How to build a custom AI assistant without code (2026 guide)
The short answer
To build a custom AI assistant without code, choose one frequent, low-risk job with written rules, design it on one page, write layered instructions with two examples, add approved and dated knowledge files, test 20 to 30 realistic cases against a release gate, try to break it, share it narrowly, and keep a portable SKILL.md copy outside any platform.
Want the tools, not just the method? The The Build Your Own AI Assistants Kit kit turns this guide into ready-to-use templates, spreadsheets and printables, with an instant download.
- Choose the job before the tool: frequent, rule-based, low-risk, no sensitive data.
- Steps and two examples beat long lists of prohibitions.
- Small, dated knowledge files are easier to find and cite than one long PDF.
- Test every version against the same set and a release gate.
- Custom GPTs and Gems are being retired in favour of skills: build new assistants as skills and keep a SKILL.md copy you own.
What are the steps to build a custom AI assistant?
A custom assistant is a saved set of instructions, knowledge files and settings that makes a general AI tool do one job the same way every time. Whatever your platform calls it, the build follows the same nine steps.
- 1. Choose a job: frequent, rule-based, low risk if wrong, no sensitive data.
- 2. Design it on one page: purpose, users, inputs, knowledge, instructions, output, guardrails, escalation.
- 3. Write instructions in layers with one typical and one hard example.
- 4. Prepare knowledge files: approved, current, one topic per file, dated on line one.
- 5. Add a name, a when-to-use description, conversation starters and a disclosure line.
- 6. Build on your organisation's approved tool and keep a portable SKILL.md copy.
- 7. Test 20 to 30 realistic cases and release only through a gate.
- 8. Red-team it, set sharing narrowly, then roll out with a short briefing.
- 9. Measure value weekly, version every change and retire on purpose.
Which jobs should you build an assistant for?
Score each candidate job from 1 to 5 on frequency, time per use, how clearly the rules are written, risk if the answer is wrong, data sensitivity and whether approved knowledge exists. Weight frequency, risk and sensitivity most. A score of 1 on risk or sensitivity should stop the build whatever the total.
The research explains why the choice matters. In a field experiment with 758 consultants, AI help on tasks inside its capabilities meant more tasks done, faster and at higher quality, but on a task chosen to be outside those capabilities, people using AI were 19 percentage points less likely to reach a correct solution. Jobs with written rules and easy-to-spot mistakes sit well inside the frontier.
Keep assistants out of decisions about people. AI systems for recruitment and selection are listed as high-risk in the EU AI Act, and Colorado's 2026 law adds notice and human-review duties when automated technology materially influences employment or health-care decisions. Drafting, answering, checking and tutoring are the safe territory.
How do you write instructions an assistant will follow?
Use seven layers: role and purpose, users and context, numbered steps, sources, output format, three or four firm boundaries with an escalation contact, and two examples. Keep it to about 800 words. If you need more, the assistant is doing two jobs or reference material belongs in knowledge files.
Examples carry a lot of weight. Research on in-context learning found that examples work largely by showing the format and the kind of input and output expected, so one perfect example of your answer format does more than a paragraph describing it. It also means the assistant will copy any mistake in your example.
Do not judge instructions by reading them. A study of prompt formatting found that small formatting changes moved accuracy by up to 76 points on the same task for some open models. Run the same test set before and after every change.
How should you prepare knowledge files?
Split long documents into topic files of a few pages, keep the source's section numbers, and put the organisation, title, version and effective date on line one so answers can cite them. Use plain text or simple Markdown and remove personal data.
Two findings justify the effort. Language models use the start and end of long inputs better than the middle. And retrieval reduces invented answers without removing them: retrieval-based legal research tools from two major vendors still produced incorrect or misgrounded answers between 17% and 33% of the time in a pre-registered evaluation. NIST calls this risk confabulation. Small, labelled files and retrieval tests reduce it; your test set catches what remains.
