GPT 6 examples: six workflows you can inspect
Inspect six worked examples across Astra, Sol and Luna, from code repairs and structured data to documents and a playable game.
6 items in this library.
GPT 6 document formatting with preservation checks
Treat formatting as a constrained transformation and inspect both the visual result and the underlying document structure.
GPT 6 document reconciliation with visible evidence
Compare documents through aligned fields and an evidence ledger so a reviewer can verify every important change.
GPT 6 frontend QA: turn a brief into evidence
Use a model to draft a test matrix, then use the real browser and target viewports to establish what actually works.
GPT 6 game prototyping: build the first playable loop
Use a constrained loop, explicit state, and a short review path before adding content or polish.
GPT-6 Luna support triage example
An inspectable eight-request batch showing strict JSON output, unknown-field handling, and a check against command-like customer text.
GPT-6 Sol pagination review with a runnable fix
A complete synthetic code review shows how to connect pagination defects to a small patch and repeatable tests.
No items match these filters. Try another category or clear the filters.
More results →Choose GPT 6 examples by the task you can check
These GPT 6 examples include four Astra exercises recorded on September 7 and two new Sol/Luna cases recorded on September 24. They use synthetic documents, a local browser fixture, and a small game. Each case pairs the task with inspectable artifacts and a check record. They are independent exercises, not reproductions of OpenAI customer benchmarks.
| Your task | Example and input | Artifact to inspect | Decisive check |
|---|---|---|---|
| Apply a document style without changing content | Document formatting: a brief, table, image, links and footnote | Finished DOCX/PDF, original and rendered pages | Wording and objects survive; exported fonts match |
| Explain changes between versions | Document reconciliation: two fictional workshop records | Located differences and unchanged controls | A sign change, currency change and missing qualification stay distinct |
| Reproduce an interface failure | Frontend QA: prompt search and clear controls | Before/after browser evidence and a small patch | Clearing an empty search restores the cards |
| Make one playable loop | Game prototyping: catch stars, score and restart | Browser prototype, source and state checks | Terminal states and restart obey the rules |
For the new models, inspect Sol’s pagination repair and Luna’s eight-request support triage. Their fixtures, actual outputs and executable checks are separate from the original Astra records.
What the local work reveals
The formatting task produced a two-page document from a one-page source. Its first PDF substituted Liberation Sans for Arial even though the editable file specified Arial. A local font configuration corrected the export; the first version remains available. This is an export issue that a document preview or a text-only answer could miss.
The reconciliation response identified eight differences and retained two unchanged control fields. It did not add unlike currencies together or interpret an omitted approval condition as proof of approval. The interesting result is the located comparison, which a reader can check against both inputs.
The frontend exercise first checked the current local prompt library. Because the focused path passed, the failure demonstration uses a clearly labeled training fixture with an intentionally broken Clear search handler. The same browser checks record zero restored cards before the fix and 15 afterward. That defect is not attributed to the production site.
The game case exposes a different boundary: a state test, an ordinary play session and a browser screenshot answer different questions. Inspect its attached check record before treating a visible score or victory screen as evidence for the full loop.
Official examples are separate context
OpenAI’s Playco account reports three themed prototypes from one grey-box foundation and 50% fewer manual fixes; one version still required a performance fix. OpenAI’s Legora account reports financial-statement review results. Those are attributed customer reports, not measurements of the exercises here. The sources were rechecked on September 7, 2026.
To adapt a case, preserve a small input, define the expected result, run the task once, and keep the first failure beside any correction. The practical guide explains how to carry that record into another task. Choose the example whose output you can independently inspect, rather than the one with the most impressive reported improvement.