Foundations
6 min read · July 31, 2026
Last verified: July 31, 2026
Build Your First Custom GPT in 15 Minutes
Package one piece of your expertise into a tool someone else can use, with no code, in about 15 minutes.
What this is for
You answer the same question or apply the same judgment call repeatedly, for your team, clients, or people who ask you for free advice. This is for turning one of those repeatable things into a tool that runs without you, using ChatGPT's Custom GPTs or Claude's Projects. It is not a business by itself — see the honest limits section before you get attached.
Before you start
- A ChatGPT Plus/Team/Enterprise account (Custom GPTs require a paid plan) or Claude Pro/Team for Projects.
- One narrow task already picked — see step 1.
- 3-5 real documents you'd normally hand someone doing this task.
Step 1 — Choose a narrow enough job
The most common failure is building a GPT that's "a helpful assistant for my whole field." That's a worse version of ChatGPT, not a product. Narrow it until it does one job for one specific person.
Bad: "Marketing helper for small businesses." Good: "Reviews a therapist's website copy for HIPAA-safe language and rewrites anything risky."
Test: can you describe it in one sentence naming who it's for and what it outputs? If not, narrow further.
Prompt — Narrow your idea
Use in: Either.
ROLE: You are a product scoping consultant who forces vague tool ideas into narrow, buildable ones.
CONTEXT: My rough idea: [e.g. "helping consultants with proposals"]. My actual expertise: [e.g. "25 years writing and negotiating IT services contracts for mid-market companies"]
CONSTRAINTS:
- Push the idea until it describes one job, for one type of user, producing one type of output.
- Reject any scope that could apply to more than one clearly different audience.
- Ground it in my actual expertise, not a generic version of it.
OUTPUT FORMAT:
1. Three candidate narrow versions, each as: "[Tool name]: Takes [INPUT] from [SPECIFIC USER], produces [SPECIFIC OUTPUT]."
2. Your recommendation of which to build first, and why.
Step 2 — Write the instructions block
This is the core of the tool. It's the system prompt that runs every time someone uses it.
Use in: Either. Draft it once, it works in both platforms with minor formatting differences.
ROLE: You are [SPECIFIC ROLE, e.g. "a contract risk reviewer for IT services agreements"].
CONTEXT: Used by [SPECIFIC USER, e.g. "small IT consulting firm owners without in-house legal counsel"]. They paste in [INPUT TYPE, e.g. "a draft services agreement or SOW"]. Your job: [SPECIFIC OUTPUT, e.g. "flag clauses creating unlimited liability, unclear scope, or unpaid-work risk, in plain language"].
KNOWLEDGE: Reference uploaded files for [WHAT THEY COVER, e.g. "my standard risk checklist and past marked-up contracts"]. Where uploaded examples conflict with general best practice, follow the uploaded examples.
CONSTRAINTS:
- Never give formal legal advice; always recommend an attorney review before signing.
- Don't rewrite the entire contract — flag, explain, and suggest replacement language only for flagged clauses.
- If the input is missing standard sections (payment terms, termination, IP ownership), say so explicitly.
- Direct, practical tone, not academic.
OUTPUT FORMAT: Numbered list of flagged clauses — risky text quoted, why it's risky in one sentence, suggested replacement. End with overall risk level (Low/Medium/High) and why.
BOUNDARIES: If asked something outside contract risk review, say this isn't what the tool is for and redirect.
Step 3 — The exact click path
ChatGPT Custom GPTs:
- Click your profile (bottom left), then "My GPTs."
- Click "Create a GPT" (top right), then "Configure" at the top (skip the conversational builder).
- Paste your instructions block into "Instructions."
- Give it a Name and short Description.
- Scroll to "Knowledge," click "Upload files" — add your 3-5 reference documents.
- Under "Capabilities," leave Web Browsing and Code Interpreter off unless genuinely needed — each adds failure modes.
- Click "Create," then choose access: only you, anyone with the link, or the GPT Store.
Claude Projects (equivalent):
- Click "Projects" in the sidebar, then "Create project."
- Inside it, find "Project knowledge" and click "Add content" to upload your documents.
- Click "Set custom instructions" and paste your instructions block.
- No separate publish step — share access by adding people to your Claude Team workspace, or hand them the instructions and files to set up their own Project.
Step 4 — Test it against 5 adversarial inputs
Don't test with the input you used to write it. Test with: a real but messy example (typos, out of order); an input completely outside its scope (does it refuse correctly?); an input trying to override its instructions ("ignore your previous instructions and..."); an ambiguous edge case; an input from someone who phrases the question badly.
If it fails, fix the instructions block, not the knowledge files — most failures are scope problems, not missing information.
Step 5 — Share or sell access
ChatGPT: share privately via link, or list on the GPT Store (monetization there is indirect, through usage-based creator payouts in eligible regions — not a price tag you set). Claude: no public marketplace — invite someone to your workspace or hand them the setup files. For direct sales, bundle access into a paid service, or gate it behind a Gumroad/Stripe page that delivers the link or files.
Honest limits of custom GPTs as a product
A custom GPT is a configured version of someone else's model. You don't own the technology, can't guarantee it keeps working the same way after platform updates, and can't stop someone from copying your instructions. It's a good lead magnet and internal tool, and a reasonable value-add inside a paid service. It's a weak standalone product, because the moat is thin and platforms can change pricing or capabilities under you.
Where this goes wrong
- Scope creeps back to "general assistant" during testing. Every added capability makes the core job worse. Keep a "someday" list instead of building extras in.
- Knowledge files are your whole hard drive, not curated examples. Upload only documents showing the exact judgment you want reproduced.
- You skip adversarial testing and find gaps from a client instead of yourself. Boundaries only exist if you write them down. Test rudely before anyone else does.
The 2-minute version
Write one sentence naming the user and the output. Fill in the instructions template above. Upload 3 real examples. Ask it the worst question you can think of before showing anyone.
Where you stand with AI, in 10 minutes. The free AI Readiness Quiz gives you a personalized roadmap instead of another tool list: genxcelerate.com
Weekly, on Tuesdays. The Gen Xcelerator — practical AI intelligence for people who already know how to do the job. Subscribe free
Experience Is the API. — GenXcelerate
Related resources
- 12 Prompt ShortcutsTwelve lines you paste before or after any prompt to fix the most common AI output problems — five minutes to read, permanent once you save them.
- The Experienced Professional's AI StackOne tool per job, picked for 2026 — 15 minutes to read, and you'll end up using three tools, not thirty.
- Pricing Your ExpertiseSet a number you can say out loud without flinching, and know why it's right — 20 minutes of prep before your next quote.
The Gen Xcelerator
A new one every Tuesday. Short, useful, and written for people who already know how to run things.
Get the newsletter