How do you test a custom AI assistant?
Write the test set before you build: 20 to 30 cases taken from real questions, about two-thirds normal, the rest edge cases and attempts to break it. Score each answer 0, 1 or 2 on accuracy, completeness, format, guardrails and tone. A case passes at 8 out of 10 with no zero on accuracy or guardrails; any zero there is critical. Release at 90% or more with zero critical failures.
You can use another model to draft scores. Strong model judges have matched human preferences over 80% of the time, but they show biases towards position, length and their own outputs, so a person should score a sample and compare.
| Version | What changed | Passed (of 30) | Critical | Gate |
|---|---|---|---|---|
| v0.1 | First draft from the canvas | 18 | 3 | NO-GO |
| v0.2 | Handbook split into 11 dated section files | 23 | 2 | NO-GO |
| v0.3 | Escalation rule for medical, leave and pay questions | 26 | 1 | NO-GO |
| v1.0 | Two worked examples added | 28 | 0 | GO |
| v1.1 | Handbook section 7 (PTO) updated to v3.3 and swapped in | 28 | 0 | GO |
| v1.2 | Rebuilt as a workspace plugin skill ahead of the 11 December 2026 GPT retirement | 29 | 0 | GO |
A worked example: is a policy Q&A assistant worth it?
Take an HR team receiving 46 handbook questions a week at about 6 minutes each: 4.6 hours a week. Building and testing the assistant takes 24 hours at a loaded cost of $48 an hour, so the build costs $1,152. Upkeep (sample reviews, file updates, retests) takes an hour or more a week. All figures are estimates; put yours in their place.
The low case matters most: if the assistant handles only 8 questions a week and upkeep takes 1.5 hours, it never pays back. In the expected case it pays back in about 20 weeks. Show the low case first when you ask for sign-off.
| Case | Questions handled per week | Minutes each | Upkeep hours per week | Net value per week | Break-even |
|---|---|---|---|---|---|
| Low | 8 | 6 | 1.5 | -$34 | Never |
| Expected | 22 | 6 | 1 | $58 | 20 weeks |
| High | 30 | 7 | 1 | $120 | 9.6 weeks |
Reading helps; measuring tells you what to work on. These AI-graded assessments on AssessAll pair with this topic:
How do you protect an assistant from prompt injection and data leaks?
Prompt injection is text that tries to override the instructions, typed by a user or hidden in a file or web page the assistant reads. Researchers have shown that instructions planted in content likely to be retrieved can hijack AI applications, and the UK's National Cyber Security Centre warns that prompt injection may never be fully mitigated because these models do not separate data from instructions.
So design for it: give the assistant no authority to act in other systems, no personal data it does not need, and a rule to trust only approved files and flag conflicts. Then run a 45-minute red-team session covering instruction extraction, impersonation, hidden instructions in uploads, pasted personal data and requests for rules that do not exist. The OWASP Top 10 for LLM Applications 2025 is a good checklist. Share only with the intended audience, and remember that on some platforms everyone with access can see a shared assistant's instructions and files.
How do you roll an assistant out and measure it?
Before anyone outside the build team uses it, run a go/no-go check: a signed one-page design, approved knowledge files, a passed release gate, a red-team session with every failure fixed or accepted in writing, tested escalation contacts, sharing limited to the intended audience, and a disclosure line telling users they are talking to an AI assistant that can be wrong. In the EU, the AI Act requires systems that interact with people to be designed so people know they are dealing with AI, unless that is obvious.
Then brief the people who will use it. Forty-five minutes is enough: why it exists, a live demo with one normal question, one escalation and one out-of-scope request, what never to paste in, paired practice, a short knowledge check and how to flag a wrong answer. Under the EU AI Act as amended in 2026, organisations that deploy AI systems take measures to support their staff's AI literacy; a recorded briefing is one practical measure.
