How to Review Codex Results and Finish Safely
- Previous: Codex hands-on practice: organize product materials
- Goal: learn to determine whether results are truly complete and safely finish the task.
- Estimated time: 8–12 minutes.
Step 4 · Complete Your First Task completed your first practice sample → this guide covers the four review dimensions, revision principles, and high-risk boundaries to wrap up safely.
💡 Core takeaway: When Codex displays “complete,” it only means this execution cycle is finished. Whether the output is truly usable must be verified by your own inspection of files and content.
1. Review from Four Angles
1.1 Scope: Did It Touch Files It Shouldn't?
Check:
- whether it only processed the designated directory;
- whether it only touched approved files;
- whether it read unrelated materials;
- whether source files were deleted, moved, or overwritten;
- whether any out-of-scope operations occurred.
1.2 Counts: Do Numbers Match Before and After?
For example:
Source files: 5
Successfully copied: 4
Needs review: 1
Failed: 0
If numbers do not match, do not let AI brush it off with “some might have been missed.” Require it to list specific filenames.
1.3 Content: Is There Evidence for Facts?
Key items to verify:
- whether product parameters have reliable sources;
- whether models are misidentified;
- whether inferences are written as facts;
- whether important fields are omitted;
- whether professional-looking but unverifiable content was generated.
For certifications, prices, delivery dates, company information, and client details, anything lacking a source must be marked as pending verification.
1.4 Operation Record: What Exactly Changed?
Ask Codex to provide:
Please generate a delivery report for this round, listing:
1. files read;
2. files created;
3. files modified;
4. files copied or moved;
5. files skipped;
6. failed items;
7. content that still requires human confirmation.
Do not simply accept an unverifiable summary like “organization completed.”
1.5 Verify with Two Evidence Layers
The AI's delivery report is the first layer of evidence. The second layer is leaving the chat interface and opening File Explorer or Finder to inspect independently.
AI delivery report
→ manually open output directory
→ check counts, names, modification times, and source files
→ both agree before accepting this round
Manually inspect at least:
- whether output file counts match the delivery report;
- whether filenames, directories, and modification times are reasonable;
- whether source files remain in their original locations;
- whether extra files appeared outside the approved task;
- whether “needs review” items were mistakenly written as confirmed facts;
- whether materials from this run appeared in the Recycle Bin or Trash.
A chat message stating “nothing was modified” or “nothing was deleted” is merely the AI's explanation of its run. It helps define what to inspect, but cannot replace File Explorer, Finder, Git, or before/after manifests as independent evidence.
The screenshot comes from an already-installed Windows 11 environment using CC Switch with the “Laifa AI” configuration for a single-file sample reference test. File Explorer independently confirmed that 2 output files matched the delivery report. Because a full before/after manifest was not generated, this remains narrow reference evidence, not proof that “nothing changed” across the entire computer.
1.6 Advanced Review: Compare Before and After Manifests
When file counts grow, manual inspection can easily miss items. A more rigorous method records relative path, file size, modification time, and SHA-256 before and after the task, then automatically compares additions, modifications, overwrites, and deletions.
This public teaching graphic was compiled from the August 3, 2026 machine file manifests and excludes internal run or repository version identifiers. Observed results at the file layer: 1, 2, and 3 inputs matched the same number of new outputs; 4 inputs produced only 3 new outputs; existing same-name outputs were not overwritten; original input SHA hashes remained unchanged.
The full run remains partial: it did not prove the real Codex UI triggered every disk write, lacked Chinese step-by-step interaction evidence, and did not prove 01-originals was never read or that no access occurred outside the test root. The proper takeaway: the file difference layer passed, while end-to-end causality and read boundaries await further retesting.
Beginners do not need to learn how to write manifest scripts right away, but should grasp the underlying principle: the AI's response is one layer of evidence; independent disk checks are another layer. The higher the risk, the stricter the second layer must be.
2. How to Revise After Finding Errors
Do not simply say:
Done wrong, do it again.
Instead, point out:
My inspection found three issues:
1. Two images for A100 were placed into B200;
2. Product use cases in README have no source material;
3. Original files should not have been renamed.
Please explain the cause and the files you plan to modify first.
