replaces it for years starting on or after January 1, 2027.

IFRS 1. The IFRS standard for a company that starts using IFRS for the first time.

Subsidiary. A company that is controlled by another company, its parent.

Jurisdiction. A country or region that makes its own rules, such as which accounting standards its companies must use.

Listed company. A company whose shares are traded on a stock market.

Preface and table of contents. The preface is the short introduction at the start of a book. The table of contents lists the book's chapters.

Prompt. The instruction or question you type to an AI tool.

Project. A workspace in Claude or ChatGPT that keeps a task's chats and files together. Rebuilt Claude projects can also start separate threads, but Nisha did not use them.

Grading and self-improvement. Asking the AI to score its own work against clear areas, then to improve it and score it again, round after round.

Source and fact check. A source is the place a fact comes from. A fact check means opening the source and confirming that it really says what the document says.

ESL. English as a second language: readers whose first language is not English.

This case was prepared for class discussion. It is not meant to show whether the choices described were good or bad, and it is not a source of accounting advice.

  1. Nisha's AI Team: Practical Log, with the draft files listed in its Appendix C.↑
Owner
Panaversity