Custom GPTs¶

Last verified: 2026-05-06 · Drift risk: medium Official sources: Creating and editing GPTs, Projects in ChatGPT

What this surface is¶

A Custom GPT is a published configuration of ChatGPT that packages a specific set of instructions, knowledge files, capability settings, and optional external API connections (Actions) into a shareable assistant. You build it through the GPT builder at chatgpt.com/gpts, and you control who can access it — yourself only, anyone with a link, or anyone on the GPT Store.

Custom GPTs are distinct from Projects. A GPT is a fixed, published tool: one instruction set, one set of knowledge files, deployed to whoever you share it with. A Project is a living workspace where context accumulates over time. If you are building something that will be used repeatedly by people who should not need to configure anything themselves, a Custom GPT is the right choice. If you are building a personal or team context hub that evolves, use a Project instead. The official FAQ describes GPTs as "best for reusable knowledge that spans functions and contexts" and Projects as "best for team alignment, faster onboarding, and knowledge capture."

Who it is best for¶

Developers or teams that want to publish a specialized assistant to internal or external users without writing any code.
Organizations that need a consistent interface for domain-specific tasks (triage, summarization, formatting, Q&A over documentation) that many people will use.
Builders who want to connect a ChatGPT interface to an existing HTTPS API without standing up a full agent system.

Prerequisites¶

A paid ChatGPT account (Plus, Pro, Business, Enterprise, or Edu). Building and editing GPTs is not available on the Free plan at the time of writing.
The GPT builder is only available on the web at chatgpt.com. The mobile apps allow using GPTs but not building or editing them.
If you plan to add Actions: an HTTPS endpoint you control, with a valid TLS certificate, that accepts JSON requests. Do not use http:// — Actions require HTTPS.

Step-by-step setup¶

Go to chatgpt.com/gpts and click Create.
The builder opens with two tabs: Create (conversational wizard) and Configure (direct field editing). Start with Configure for precise control.
Fill in Name — this appears in search results and at the top of the chat. Be specific: "Cardiovascular Literature Triage" rather than "Med GPT."
Fill in Description — one or two sentences explaining purpose, audience, and typical tasks.
Write Conversation starters — three to five realistic example prompts that demonstrate how to use the GPT. These appear to users when they open it.
Write Instructions — see the section below.
Upload Knowledge files if needed.
Enable or disable Capabilities (web search, image generation, code interpreter).
If connecting to an API, set up Actions — see below.
Click Create to save and publish. To update an existing GPT, click Update after making changes. Changes save to a draft automatically while you edit.
Use the built-in Preview panel to test the GPT before sharing it.

Builder UI walkthrough¶

Configure tab¶

Name, Description, Conversation starters

These are user-facing fields. The Name appears everywhere the GPT is referenced. The Description surfaces in GPT Store listings and previews. Conversation starters show on the opening screen and should reflect realistic, high-value tasks.

Instructions

Instructions define how the GPT behaves: what it should do, how it formats responses, what it should decline to do, and in what tone. Per the official docs: "For multi-step workflows, use explicit step structure, for example: 'When X happens → do Y,' and separate sections with clear delimiters." Positive directives ("Return a JSON object with keys...") work better than long prohibition lists. Include brief examples of acceptable and unacceptable outputs when the GPT must apply specific classifications.

Knowledge

You can attach up to 20 files, each up to 512 MB, per the docs. Use knowledge files for reference material the GPT should draw from (documentation, handbooks, lookup tables). Put behavioral rules in Instructions, not in files. After uploading, test in Preview to confirm the GPT actually uses the content as expected.

Capabilities

Four built-in capabilities you can toggle: Web Search, Image Generation, Canvas (longer structured content editing), and Code Interpreter & Data Analysis. Enable only what the GPT actually needs — unused capabilities increase surface area for unexpected behavior.

Actions

Actions let the GPT call external APIs you define via an OpenAPI schema. Use Actions when the GPT needs to retrieve live data or trigger operations in external systems. A GPT can use either the connected Apps system or Actions, but not both at the same time. See the Actions setup section below.

Worked example: a literature triage GPT¶

Goal: A GPT that accepts a structured abstract (PMID-style format, but no real PHI) and returns a JSON triage decision.

This example uses entirely synthetic text and does not reference real trials, real patient data, or real medical recommendations.

Instructions:

You are a research triage assistant for a systematic review team.

Input format: You will receive one or more abstracts in the following structure:
  TITLE: <title>
  AUTHORS: <authors>
  SOURCE: <journal, year>
  ABSTRACT: <full abstract text>

Your task for each abstract:
1. Assess relevance to the review topic: "interventions for type 2 diabetes in adults aged 65+."
   - INCLUDE: RCTs and systematic reviews with human participants aged ≥65 with T2D.
   - EXCLUDE: Animal studies, non-English, case reports, editorials, studies with mean age < 60.
2. Return a JSON object with these keys:
   - "pmid_placeholder": the value provided in the SOURCE field if a DOI or ID is present, else null
   - "decision": "INCLUDE", "EXCLUDE", or "UNCERTAIN"
   - "exclusion_reason": brief string if EXCLUDE, else null
   - "population_flag": true if age of population is explicitly stated, else false
   - "study_design": one of "RCT", "Systematic review", "Observational", "Other", "Not stated"
3. Do not add commentary outside the JSON. Return one JSON object per abstract, wrapped in a JSON array if multiple abstracts are provided.
4. Do not make clinical recommendations. Do not infer data not present in the text.

