Open the agent you used for Part 1 and ask whether it still has the OpenRouter connection. We used GrokBot for the working demo.

We first opened a different app and had to go back to the right place. Check this before asking the agent to build your skills. A connection in one app isn't automatically available in another.

Keep your API key out of the skill files and any repository you share. The bundle needs instructions for using the connection, not a copy of your credentials.

Ask your agent to build skills around the images you actually make. We picked newsletter ads and studio product photos, then asked for a separate skill to check OpenRouter for model candidates.

That model-checking skill had three groups to consider: cheap generations, affordable options with better quality, and premium generations. Keeping it separate means you can refresh the model choices without rebuilding every image workflow.

Here's an adaptable version of the request:

Help me build a custom image skill bundle using OpenRouter. Package it like a GitHub skill repository.

Make a skill that checks OpenRouter for cheap, affordable-quality, and premium image models. We'll run that weekly or monthly.

Then make brainstorming skills for [your common image jobs]. Generate different variations and tell me which model made each one. Include a step for choosing a direction before using a more expensive model for the final image.

We asked for that final-image step as part of the design. Our test here focused on the brainstorming skill, so start there before spending on a premium pass.

Pro tip: Package your skills together as if you were going to share them on GitHub. Give them useful names and keep the tests with them. It's much easier to reuse work you've saved properly than find it in an old chat.

Once the agent has created the skills, pick one to test. We started with newsletter ads rather than testing the whole bundle at once.

An evaluation, or eval, is a repeatable test. It gives you a way to compare changes to a skill instead of deciding whether the latest result feels better.

Ask the agent to test the choices that affect the images:

Write real evals in code for the newsletter ad brainstorming skill. Test the prompts, model selections, and number of variations. Run the evals and use the results to improve SKILL.md. Send me the generations and scoring results when they're done.

We wanted the agent to test its work and revise the skill from the results. It returned the generated images and scores for us to inspect.

For a useful comparison, also ask what the skill improves over a regular prompt. If it adds steps without improving the images or saving you time, keep working on it.

Open the generated images next to the scores. Check whether you agree before accepting the updated skill.

Our agent gave the models 10 out of 10. We gave one of those results 5.5. Some images were usable starting points, but others had poor layouts. The scores were far too generous.

Tell the agent what failed in the image and have it improve the test. A newsletter ad needs a layout you can use, so a test that gives a bad layout full marks needs changing too.

The results suggested four variations by default and six when we wanted more options. Treat that as a starting point for this skill, then compare it with how many images you actually find useful.

Pro tip: Ask for the images as well as the score report. We would have missed the weak results if we'd accepted the 10s without looking.