Try Sol and Luna. Check the work yourself.

The files behind the walkthrough: five creator exercises and a separate Sol/Luna comparison, with original prompts, fictional inputs, saved outputs and the notes that catch the mistakes.

Watch the walkthrough · Read the companion article

Download the Sol & Luna practice pack (ZIP) — free, no sign-up. Includes 42 exercise files and a START-HERE guide: prompts, fictional inputs, saved model outputs and review notes.

Use this alongside the five-test companion walkthrough and the Sol vs Luna video. The article explains what happened; these files let you repeat the exercises and check your own results.

Start with one small task

  1. Download and extract the ZIP, then open START-HERE.md.
  2. Start with Luna/05-structured-extraction/. Open SOURCE-RECORDS.md and TEST-PROMPT.md in a text editor.
  3. Start a fresh model conversation. Supply only the source records and task prompt. Keep the saved output and QA notes out of the conversation.
  4. Save your first answer before asking for corrections. Record the model, effort setting and date you used.
  5. Check that all 12 records are present, the five CSV columns match the prompt, and blanks stay blank. Then compare against the saved output and QA notes.

The original five exercises used GPT-6 Luna Medium in ChatGPT Work. You can try the same prompts with another available model, but label that as your own run. Model access or API usage may cost money; the practice files are free.

Pick an exercise

The folder names preserve the original project order. You can work through them in the order below.

Exercise Folder and files to give the model What to check yourself
Registration extraction Luna/05-structured-extraction/: SOURCE-RECORDS.md + TEST-PROMPT.md 12 rows, exact columns, preserved blanks, and the reasons for following up on R-105, R-106 and R-108.
Client inquiry triage Luna/01-client-intake/: INPUT-DATA.md + TEST-PROMPT.md 20 inquiries, decisions supported by exact quotes, and reply drafts that ask only for missing information. Read the QA notes about the corrected answer key and borderline case.
Audience feedback Luna/03-feedback-analysis/: SYNTHETIC-RESPONSES.md + TEST-PROMPT.md 22 responses. If using exclusive themes, each ID appears once and the counts total 22. Compare the initial output with RECONCILED-OUTPUT.md.
Transcript repurposing Luna/04-transcript-repurpose/: SOURCE-TRANSCRIPT.md + TEST-PROMPT.md Three Shorts drafts, a community post and a newsletter section. Check source fidelity and read scripts aloud to check duration.
Product catalog copy Luna/02-product-image-catalog/: PRODUCT-FACTS.md + TEXT-ONLY-PROMPT.md Four products, facts supported by the input, and alt-text status Needs image. This is a text-only exercise; no product photos are included.

Each exercise also includes saved outputs and QA notes. These are review materials, including mistakes and corrections, rather than perfect answers to copy into the model.

Try the later Sol/Luna comparison

This is a separate exercise with a more explicit service policy than the original client-intake test.

  1. Open Demo/inputs/SERVICE-GUIDE.md, CLASSIFICATION-PROMPT.md and shared-challenge.json.
  2. Give those same three files to Sol and Luna in separate fresh conversations. Keep Demo/qa/ and both models' saved outputs out of the inputs.
  3. Save each answer. Compare all five decisions, evidence quotes, missing facts and draft replies.
  4. Try routine-inquiries.json with Luna using the same policy and classification prompt.
  5. After your run, open Demo/qa/answer-key.json, REVIEW-NOTES.md and editorial-review.json. The original outputs and run notes are in Sol/06-workflow-designer/ and Luna/06-batch-intake/.

Both recorded agents used the same predefined policy at Medium effort in Codex. Luna did not consume Sol's generated rubric. Matching an expected category does not establish that the reply is ready to send. These files do not include measured latency, token usage, API bills or a real human policy approval.

Keep a simple run log

Copy this into a new text file for each exercise:

Exercise:
Model and effort:
Date:
Input files and any changes:
First output saved as:
Checks passed:
Missing records, unsupported claims or other errors:
Correction prompt:
Corrected output saved as:
What I would still review before using this workflow:

All businesses, inquiries, registrations and audience responses in the pack are fictional. The ZIP contains Markdown and JSON files; it does not include software to install, the full interactive demo, product images or video footage.

Download the practice pack · Read the full walkthrough · Watch the video

More free resources