GPT-6 Astra · Creative workflows
GPT-6 Astra Skills: Save a Repeatable Workflow in Codex
A skill is useful when the same sequence of work keeps recurring. For a creative team, that might be turning product facts and references into an image brief, an edit instruction and a review checklist. Save the stable process once; keep each campaign's changing facts in its own input.
Skills belong to the agent environment, not exclusively to Astra. OpenAI documents them as reusable instructions, resources and optional scripts for ChatGPT and Codex. They can help an Astra workflow stay consistent, but they do not grant model access or add an unavailable image-generation service.
Separate a reusable process from a one-off prompt
A one-off request contains both method and facts: ‘prepare a launch brief for this lamp using these references.’ A reusable skill keeps the method and asks for the facts each time. That prevents last month's product color, audience or price from leaking into the next campaign.
Choose a narrow job with a recognizable input and output. ‘Do marketing’ is difficult to trigger and evaluate. ‘Prepare a visual brief from supplied product facts and reference images’ is specific enough to use consistently.
Start with an instruction-only SKILL.md
OpenAI requires a SKILL.md file with name and description metadata. A skill may also contain resources and scripts, but the visual-brief example does not need executable code. Its value comes from a clear workflow and an observable output contract.
The description should explain when the skill applies. The instructions should say what to read, what to produce and how to handle missing facts. Do not claim the agent has generated an image when it has only written a prompt.
---
name: campaign-visual-brief
description: Prepare an image or video brief from supplied product facts and visual references. Use for campaign asset handoffs, not for publishing or purchasing.
---
Read the supplied product facts and references. Separate product identity, art direction and layout references. Ask about missing facts that materially change the brief; label other assumptions. Produce a short audience statement, asset table, one image prompt, one controlled edit prompt and an acceptance checklist. Preserve exact supplied claims. Do not invent prices, testimonials or test results. Identify any requested generation or editing step that remains unexecuted.Place and invoke it in the right environment
For repository-scoped Codex use, the documentation describes .agents/skills directories. Put this example in .agents/skills/campaign-visual-brief/SKILL.md in the intended repository. That lets the workflow travel with the project if the team chooses to check it in.
In Codex CLI or the IDE extension, the documentation describes /skills or a dollar-sign mention for explicit invocation. Interfaces differ, so use the picker your product exposes. Save the downloaded text as SKILL.md in its own directory; downloading it alone does not install or activate the skill.
$campaign-visual-brief
Prepare a launch image brief for the fictional Still lamp. Audience: people improving a small home-office desk. Use the attached image for product identity. Deliver a website hero and one square social variation. Keep the headline editable outside the image. Mark product specifications as unknown unless supplied.Evaluate the output with a fresh campaign
Test the skill on a second product with a different shape and audience. Check that it uses the new facts, assigns reference roles and produces a brief specific enough to execute. If it repeats the lamp's details, the stable instructions contain information that belongs in the task input.
Keep a small change log for the skill. When a revision fixes a recurring omission, rerun one representative brief. You do not need to build a large automation system before proving that a short instruction file saves useful work.
| Check | Passing result |
|---|---|
| Trigger | The description fits the requested task |
| Input isolation | Only the current product's facts appear |
| Missing facts | Unknown specifications stay visibly unresolved |
| Output contract | Brief, prompts and checklist are all present |
| Execution honesty | Planned steps are distinct from completed actions |
Read community skills as code and instructions
Public skill repositories can offer useful ideas about structure and wording. Inspect the files before adopting them, especially if they include scripts or network access. A repository title or model-specific branding is not proof that its workflow is effective for your team.
The repository below is an example of community packaging. We have not installed or audited it. The starter in this guide is independently written and instruction-only; it does not incorporate that repository's implementation.
Keep the workflow close to the actual handoff
Once the brief is approved, send its product reference, layout constraints and acceptance checklist to the person or tool producing the assets. Save the selected outputs alongside the brief so future revisions have a stable starting point.
Ciyo can be the visual production step after that handoff. The skill itself does not integrate Ciyo or invoke its tools; it prepares an input you can carry into the product. Add automation only after you know the manual handoff works.
Put your visual brief to work
Use your saved visual-brief workflow to prepare consistent inputs, then open Ciyo to work on the campaign's images or videos.