Custom GPTs: agents in your voice.
Custom GPTs let you package a system prompt, knowledge files, and tool actions into a reusable agent. Done well, they save hours per week. Done badly, they're just ChatGPT with a costume. This lesson is the build pattern that produces GPTs you actually use.
Workflow 01 The 4-section system prompt
1
Write a system prompt that actually changes behavior
Most Custom GPTs fail because their system prompts are too short and too vague. The 4-section pattern below produces dramatically better results.
The prompt that works
System prompt structureSystem prompt for a Custom GPT called 'Email Drafter':
## Identity
You are an assistant that drafts emails for [user's role]. Your one job is to produce email drafts the user can send with minor edits.
## Voice
Write in this voice:
- Short sentences. Direct.
- No 'I hope this finds you well' or filler.
- One specific call to action per email.
- End with a concrete next step, not a question.
- Length: 3-5 sentences unless asked for more.
[Paste 2-3 example emails of yours here for reference]
## Process
When the user gives you a recipient and a goal:
1. Ask any clarifying questions if the goal is ambiguous (max 1 question).
2. Draft three variations: concise, friendly, direct.
3. After the user picks one, refine if requested.
## What you don't do
- You don't send emails. You only draft them.
- You don't claim to know the recipient unless told.
- You don't insert emoji or AI-clichés like 'circling back.'
Best use cases
- Email or message drafting in your voice
- Code reviewer that uses your team's style
- Customer support drafter for your product
- Sales follow-up generator for your industry
Test with examples that should fail (off-topic prompts, requests outside the GPT's scope). A good system prompt has the GPT politely decline or redirect — not improvise.
Time savings: A focused GPT: 5 min/use savings × 20 uses/week = 100 min/week per GPT.
Workflow 02 Knowledge files vs. system prompt
2
When to upload files vs. when to paste content
Custom GPTs let you attach knowledge files (up to 20). They're searchable but the GPT doesn't 'memorize' them. Knowing the difference saves you from misusing the feature.
The prompt that works
When to use whatKnowledge files are best for:
• Reference data the GPT searches when relevant (product catalogs, policies, specs)
• Long documents (50+ pages) that wouldn't fit in a system prompt
• Data that updates periodically — you swap the file, prompt stays the same
System prompt is best for:
• Behavior and voice (always loaded)
• Process and rules (always applied)
• Examples of good output (always referenced)
Don't put behavior rules in a knowledge file — they won't be consistently applied. Don't put 200-page product manuals in the system prompt — they get truncated.
Best use cases
- Product-specific assistants with catalogs as knowledge
- Support GPTs that reference up-to-date docs
- Legal GPTs with case-law references
- Internal GPTs that need company-specific data
Knowledge files are not 'memory.' The GPT searches them on each query — it doesn't reliably recall facts you'd expect it to from a file uploaded weeks ago. Verify behavior with test prompts.
Time savings: Right placement: avoids hours of 'why isn't it following the rule' debugging.
Workflow 03 Actions — wire up real APIs
3
When you need the GPT to do, not just say
Actions let your Custom GPT call external APIs. This is where they go from 'smart text generator' to 'agent that does something.'
The prompt that works
Actions setupAdd an Action to your GPT:
1. Configure the API endpoint (OpenAPI spec required)
2. Set up authentication (OAuth, API key, or none)
3. Describe what the action does in plain English (the GPT uses this to decide when to call it)
4. Test extensively with edge cases
Examples of useful actions:
• Lookup company info via Crunchbase API
• Fetch current stock price via Polygon API
• Create a Linear ticket from the GPT
• Query a read-only database via your custom API endpoint
• Send a Slack message to a specific channel
For any action that creates, modifies, or sends — require user confirmation in the prompt. 'Show the user what you'd do, then wait for them to say 'go.'
Best use cases
- Lookup-only actions (CRM lookups, stock prices, weather)
- Create-with-confirmation actions (file tickets, draft slack)
- Query-only DB actions (read your data without modification)
- Integration with your team's internal tools
Never wire up an action with write/delete permissions to production without user confirmation in the prompt. GPTs occasionally call actions you didn't expect.
Time savings: Action-enabled GPT: replaces an entire tool for specific workflows.
Build a Custom GPT you'll actually use weekly
Pick one repetitive task you do 5+ times a week. Use the 4-section pattern to write the system prompt. Add 2-3 knowledge files if needed. Test it. Iterate. By end of week, you should be using it daily without thinking.
What you can do now
- Use the 4-section system prompt structure (Identity, Voice, Process, Don'ts)
- Choose knowledge files for reference data, system prompt for behavior
- Add actions only when the GPT needs to do, not just say
- Always require confirmation for actions that create or modify
- Test edge cases — off-topic prompts, ambiguous requests, hostile users