Build a knowledge-based custom GPT in ChatGPT’s web GPT builder: open GPTs → Create, choose the conversational builder or configuration view, put behavior rules in Instructions, upload reference files under Knowledge, then test realistic prompts in Preview. The important limitation in 2026 is access: new GPT creation and publishing are not available on personal Free, Go, Plus, or Pro accounts, while Business, Enterprise, and Edu users depend on workspace permissions.
Check access before you design the GPT
GPT creation, editing, and publishing are web-based. Mobile ChatGPT apps can use GPTs but cannot create them. OpenAI’s current guidance says new creation and publishing are unavailable on personal ChatGPT accounts, including Free, Go, Plus, and Pro. Existing GPTs can still be used, and editing an existing GPT depends on your plan and permissions.
In Business, Enterprise, and Edu workspaces, an administrator may control whether members can create, edit, or publish GPTs. Check your workspace settings and announcements before building a workflow around a GPT. OpenAI’s current notice says retirement is planned for December 11, 2026 for affected Enterprise workspaces; a migration experience is targeted for September 17, 2026, but may not appear for every workspace at the same time.
Instructions versus Knowledge: put each kind of information in the right place
Instructions define how the GPT behaves. Use them for its role, tone, task sequence, safety boundaries, formatting requirements, and rules such as “ask one clarifying question before drafting.” Knowledge files are reference material the GPT can consult during a conversation: documentation, handbooks, guides, policies, product catalogs, or internal procedures.
#1 Best Overall
A practical split looks like this:
| GPT area | What belongs there | Example |
|---|---|---|
| Instructions | Behavior, workflow, tone, constraints, output format | “Answer from the uploaded policy first. If the policy is silent, say so.” |
| Knowledge | Facts and source material the GPT should retrieve | Employee handbook, API reference, troubleshooting guide |
| Conversation starters | Example requests users can click | “Summarize the refund rules for a customer” |
If the GPT should cite or quote files, say exactly how in Instructions—for example, “Name the document and section after each answer; if no passage supports the answer, state that the files do not establish it.” Uploading a file does not guarantee that every answer will be correct, so testing and source design matter.
Prepare a retrieval-friendly knowledge base
Use readable, text-forward files
Start with clean documents whose important content is selectable text. Headings, short sections, descriptive filenames, and consistent terminology make it easier to identify the relevant passage. If a PDF is mostly screenshots, scans, or complex slide layouts, convert important information into text or add text annotations. Do not rely on an image alone for a rule the GPT must apply.
Respect current file limits
OpenAI’s current documented limit is 20 files per GPT, with each file up to 512 MB. Supported document, spreadsheet, image, text, and code types can vary by model; some formats require the Code Interpreter & Data Analysis capability. Treat these as current product limits and verify them again when you publish a time-sensitive workflow.
Organize files for maintenance
- Use one authoritative version of each policy instead of uploading conflicting drafts.
- Put the effective date and owner in the document itself.
- Separate reference material by task when a single giant file would be difficult to navigate.
- Remove obsolete or duplicated files before testing.
Build the GPT step by step
- Open the builder. In ChatGPT on the web, open the GPTs area and select Create. If Create is missing, your account or workspace does not currently have permission.
- Choose a build route. The conversational builder drafts a GPT from a natural-language description. The configuration view lets you enter each setting directly. You can start conversationally and then refine the individual fields.
- Set the identity. Enter a clear name and description that tell users what the GPT does and who it is for. Add conversation starters that represent real requests rather than vague prompts such as “Ask me anything.”
- Write Instructions. Define the role, process, tone, allowed sources, uncertainty behavior, and output format. Include what to do when no uploaded passage answers the question.
- Upload Knowledge. Add your reference files in the Knowledge section. Confirm that each file uploaded successfully and that its contents are current.
- Choose capabilities only when needed. GPTs can combine instructions, knowledge, and selected capabilities. Add tools after the core behavior works; otherwise a tool failure can hide a retrieval or instruction problem.
- Test in Preview. Run representative questions, edge cases, and requests that should be refused or qualified. Check whether the answer uses the right document, follows the required format, and distinguishes supported facts from gaps.
- Refine, then publish or share. Improve Instructions and files based on Preview results. Workspace sharing and publishing options depend on your plan and administrator settings.
A robust Instructions template
Adapt this template to your domain rather than pasting it unchanged:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Role: “You are the support assistant for [team/product].”
- Source priority: “Use Knowledge files as the primary source. Do not invent a policy, price, date, or technical requirement.”
- Workflow: “Identify the user’s task, find the relevant section, answer in numbered steps, and list prerequisites.”
- Uncertainty: “If the files do not establish an answer, say that plainly and identify what information is missing.”
- Citations: “For factual claims, provide the filename and heading or page when available. Quote only short passages.”