If an abstract is ambiguous, set "decision" to "UNCERTAIN" and explain in "exclusion_reason".

Test input (synthetic):

TITLE: Effects of SynthDrug-A on HbA1c in Older Adults with Type 2 Diabetes
AUTHORS: Smith J, Doe A, Lee B
SOURCE: Synthetic Journal of Endocrinology, 2025. ID: SYNTH-2025-001
ABSTRACT: Background: Glycemic management in adults aged ≥65 with T2D presents challenges.
Methods: RCT, n=280, mean age 71.2. Intervention: SynthDrug-A 5mg vs placebo.
Duration: 52 weeks. Primary outcome: HbA1c change from baseline.
Results: Mean HbA1c reduction 0.8% (95% CI 0.5–1.1), p<0.001.
Adverse events: hypoglycemia 4% vs 2%. No severe events.
Conclusion: SynthDrug-A reduced HbA1c with an acceptable safety profile.
Limitations: Single center. No COI disclosed.

Expected output:

[
  {
    "pmid_placeholder": "SYNTH-2025-001",
    "decision": "INCLUDE",
    "exclusion_reason": null,
    "population_flag": true,
    "study_design": "RCT"
  }
]

Test this in the Preview panel before publishing. Verify that the GPT returns valid JSON, handles edge cases (missing age, case report), and declines to add clinical commentary.

Actions: connecting to an external API¶

Actions are defined using an OpenAPI 3.x schema that you paste into the GPT builder's Actions editor. A minimal schema structure:

openapi: "3.1.0"
info:
  title: My Tool API
  version: "1.0"
servers:
  - url: https://api.yourdomain.com
paths:
  /search:
    get:
      operationId: searchRecords
      summary: Search the internal record database
      parameters:
        - name: query
          in: query
          required: true
          schema:
            type: string
      responses:
        "200":
          description: Search results

Key requirements:

The servers.url must use https://. Plain http:// is not accepted.
Do not include authentication tokens or API keys directly in the schema text. Use the Authentication section of the Actions builder to configure bearer tokens, API keys, or OAuth separately. Never paste a real credential into the schema field.
The GPT calls these endpoints on behalf of the user when the conversation triggers the relevant operation. The user may see a confirmation prompt before the action runs, depending on the action type.

When to use a Custom GPT vs a Project¶

Criterion	Custom GPT	Project
Will multiple people use a fixed tool?	Yes — use a GPT	No — use a Project
Does context accumulate over time?	No	Yes — use a Project
Do you need live API integration?	Yes — Actions	Via connected Apps (limited)
Does the team need to see each other's work?	No	Yes — use a Project
Is this a one-off personal workspace?	No	Yes — use a Project

Limits and gotchas¶

Knowledge file limits: 20 files, 512 MB each, per the docs. Complex layouts in PDFs reduce retrieval reliability.
Actions require HTTPS. There is no exception for local or staging endpoints. Use a tunnel tool (e.g., ngrok) for development testing, but replace with a real HTTPS endpoint before sharing the GPT.
Actions and Apps are mutually exclusive. A GPT can use one or the other, not both simultaneously.
Version history is available from the three-dot menu in the editor. If you restore an older version that had Actions configured, you may need to reconfigure authentication afterward, per the docs.
Mobile users cannot build GPTs. They can use them.
JSON output is not guaranteed to be valid JSON without careful instruction engineering and testing. Use the Preview panel thoroughly. Consider adding explicit instructions to "return only valid JSON with no surrounding text."
Model selection: You can set a Recommended Model for your GPT, but users with access to other models can switch. If your JSON schema relies on a specific model's behavior, document that.
GPT Store visibility and discoverability change with product updates. Verify current sharing settings in the builder before publishing to a wide audience.

Confirmed by docs vs. practical inference¶

Claim	Source
20 files, 512 MB each for Knowledge	Confirmed — official docs
Actions require HTTPS	Confirmed — official docs
Actions and Apps are mutually exclusive	Confirmed — official docs
Restoring old Action versions may break auth	Confirmed — official docs
Builder only on web, not mobile	Confirmed — official docs
JSON output reliability with instruction engineering	Practical inference — valid JSON output requires explicit prompting and testing; no documented guarantee
Triage accuracy on synthetic abstracts	Practical inference — accuracy depends on model version and instruction clarity; not documented

Cost and rate-limit notes¶

Building GPTs is included in paid plan subscriptions at no separate charge. When users interact with your GPT, their conversations consume from their own plan's message quota. If your GPT uses Code Interpreter, image generation, or Actions with external APIs, those tool calls may use additional credits or count toward tool-use quotas — the exact accounting follows standard ChatGPT pricing for the capability in question. Publish to a limited audience first to estimate usage before a broad release.

Where to go next in this guide¶

If you need multi-step automation, tool calling, or want to run agents in code, see OpenAI API and Agents SDK.
For a persistent team context hub rather than a reusable tool, see ChatGPT Projects.
For a terminal-based coding agent, see Codex CLI.