Day 2 Summary Codex AI Auditor PM Automation

Day 2. Codex Client-Based Auditing Practice Summary

Master how to combine Codex Workspace's reasoning with the key design control work that causes bottlenecks in practice: finding planning doc mismatches, analyzing the impact of requirement changes and extracting decision logs from unstructured messenger chats.

Part 1. Background
Part 2. Lab goals
Part 3. Follow along
Part 4. How it works
Part 5. Teaching guide

Why Codex Client for these three tasks?

In typical IT service projects, the biggest source of lost time and rework is **"planning/design mismatches caused by miscommunication"**. Let's see the difference between a general chat model like ChatGPT and Codex Client, which holds a whole project's files and keeps context.

The key difference between ChatGPT and Codex Client

Compared Regular ChatGPT (single chat) Codex Client (Workspace-based)
Analysis range Compares single or fragmented copy-pasted text Checks consistency across all docs in the project folder
History tracking Risk of losing earlier file context as the chat goes on Syncs mismatches between before/after requirements and downstream specs (API, UI) in real time
Cleaning transcripts Records only simple meeting results, or risks inventing owners Carefully separates the reasons (Why) and open items (Action Items) in unstructured chats

Productivity of using AI as a Reviewer

Prevention in software engineering

Fixing an error at the design stage is **up to 100 times** cheaper than fixing it after development and testing. AI catches design contradictions right away and prevents rework.

Thoroughness as a Reviewer

In seconds, AI accurately points out, based on facts, things that are tedious for people to check one by one, like missing inactive rules or crossed variable mapping (e.g. rater reversed).

Today's final output

Task 01. Find design mismatches

A mismatch summary report table sorting contradictions and gaps between meeting notes, requirements and screen design (output: consistency-report.md)

Task 02. Manage change impact

A change impact analysis table mapping specific action items to apply to screen UI plans and back-end API specs for a new requirement change (output: impact-analysis.md)

Task 03. Track decisions

A Decision Log and Pending Items table separating final agreements (Why) from pending matters needing review, taken from unstructured spoken/chat records (output: decision-log.md)

Mission 1 (Task 06). Impact analysis

An impact analysis report on the screens, APIs, DB and docs affected by adding phone verification at sign-up (output: impact-analysis-report.md)

Mission 2 (Task 07). Write an issue report

A project issue report summarizing current progress and key issues from a combined analysis of project material (output: project-issue-report.md)

Input / Output data pipeline

A map of the whole workflow: input files go through Codex Client and are refined into result tables during the lab.

01
Find mismatches
02
Change impact
03
Decision history
M1
Impact analysis
M2
Issue report
Task 01. Mismatch detection data flow
šŸ“„ Input (planning/design docs)
• meeting-notes.md (meeting notes) • requirements.md (requirements doc) • screen-specification.md (screen spec)
Codex Client āž”
šŸ’¾ Output (summary report)
| Type | Contradiction found | Impact | Source | Recommended action | | Policy | Refund fee differs on account deletion | High | Meeting notes | Align logic | | Feature | Wrong rater | High | Meeting notes | Fix FR-004 |

STEP 1. Task 01 - Find mismatches

After creating the lab files in the project folder, send the instruction below in the Codex Client prompt box to audit for contradictions.

šŸ” Codex request prompt (review instruction)

Compare and analyze the three docs 'meeting-notes.md', 'requirements.md' and 'screen-specification.md' in the project folder.

Find and list items that contradict or don't match between the docs, or items that exist in one doc but are missing from another.
Output the result to meet these requirements.

1. Organize the result as a Markdown table with the columns 'Type', 'Finding and contradiction', 'Impact', 'Highest-authority source doc' and 'Recommended action'.
2. Mark items that could cause serious rework when entering development or testing as 'high' impact.
3. Don't invent answers; base the comparison only on facts from the doc text.
  • Check that it caught the 10% penalty fee policy from the meeting notes missing in the requirements doc.
  • Check that it caught the policy mismatch across 3 docs for account deletion (auto-cancel vs keep stored vs hide without pop-up).
  • Check that it caught the logic error in FR-004 where the rater is written as Expert instead of Customer.

STEP 2. Task 02 - Change management

Prepare the 2 requirement files (v1.0 vs v2.0) and the existing design/API specs, then send the change impact mapping instruction.

šŸ”„ Codex request prompt (change and impact analysis instruction)

First compare 'requirements-before.md' and 'requirements-after.md' in the project folder and analyze which policies changed.

Then analyze where the changed policies must be reflected in 'screen-specification-v2.md' and 'api-specification.md', which are still at v1.0.

Output the result as a Markdown table in this format:
1. 'Change item', 'Before', 'After', 'Impact area (screen/API)', 'Importance', 'Action (specific fix guide)'
2. Also mark the importance of each needed fix, based on the v2.0 requirements, as 'high/medium/low' in a column.
  • Check whether it set a change guide for the login screen UI and API parameter (oauth_token) for Google social login.
  • Check whether it covered replacing the sign-up UI and adding required API Payload parameters (phone, verification_code) for SMS verification.
  • Check whether it covered designing a new dashboard screen and mapping an API role field for Admin permissions.

