Getting Started with ChatGPT Custom GPTs

When to use

Triggers:

• "What's a Custom GPT?" / "How do Custom GPTs work?"
• "Create a Custom GPT for [task]"
• "Set up a GPT for my team"
• "How do I share a GPT?"
• "Should I use Claude Skills or Custom GPTs?" / "Which is better?"
• Any AU SME operator using ChatGPT who wants to go beyond plain chat

Don't fire for:

• Claude-only questions (use getting-started-with-claude-skills)
• OpenAI API programming (developer-level, out of scope)
• ChatGPT plugin development (mostly deprecated by Actions in Custom GPTs)

The three layers ChatGPT gives you (April 2026)

ChatGPT has three ways to personalise / persist work:

1. Custom Instructions

Settings-level personalisation. Two free-text fields:

• "What would you like ChatGPT to know about you?" (your context)
• "How would you like ChatGPT to respond?" (your style preferences)

Applies to every conversation. Useful but blunt.

2. Projects

Scoped workspaces introduced in 2024. A Project contains:

• Its own set of instructions
• Attached files (docs, PDFs, spreadsheets)
• Its own conversation history

You start a new chat "inside" the Project; it inherits the instructions and files.

Projects are the closest ChatGPT analog to Claude Projects. They're not skills — they're contextual containers.

3. Custom GPTs (the closest ChatGPT analog to Claude Skills)

A Custom GPT is a packaged assistant:

• Name + description + avatar
• System prompt (the "instructions" field)
• Attached knowledge files
• Optional Actions (API calls to external services)
• Optional Code Interpreter / browsing / DALL-E access

Create one once, use it many times, share with your team or make it public.

Custom GPTs vs Claude Skills — the honest comparison

| | Custom GPT | Claude Skill | |---|---|---| | Discovery | GPT Store (public), ChatGPT Business workspace (private) | Local folder, shared GitHub repo, Claude.ai Project | | Format | Built via a GUI + system prompt | Plain Markdown file (SKILL.md) | | Reference files | Attached via UI, limited to GPT's knowledge files | references/*.md + scripts/*.py in the same folder | | Version control | Limited (GPT editor saves versions; no Git) | Native — put the folder in Git, done | | Composition | GPTs don't chain naturally; you call one at a time | Skills reference each other via ## Related skills | | Team sharing | ChatGPT Business workspace (works, paid feature) | Private Git repo, Claude.ai Project files | | Automatic triggering | You pick the GPT explicitly | Claude picks skills based on description triggers | | Multi-file project | Knowledge files (read-only context) | Full folder structure; scripts can be executed | | Best for | Self-contained assistant with a specific persona | Procedural knowledge composed across a library |

When Custom GPTs shine

• A customer-facing assistant with a specific persona (help desk, brand voice, specialist adviser)
• A team-internal assistant that wraps a large knowledge base (policies, product docs)
• Quick GUI-driven setup where editing Markdown feels heavy
• You want a GPT in the public store for discoverability

When Claude Skills shine

• Multiple related workflows that reference each other (a library, not a single assistant)
• Procedural knowledge that should be auto-triggered based on the user's task
• Version-controlled, reviewed, code-shaped artefacts
• Cross-tool reuse — the same SKILL.md works in Claude.ai, Claude Code, and the Claude API
• Composition — a skill that calls another skill that calls another

Honest take

Custom GPTs are great for packaged personas; Claude Skills are great for procedural knowledge. For an AU SME running a team: Projects + Skills + shared Git is a tighter loop than Custom GPTs for anything complex, but Custom GPTs remain the fastest way to get a specific assistant in front of customers or teammates via the ChatGPT UI.

Many operators run both. That's fine.

How to create your first Custom GPT

In ChatGPT (Plus / Business / Enterprise):

1. Menu → Explore GPTs → + Create
2. Configure tab (not Create tab):
   • Name: specific and searchable
   • Description: 1–2 sentences on what it does
   • Instructions: this is the core — 500–2000 words on:
     • What the GPT is (role, expertise)
     • Common tasks it handles
     • Tone and style rules
     • What NOT to do
     • Step-by-step for the main workflow
   • Conversation starters: 3–4 example prompts
   • Knowledge files: upload reference documents (policies, templates, past work)
   • Capabilities: enable Web Browsing / Code Interpreter / DALL-E / Actions as needed
   • Actions: optional — connect to an external API (this is where GPTs extend beyond chat)
3. Test in the right pane
4. Save — private, workspace-shared, or public

Total time: 30–45 minutes for a first working version.