Only modify these three points; do not reprocess other confirmed items.
Output an updated change list after modifying.
The core principle:
Point out specific problems, narrow the rework scope, and check once more.
3. Safe Tasks for Further Practice
After completing the foundation path, you can continue practicing:
- organizing copies of product materials;
- compiling Markdown, TXT, or small CSV files;
- comparing two document versions;
- generating file inventories;
- checking for missing materials;
- drafting articles for websites;
- copying and classifying images under established rules;
- creating lists of unresolved questions for review.
Always follow: sample first, human confirmation, then expand scope.
4. Actions That Require an Explicit Stop
High-Risk File Operations
- deleting files;
- overwriting original files;
- bulk moving or renaming;
- cleaning up the entire computer;
- modifying system settings.
High-Risk External Actions
- sending emails or messages;
- uploading materials;
- publishing websites;
- submitting forms;
- creating, modifying, or deleting business data.
Account and Cost Actions
- entering or modifying API keys;
- switching paid services;
- topping up balances;
- modifying account permissions;
- running bulk tasks that may incur fees.
These actions must never be approved automatically just because the AI “looks confident.”
5. What If Codex Makes a Mistake?
Follow these steps in order:
- Stop the current task immediately, do not expand scope;
- Review the delivery report and file modifications;
- Confirm that original materials are still intact;
- Restore from copies or version control history;
- Identify whether the issue was an unclear brief, excessive scope, or execution error;
- Narrow the scope and re-run a small sample;
- Take over manually whenever uncertain.
Do not repeatedly send “try again” while the outcome is unclear, which can cause duplicate modifications.
6. Beginner Path Completion Check
If you can accomplish everything below, you have completed Codex beginner onboarding:
- can install and sign in to Codex through official channels;
- can create a dedicated working directory;
- avoid choosing the entire Desktop or hard drive from the start;
- can clearly write task goal, scope, steps, and red lines;
- ask AI to inventory and restate first;
- run a small sample first;
- check scope, counts, content, and operation records;
- know to confirm before deletion, sending, upload, or publishing;
- know to stop rather than blindly retry when results are uncertain.
📋 Common Pitfalls
| Pitfall | Consequence | How to avoid |
|---|---|---|
| 🔴 Telling AI to "fix everything along the way" upon finding errors | AI broadens the scope and breaks previously working results, creating secondary bugs | Point out specific files and issues, restrict fixes only to those points, and require a diff report |
| 🔴 Skipping disk-level review outside the AI chat interface | Hidden accidental file moves, deletions, or corruptions slip undetected into production | Combine the AI delivery report with independent File Explorer/Finder checks on counts and timestamps |
| 🟡 Accepting completion based purely on AI claiming "done" | Summaries may gloss over omitted items or errors, giving a false sense of success | Systematically verify all four dimensions: scope, counts, factual basis, and operation records |
| 🟡 Repeatedly sending "retry" commands when errors occur | Piles up duplicate edits and creates conflicting versions, complicating recovery | Stop the task immediately, inspect the file system, and restore from backups or version control |
| 🟢 Accepting unsupported AI inferences as verified facts | Specs, pricing, and delivery dates become distorted, causing downstream business damage | Mark any claim lacking a direct cited source as "needs review" and defer to human decision |
Congratulations on finishing the Codex five steps! You have mastered the four universal methods: safe workspaces, clear task briefs, small samples first, and four-dimensional review. Now return to AI Learning Paths to choose your next practical scenario.
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:
- the step you are learning;
- the task you want to complete;
- the full error or message you see;
- what you already tried.
Redact account details, API keys, customer information, and other sensitive data before sending a screenshot.
8. Where to Go After Finishing the Foundation Path?
You do not need to follow a fixed numbering to learn everything; choose based on real work needs.
Only Want to Continue Processing Files and Documents
Return to AI learning paths, and select everyday work scenarios next.
Need to Manage Multiple API Configurations
Elective: CC Switch installation and Codex configuration guide.
Want to Learn Prospecting or Website Content
Laifaxin, AllinCMS, and SEO will be developed as separate advanced tracks, not forced into the required beginner curriculum.