How to Review Codex Results and Finish Safely

Codex Beginner Path · Step 5 of 5

• Previous: Organize product materials
• Goal: decide whether the result is correct, request targeted revisions, and close the task safely.
• Estimated time: 10 minutes.

💡 A message that says “done” is not proof. Completion means the output, source files, and operation record all pass your checks.

1. Review from Four Angles

1.1 Scope

Confirm that only the allowed workspace and files were used, the sample or batch size was respected, and no unrelated folder entered the task.

1.2 Counts

Compare source count, output count, successful items, skipped items, and needs-review items. Any mismatch needs an explanation.

1.3 Content

Check filenames, classifications, fields, units, language, formatting, and unsupported claims. Product specifications, certifications, prices, lead times, and customer details require evidence.

1.4 Operation Record

Ask Codex for a delivery report that can be checked:

Create a delivery report for this run. List:

1. files read;
2. files created;
3. files modified;
4. files copied or moved;
5. files skipped;
6. failed items;
7. content that still needs human confirmation.

Do not accept a summary such as “organization completed” when it cannot be matched to specific files.

1.5 Verify with Two Evidence Layers

The AI's delivery report is the first evidence layer. The second is independent human inspection outside the chat in File Explorer or Finder.

AI delivery report
→ open the output folder yourself
→ compare counts, names, modification times, and source files
→ accept the run only when both layers agree

At minimum, verify:

• the number of visible output files matches the delivery report;
• filenames, locations, and modification times are reasonable;
• source files still exist in their original location;
• no unexpected file appeared outside the approved task;
• a needs-review item was not presented as a confirmed fact;
• the Recycle Bin or Trash does not contain material from this task.

An AI report is not disk-level proof

A chat message saying “nothing was modified” or “nothing was deleted” is the AI's own account of the run. It helps define what to inspect, but it cannot replace File Explorer, Finder, Git, or a complete before/after manifest as independent evidence.

Boundary of the Windows reference evidence

The screenshots came from an already-installed Windows 11 environment using CC Switch with the third-party Laifa AI configuration for a one-file sample reference test. File Explorer independently confirmed that the two visible output files matched the AI report. However, no complete before/after manifest was generated. This is therefore partial reference evidence: it does not prove that the entire computer was unchanged, that files outside the visible output folder were untouched, or that OpenAI's official ChatGPT sign-in was verified.

1.6 Advanced Review: Compare Before and After Manifests

For a larger task, visual inspection can miss changes. A stronger method records each file's relative path, size, modification time, and SHA-256 before and after the task, then compares additions, modifications, overwrites, and deletions.

This public teaching graphic was derived from the August 3, 2026 machine file manifests and contains no internal run or repository-version identifier. At the file-difference layer, one, two, and three inputs corresponded to the same number of observed additions; four inputs corresponded to three observed additions; an existing same-name output was not overwritten; and source-file hashes did not change.

File-layer results do not mean the whole tutorial passed

The full run remains partial. It did not prove that the real Codex UI caused every disk write, it did not include Chinese step-by-step interaction evidence, and it did not prove that 01-originals was never read or that nothing outside the test root was accessed. The correct interpretation is: the file-difference layer passed, while end-to-end causality and read-boundary evidence still require retesting.

A beginner does not need to write a manifest script on the first day. The transferable principle is enough: the AI report is one evidence layer; independent disk inspection is another, and the second layer should become stricter as risk increases.

2. Give Precise Revision Instructions

Avoid “do it again.” State the location, problem, expected rule, allowed scope, and what must remain unchanged.

Please revise only the three sample files.

Problem: Sample 2 was classified using its filename, but the model inside the specification says Product B.
Expected rule: use the content of the specification as evidence; otherwise mark needs review.
Do not touch source files or the remaining files.
Return an updated sample table and change log, then stop.

3. Safe Tasks for Further Practice

Good next exercises include renaming copies under a confirmed rule, extracting fields from a few documents, finding missing material, preparing website drafts, copying images into categories, and creating a questions-for-review list.

Continue to use: sample first, human confirmation, then a larger scope.

4. Actions That Require an Explicit Stop

File operations

Deletion, overwriting, bulk moving or renaming, computer cleanup, and system-setting changes.

External actions

Sending email or messages, uploading material, publishing a website, submitting forms, and creating or deleting business data.

Accounts and money

Entering or changing API keys, switching paid services, adding funds, changing permissions, or starting a batch job that may incur charges.

AI confidence is not authorization.

5. If Codex Makes a Mistake

1. Stop the current task immediately.
2. Review the report and changed files.
3. Confirm that original material still exists.
4. Restore from copies or version history when needed.
5. Identify whether the cause was unclear instructions, excessive scope, or an execution error.
6. Reduce the scope and run a new sample.
7. Take over manually when the state is uncertain.

Do not repeatedly send “try again” while the file state is unclear.

6. Beginner Path Completion Check

• Install and sign in through official channels.
• Create a dedicated workspace.
• Avoid opening an entire Desktop or drive.
• State goal, scope, steps, red lines, and deliverables.
• Request inventory and restatement first.
• Validate a small sample.
• Review scope, counts, content, and operations using both the AI report and independent disk inspection.
• Confirm deletion, sending, upload, or publication separately.
• Stop instead of guessing when the result is uncertain.

7. How to Ask the Community for Help

If you are stuck on workspaces, task briefs, review, or safety boundaries, scan the QR code and add us on WeChat. Send “Codex” plus:

1. the step you are learning;
2. the task you want to complete;
3. the full error or message you see;
4. what you already tried.

Redact account details, API keys, customer information, and other sensitive data before sending a screenshot.

8. Choose the Next Topic

You do not need to continue through one permanent global sequence:

• for documents and files, return to the AI learning paths;
• for multiple API configurations, choose the optional CC Switch guide;
• prospecting, Laifaxin, AllinCMS, and SEO will be developed as separate advanced paths rather than added to the beginner requirement.

← Return to Step 4: organize product materials