STEP 3. Task 03 - Decision tracking

Prepare the unstructured dev chat (messenger-chat.md), then automatically separate and extract the decision log and follow-up tasks.

šŸ“‹ Codex request prompt (decision extraction instruction)

Read the 'messenger-chat.md' transcript in the project folder and write a Decision Log and Pending Items for project management and tracking.

Analyze it under these conditions.
1. For settled matters, write a 'Decision Log' table. Columns must include 'Decision item', 'Decision', 'Reason/background', 'Decision date', 'Owner' and 'Status'.
2. Put matters not yet decided, needing follow-up or on hold into a 'Pending Items' table. Columns: 'Pending item', 'Follow-up action', 'Owner' and 'Due date'.
3. Don't make up dates or decision owners not based on the transcript text.
  • Check that it extracted the reason (review delay, etc.) and owner for dropping Apple login and launching with Google login only.
  • Check that it worked out the reason for holding real-time bank transfer, and the owner (dev team) and follow-up for reviewing overseas PayPal.
  • Check that it included the reason for approving 2-tier Super/Sub Admin permissions to meet the security audit.

Mission 1. Sign-Up Change Impact Analysis (Task 06)

For the customer's change request "add phone verification at sign-up," use Codex to analyze the impact on the whole system (screens, API, DB, docs) and write a report.

šŸ” Codex request prompt (impact analysis instruction)

If phone verification is added to sign-up, find every screen, API, DB and doc affected and write an impact analysis report.
  • Check that it detected the sign-up screen (UI) impact as 'high' with a 'fix screen' action.
  • Check that it detected the login/sign-up API impact as 'medium' with 'fix API'.
  • Check that it mapped the User DB impact as 'high' with an 'add column' action.
  • Check that it mapped the admin screen impact as 'low' with an 'add view field' action.

Mission 2. Write a Project Issue Report (Task 07)

Analyze raw project material — meeting notes, change history, Action Items, issue lists — and automatically write a report on current progress and key issues.

šŸ“‹ Codex request prompt (issue report instruction)

Summarize the current project's main issues and write a project issue report including urgency and priority.
  • Check that done (login) and in-progress (sign-up) status are sorted correctly.
  • Check that the payment API schedule delay is analyzed as 'high' priority, owned by the PM.
  • Check that the upcoming requirement change risk is mapped as 'high' priority, owned by the planning team.
  • Check that the open admin permission policy item is correctly set as 'medium' priority, owned by the planning team.

Semantic Diff

This isn't a simple line-by-line text comparison or change detection. It recognizes how the core logic of a business requirement was revised (e.g. email verification āž” SMS phone verification), and follows the dependency all the way to the variable mapping tables in dependent files (front-end UI wireframe spec and back-end REST API spec) to map a fix guide.

Capturing context inside unstructured chats

From informal group-chat talk like "let's talk about Apple at the next meeting" or "bank transfer is on hold for now because of the PG review," AI semantically extracts the reasons that led to the final agreement (Apple developer annual account review schedule, PG contract approval delay, etc.). This way, even if project members change, the background (Why) of meeting-room decisions is fully remembered.

How the three Tasks work together

Planning/design review isn't a one-off task. It acts as a design control system that closes the gap between planning and development across the IT project lifecycle.

1. Find contradictions (Task 01)

Spot crossed conflicts or holes between existing plans, screen designs and meeting notes in seconds.

2. Track decisions (Task 03)

Report the reasons in meeting chats that led to approving requirement changes, plus open work.

3. Apply control (Task 02)

Final control so decision results are mapped to downstream UI screens and API specs with nothing missed.

Course design specs

Item Detailed course spec
Difficulty ā˜…ā˜…ā˜…ā˜†ā˜† (intermediate)
Lab time 1h 30min ~ 2h (including combined lab and feedback)
Real-world use ā˜…ā˜…ā˜…ā˜…ā˜… (IT project consistency checks and rework prevention)

Final course Task for the planning/design stage

Task Topic AI use
Task 01 Find mismatches Comparing and checking differences between docs
Task 02 Finding gaps Check for missing features, policies and exceptions
Task 03 Change management Change history management and impact analysis
Task 04 Doc version control Compare versions and manage the latest doc
Task 05 Decision tracking Manage meeting decisions and Action Items
Task 06 Impact analysis Analyze impact of change requests
Task 07 Project issue report Automatic project status and issue reporting

Learner questions and discussion topics

"What's the cost difference between doing these three Tasks right after a meeting and having people compare docs by hand during or after development?"

→ Checking specs early can save up to about 100 times the cost and effort compared with handling rework by tearing up code after development is done, strongly protecting the launch schedule.

"How well does AI's impact analysis catch minor fields developers might miss (e.g. API Request Payload)?"

→ You'll see how accurately it identifies input parameters people easily forget in API specs on a busy schedule, like social login tokens and phone number formats for SMS verification.

← Enter Day 2 Classroom View full roadmap