First Codex Task: Organize Product Materials
- Previous: Write a clear AI task brief
- Goal: use a few copies to complete the workflow: inventory → sample → confirm → organize → deliver.
- Estimated time: 10–20 minutes.
💡 Why this exercise? Organizing product materials is relevant to international trade, but it does not send outreach, change a business system, or spend money. That makes it suitable for a first real task.
1. Prepare Test Materials
In the Codex-Test workspace created in the previous lesson, prepare:
Codex-Test/
├── 01-originals/
│ ├── A100-specs.txt
│ ├── A100-front.jpg
│ ├── A100-side.jpg
│ ├── B200-spec.md
│ └── unknown-model.jpg
├── 02-to-process/
│ ├── A100-specs.txt
│ ├── A100-front.jpg
│ └── B200-spec.md
├── 03-output/
└── 04-needs-review/
Use copies. Keep backup/reference copies in 01-originals for human comparison only, then copy at most three test files into 02-to-process. AI may read and process only 02-to-process in this task; it must not read or process 01-originals. If you do not have suitable real material, create a few fictional files containing example text.
The purpose of the first exercise is to learn the workflow, not to maximize volume. Do not include an irreplaceable original, account credentials, contracts, or a complete customer database.
2. Ask Codex to Inventory First
Copy this prompt:
First inventory the files in `02-to-process`.
Do not read or modify `01-originals`. For now, do not modify, move, rename, overwrite, or delete any file.
Report:
1. filename;
2. file type;
3. likely product model;
4. evidence for that conclusion;
5. anything that cannot be determined.
Separate direct facts, inferences based on filenames, and items that need human confirmation.
Check whether the file count and model interpretation are correct. Correct the inventory before moving on if you find an error.
This Windows reference test occurred before the folder rules were unified. It read one fictional source file from 01-originals. The screenshot demonstrates only the inventory method below; it does not verify the current rule that 02-to-process is the only AI input. That revised boundary still needs a separate retest.
The AI did not invent extra files to match an example count. It separated the available information into:
- Facts: content directly supported by the file;
- Inferences: conclusions based on a filename or surrounding context;
- Needs review: items without enough evidence and therefore requiring a person.
Reporting the number of files that actually exist is more important than producing a list that merely looks complete.
3. Ask for an Organization Plan Before Action
Based on the inventory I have confirmed, design an organization plan.
Requirements:
- create folders by product model;
- leave source files unchanged;
- place organized results in `03-output`;
- list files with an unconfirmed model under `04-needs-review`;
- include a README.md in each product folder describing the included material;
- create `product-material-index.md` in the output root.
Show the proposed directory structure first. Do not execute the plan yet.
You should see a structure similar to:
03-output/
├── A100/
│ ├── README.md
│ └── material copies
├── B200/
│ ├── README.md
│ └── material copies
└── product-material-index.md
Review the folder structure and rules first. A detailed plan is not evidence that any file has already been created.
4. Approve at Most Three Samples
Do not approve the full batch immediately after the plan looks reasonable.
The plan is approved for a sample. Process at most 3 files whose product model can be confirmed.
If fewer than 3 files can be confirmed, process the actual confirmable number. Do not invent files or expand the scope just to reach 3.
Requirements:
- copy only; do not move source files;
- do not overwrite existing files;
- do not process files whose model cannot be confirmed;
- stop after the sample;
- list every file created or copied in this run;
- if an output already exists, skip it and report the skip instead of overwriting it.
Only one source file could be confirmed in this reference test, so the exercise used a controlled one-file adaptation and processed one file. “At most three” is a ceiling, not a quota. If an output with the same name already exists, the correct behavior is to skip it and report the conflict, not overwrite the previous result.
These screenshots came from an already-installed Windows 11 environment using CC Switch with the third-party Laifa AI configuration. They demonstrate only this narrow method: specified one-file inventory → plan without execution → controlled one-file sample → no overwrite of an existing output. They do not verify the revised 02-to-process-only boundary, OpenAI's official ChatGPT sign-in, a full batch, real business data, or complete Codex compatibility. No complete before/after manifest was generated, so this evidence remains a partial reference run, not a full pass.
5. Review the Sample Yourself
Open the output folder and check:
- whether the product model is correct;
- whether images are in the correct folder;
- whether every source file still exists;
- whether each README contains only supported information;
- whether a filename-based inference was presented as an official specification;
- whether the AI exceeded the sample size you approved;
- whether anything outside the task scope was changed.
If you find a problem, respond precisely:
The sample needs these corrections:
1. Copy images without renaming them.
2. In README files, list filenames only; do not infer product uses.
3. Put files with an unknown model in the needs-review list.
Restate the revised rules first, then regenerate only the sample.
6. Process the Remaining Material Only After Approval
Send the following only after the sample passes:
The sample is approved.
Process the remaining files using the same rules, while continuing to follow these boundaries:
- do not delete source files;
- do not move source files;
- do not overwrite existing files;
- put uncertain items in needs review;
- do not send, upload, or publish anything externally.
When finished, update `product-material-index.md` and create `processing-report.md`.
The report must list:
1. files read in this task;
2. files copied;
3. files created;
4. items that still need confirmation;
5. failures or skipped items.
7. The Transferable Skill Is Not File Organization
The method behind this exercise applies to many other jobs:
Prepare copies
→ inventory first
→ propose a plan
→ create a small sample
→ review it yourself
→ expand the scope only after approval
→ produce a delivery record
Use the same sequence later for Word documents, PDFs, spreadsheets, website content, or other AI tools.
8. Completion Check
- Only test material or recoverable copies were used.
-
01-originalsremained for human comparison only, and AI used only02-to-processas input. - Inventory came before any batch organization.
- The proposed directory structure was reviewed first.
- At most three confirmable files were processed as samples; if fewer existed, the actual number was used without inventing files.
- Remaining files were processed only after human review.
- Source files were not deleted, moved, or overwritten.
- Existing outputs were skipped and reported instead of overwritten.
- Uncertain content went to needs review.
- The final index and processing report were delivered.
9. Next Step
Codex may have finished executing, but “execution finished” does not mean “the result is correct.” The final lesson covers formal review and a safe closeout.