- Safety: “Do not expose confidential content to an unauthorized requester. Ask for clarification when identity or scope is unclear.”
- Format: “Use headings for separate tasks, tables for comparable values, and code blocks for commands.”
Test retrieval instead of assuming it works
Create a small test set before inviting users. Include a direct lookup (“What is the cancellation period?”), a synthesis question requiring two files, a paraphrased question using different terminology, an unsupported question, and a deliberately conflicting or outdated term. For each result, check:
- Did the GPT select the authoritative file?
- Does the answer preserve qualifiers such as dates, regions, or exceptions?
- Does it cite or identify the source in the format you requested?
- Does it admit when the Knowledge base has no answer?
- Does it follow the requested tone and structure?
If a file is ignored, first confirm that the information is actually present. Try a narrower prompt that names the topic or document. Simplify complex PDFs or slide layouts, and convert image-only content into text or annotations. These steps improve diagnosability; they do not guarantee retrieval of every passage.
Connecting services: capabilities, apps, and actions
A GPT can use selected capabilities and can connect to outside services through apps or custom actions. Apps and custom actions are alternatives: a GPT cannot use both in the same configuration. Custom actions are an advanced path requiring API details, authentication information, and an OpenAPI schema. Use them when the GPT must call a service, not merely answer from uploaded material.
Keep the boundary clear: a custom GPT runs inside ChatGPT. It is not an embeddable assistant for your public website or application. If your product needs an assistant inside its own interface, use the API route instead.
Recommended Free Tools
Troubleshooting common failures
Create button or publishing option is missing
Cause: personal-account restrictions, a workspace policy, or insufficient permission. Fix: verify the account type, ask a workspace administrator, and check current announcements. Do not assume upgrading a personal plan restores creation while the restriction is in effect.
The GPT gives plausible but unsupported answers
Cause: instructions do not require evidence, files conflict, or the source is absent. Fix: add an explicit “do not invent” rule, identify the authoritative file, require source names or sections, and test unsupported questions.
A PDF is uploaded but key content is missed
Cause: scanned pages, image-based tables, or complicated slide layouts. Fix: provide a text-forward version, add text annotations, simplify the layout, and ask focused prompts that point to the relevant section.
The file will not upload
Cause: a file exceeds 512 MB, the 20-file limit has been reached, or the type is not supported for the selected model and capabilities. Fix: split or compress the source, remove duplicates, and check whether Code Interpreter & Data Analysis is required for that format.
An action fails even though the GPT answers normally
Cause: invalid authentication, an incomplete OpenAPI schema, or service-side errors. Fix: validate the schema and credentials independently, test a minimal request, and keep a non-action fallback response in Instructions.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Or skip the browser setup
If you need screenshots of documentation, test pages, or GPT-related web content rather than a custom GPT itself, ScreenshotNeo provides a one-request website screenshot API. It accepts cookie and consent banners as a visitor and removes more than 60 known consent platforms, newsletter popups, and chat widgets before capture; each cleanup step can be disabled. Bot checks or CAPTCHAs, blank pages, timeouts, failed loads, and cache hits are not billed, and response headers report the page verdict and billing status.
Use the API documentation at https://screenshotneo.com/docs/ for authentication and options. A basic call is:
curl -G "https://api.screenshotneo.com/v1/shot" -d access_key=YOUR_API_KEY --data-urlencode url=https://stripe.com -o shot.webpPython:
import requests
r = requests.get("https://api.screenshotneo.com/v1/shot", params={"access_key": "YOUR_API_KEY", "url": "https://stripe.com"}, timeout=90)
open("shot.webp", "wb").write(r.content)Node.js:
const q = new URLSearchParams({ access_key: 'YOUR_API_KEY', url: 'https://stripe.com' });
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);ScreenshotNeo also offers an MCP server with take_screenshot, get_page_info, and capture_pdf tools for Claude, Cursor, and other MCP clients. Plans include 1,000 screenshots per month free with no card; paid plans start at $5 for 3,000. Create a free ScreenshotNeo account.
Cost, reliability, and maintenance decisions
Knowledge limits are capacity limits, not accuracy guarantees. Keep a change log for source files, retest after replacing a policy, and remove superseded versions. For sensitive material, review who can access the GPT and whether sharing settings match the source permissions. Recheck account availability, file limits, and Enterprise notices before a long-lived rollout because these product conditions can change.
Frequently Asked Questions
Can I build a custom GPT from the ChatGPT mobile app?
No. Mobile apps can use GPTs, but creation and editing are web-based.
Can one GPT use both an app and a custom action?
No. A GPT can use apps or custom actions, not both at the same time.
Is a custom GPT suitable for embedding in my SaaS product?
No. GPTs run inside ChatGPT; an assistant embedded in your own product is an API use case.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