The instructions field — the heart of a GPT

The 500–2000 word system prompt does most of the work. Structure:

# [GPT Name]

You are [role]. Your job is to [what it does].

## When someone starts a conversation
1. Ask [opening question]
2. Based on answer, [branch]

## How you respond
- Tone: [direct / warm / technical / …]
- Length: [brief / thorough / whatever]
- Format: [bullets / paragraphs / tables / …]

## Always do
- [thing 1]
- [thing 2]

## Never do
- [thing 1]
- [thing 2]

## When the user asks about [edge case]
[specific handling]

## If you don't know
Say so. Don't fabricate.

## Sign-off
[how every message ends — signature block if applicable]

Test it by pretending to be the user. Find the places it goes off-script. Tighten the instructions until it doesn't.

Common beginner mistakes

1. Treating it as a chatbot, not a skill. A good GPT is procedural — given input X, produces output Y. Chatbots ramble.
2. Over-relying on knowledge files. 50 uploaded PDFs doesn't make a good GPT. Focus the knowledge on 3–5 files the GPT will actually use.
3. No trigger phrasing in the description. Users find GPTs by searching; if the description doesn't include the obvious search terms, they won't find it.
4. Cute names. "Billy the Bookkeeper Bot" is less findable than "AU SME Bookkeeping Assistant".
5. Never revising. Ship v1, use it for 2 weeks, note where it fails, revise. A GPT that's never been revised is rarely good.

Sharing across your team

ChatGPT Business workspace

• Admin invites team members to the workspace (paid tier)
• Custom GPTs created in the workspace are shared with team members
• Optional: make specific GPTs private to their creator within the workspace

ChatGPT Enterprise

• SSO + admin controls + data residency
• Custom GPTs, Projects, and data isolation per organisation

Free / Plus

• Can create Custom GPTs but sharing is public (via a link) or private-to-you
• No team-scoped sharing

If your team is serious about GPTs, ChatGPT Business is the practical minimum. Check chatgpt.com/pricing for the current price per user.

Migrating Custom GPTs to Claude Skills (or vice versa)

Common request. The good news: the concept maps cleanly.

Custom GPT → Claude Skill

1. Copy the GPT's instructions into a new SKILL.md body
2. Add YAML frontmatter:

   ---
   name: [slug-form]
   description: [GPT description with trigger phrases]
   ---
   

3. Upload the GPT's knowledge files as references/*.md in the same folder
4. If the GPT has Actions, document them in the skill; you'll wire them up separately via MCP connectors or tool use in the Claude API
5. Test in a Claude.ai Project

Claude Skill → Custom GPT

1. Copy the SKILL.md body into the GPT's instructions field
2. Upload the skill's references/*.md as knowledge files
3. Write a 1–2 sentence description for the GPT Store / workspace listing
4. Add 3–4 conversation starters from the skill's examples
5. Test in the GPT builder's preview pane

Both directions preserve the core procedural knowledge. You'll lose GPT-specific features (Actions, Code Interpreter) or Claude-specific ones (skill composition) in the migration; plan around it.

Plan selection (checked September 2026)

Rough guide:

• Free: Can use Custom GPTs others built (public ones), can't create your own
• Go and Pro: two more individual plans. Check chatgpt.com/pricing for what each includes and costs.
• Plus: Create + use Custom GPTs, GPT Store access
• Business (formerly Team, from 2 users): Workspace-shared GPTs, admin console, increased limits
• Enterprise: SSO, data residency, audit logs, SAML, dedicated support

Prices change; check chatgpt.com/pricing for current AU pricing.

Measurement

Same discipline as with Claude Skills. Use ai-roi-measurement-for-sme to measure whether a Custom GPT is actually saving time / improving quality. GPTs, like skills, underperform when nobody measures them.

What this skill does NOT do

• Program OpenAI Actions. That requires an API schema and a real backend. Developer-level work.
• Replace Claude Skills for Claude users. Different platforms; use each for what it's good at.
• Advise on data residency / compliance. For AU compliance-sensitive clients, ChatGPT Enterprise or Claude Enterprise (US or AU-hosted via Bedrock) are the options. Get proper advice for each case.

Tier access

Free. Foundational content. Whether you use ChatGPT, Claude, or both, understanding both is free-tier value.

Related skills

• getting-started-with-claude-skills — sibling primer for Claude Skills
• thl-chatgpt-to-claude-migration — if you're moving off ChatGPT and need the full migration playbook
• ai-roi-measurement-for-sme — measure whether your GPTs / skills are paying off

References

• OpenAI — Creating a GPT
• OpenAI: ChatGPT Business overview
• OpenAI — Projects in ChatGPT