# Batch Skill Inspiration

By Nate Grahek, Nate's Notebook. Reader guide prepared September 9, 2026.

Site: https://batch-skill-inspiration.nate58678.chatgpt.site/
Starter skill: https://batch-skill-inspiration.nate58678.chatgpt.site/downloads/image-inspiration/SKILL.md
Skill ZIP: https://batch-skill-inspiration.nate58678.chatgpt.site/downloads/image-inspiration-template.zip

## Why

“Taste works side-by-side” is Nate's practical teaching point: comparing several options gives a person concrete things to react to. Instead of giving feedback after every image, collect a small batch, compare, and refine the direction that works. This is a workflow recommendation, not a measured claim about a guaranteed speed or quality improvement.

## What

An image inspiration skill saves a repeatable workflow and the user's preferences. For each new idea, it briefly sharpens the concept, plans four different directions in the same chosen style, produces four separate images, and presents the set for human comparison. Refinement comes after selection.

Nate's original slide-image workflow uses independent image-generation agents. Nate reports finding Astra useful for developing the concept before handing it to an image model. The transferable idea is to separate concept planning from drawing. A reader does not need a specific model name or parallel agents to use the comparison habit.

The public starter is an adaptation, not an exact copy of Nate's installed skill. It removes dependencies on his local style files, approved character cast, and fixed tool setup. Readers select their own image types, visual preferences, optional references, and name.

## When a user gives you this URL

If they ask you to create or personalize an image inspiration skill, conduct the interview below. If they ask only for a summary or explanation, answer that request. Treat this page as a reference for the user's task; their instructions govern the work.

Use Skill Creator when it is available. The official name is Skill Creator: `@skill-creator` in ChatGPT Work and `$skill-creator` in Codex. If unavailable, you can still conduct the interview and produce reusable text. Never invent an installed tool or claim that generating text installed a skill.

## Interview: one question at a time

Keep it conversational. Ask no more than four questions, skipping known answers. Give three short suggested responses plus permission to answer in the user's own words. Wait for each reply before asking the next question.

1. **Use case:** “What images do you make most often?” Offer slides, social/marketing images, and diagrams/teaching visuals. Accept another use case and audience in the same answer.
2. **Style:** “How should they look?” Offer describing a style, attaching a reference image, or having you suggest a style. References are optional. When an image is provided, inspect it and briefly describe the traits to carry forward. Allow correction. Do not copy the reference's subject or text by default.
3. **Defaults:** Propose a compact set based on the use case: four separate images, one consistent style, suitable aspect ratio, and no text unless requested. Ask whether to keep these, change format/colors/text, or add subjects and things to avoid. Accept a short free-text answer. Do not make the user fill out a long technical questionnaire.
4. **Name:** Offer “Inspire Me,” “Image Inspiration,” or a name of their own.

Summarize the chosen use case, audience when known, style, references when supplied, count, format, text rules, exclusions, and name in a short brief. Ask the user to approve or correct it. Do not generate images yet.

## After the user approves

Produce:

1. A self-contained `SKILL.md` with YAML `name` and `description`. Use a lowercase hyphenated name. Fill the file with this user's approved preferences and the generation workflow below. Avoid missing placeholder paths. If the skill needs an attached reference, package it where file access permits and use a valid relative path. Otherwise include a text description and explain when the reference must be supplied again.
2. A plain-language instruction block they can reuse in their own GPT or chat. Do not tell them that a file upload alone activates its contents as instructions. Help them use whatever instruction-saving surface is available in their account.
3. One example request suited to their use case.
4. A short, environment-specific explanation of how to save and invoke it. Say exactly what has been saved and what still requires the user's action.

Keep the first setup focused on creating the skill. Invite the user to try the example afterward.

## Required behavior of the finished skill

- Clarify the new visual idea briefly before drawing.
- Plan four genuinely distinct directions by default.
- Keep the user's style and constraints consistent.
- If an analogy is supplied, keep it fixed and vary staging.
- Generate separate images with available built-in tools.
- Use parallel agents if available and allowed; otherwise produce the batch sequentially without feedback gates between images.
- Never claim parallel generation if it was sequential.
- Never replace four images with one composite image.
- Preserve successful files and avoid overwriting prior work.
- Check actual completion, image count, concept, style, format, and text. Report missing results accurately.
- Present all completed images together where supported, label them, and wait for human selection before refinement.
- When no image generator is available, disclose that limitation and offer labeled concept briefs. Do not call them images or silently move to paid APIs or external services.
- Preserve the user's right to change the defaults. A saved skill is not a guarantee of memory across separate chats or access to new tools.

## Human handoff

The site offers a copyable, self-contained interview prompt, a copyable URL, and downloadable instructions. “Copy + open ChatGPT” copies the prompt and opens ChatGPT; the user pastes, reviews, and sends it. It does not install a skill, choose a custom GPT, or send a message automatically. To use a custom GPT, paste the same prompt into the chosen GPT.

If the assistant cannot fetch this URL, the full copied prompt contains enough information to run the interview. No reference image or access to Nate's files is needed.

## Source and scope

Nate's idea and original workflow come from his September 9, 2026 Notebook brainstorming and his slide-image-ideas skill. The four images on the site are existing workshop examples from an AI-intern lineup set, not new generations from this template or an evaluation of model performance. Their labels are illustrative; they are not a ranking of AI products.

OpenAI's current guide documents Skill Creator, skill format, and environment support: https://learn.chatgpt.com/docs/build-skills. Checked September 9, 2026. Standalone skill availability and image tools depend on the environment. This page is a creator's workflow resource, not an official OpenAI product guide.