Finally, measure value, not just use. Track five numbers every Friday: people who used it, questions still handled by hand, escalations by design, answers flagged as wrong, and a person's review of ten answers. In the worked example above, weekly users reached 109 by week 12 while handbook questions reaching the HR inbox fell from 46 to 19 a week, and 77 of 80 reviewed answers were correct. Part of that fall came from a new intranet page, and an honest report says so.
What is changing on ChatGPT, Claude, Gemini and Copilot?
OpenAI has scheduled the standard retirement of custom GPTs for 11 December 2026 (28 February 2027 for FedRAMP workspaces and approved deferrals; dates may vary by workspace) and is moving them to plugins, where a GPT's instructions become a skill; custom actions and the selected model do not carry over, and conversation starters may not copy. Google began moving Gems on personal accounts to skills in November 2026, with Workspace business accounts following in March 2027 and education in June 2027. Claude offers projects and custom skills built on a SKILL.md file, and Microsoft Copilot accepts SKILL.md uploads as skills (with Copilot Premium) alongside agents built in Agent Builder. For a new assistant, build a skill, not a GPT or a Gem.
The common thread is the open Agent Skills format: a folder with a SKILL.md file whose frontmatter has a lowercase-hyphen name and a description of what it does and when to use it, followed by the instructions, with optional reference files. Keep your assistant in that form in your own storage, with its knowledge files and test set, and a platform change becomes a retest rather than a rebuild. Check each vendor's help pages for current menus and limits.
What are the most common mistakes?
- Building for the loudest request instead of the best-scoring job.
- Opening the builder before writing the one-page design.
- Writing fourteen 'never' rules instead of steps and examples.
- Uploading one long PDF and old versions.
- Testing with three friendly questions.
- Skipping the red-team session because it is internal.
- Sharing before the release gate, or more widely than the audience.
- Editing a live assistant without a version and a retest.
- Measuring logins instead of questions answered correctly.
- Keeping the only copy inside one platform.

Build your first assistant on tests you trust
The Build Your Own AI Assistants Kit gives you the design canvas, instruction template, test sets, red-team prompts, data boundary policy and 12 SKILL.md blueprints used in this guide.
Sources
- OpenAI Help Center: GPTs in ChatGPT (retirement dates)
- Google Gemini Apps Help: About the transition from Gems to skills
- Agent Skills: Specification
- Harvard Business School: Navigating the Jagged Technological Frontier (Working Paper 24-013)
- Microsoft Support: Create and edit skills with Copilot
- arXiv: Sclar et al., Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design
- ACL Anthology: Liu et al., Lost in the Middle
- arXiv: Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools
- arXiv: Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
- NIST AI 600-1: Generative AI Profile
- National Cyber Security Centre: Prompt injection is not SQL injection
- OWASP Gen AI Security Project: Top 10 for LLM Applications 2025
More from the Bodhih family
Questions people ask next
Can I still create a custom GPT?
New GPT creation is not available on personal ChatGPT accounts and has ended for affected Enterprise workspaces; existing GPTs are scheduled to retire on 11 December 2026 for most accounts. Check your account notice and OpenAI's help pages, and build new assistants as plugin skills (or projects for small-team work) where your plan allows.
What is a SKILL.md file?
The core of an Agent Skills folder: YAML frontmatter with a name and a description of what the skill does and when to use it, then instructions in Markdown. Claude, Gemini and Microsoft Copilot document SKILL.md uploads, and ChatGPT plugins hold skills, which makes it a practical portable copy of your assistant.
How many test cases do I need?
Start with 20 to 30 realistic cases, including edge cases and attempts to break it. Run the same set against every version so you can compare results.
Should a custom assistant use personal data?
Only if your organisation's approved tool is approved for that data and the job needs it. Most assistants can be designed so they never ask for personal data. Regulators in Australia and Canada advise against putting personal information into tools that are not managed or approved for it.
Can an AI score my assistant's test results?
It can draft scores, but model judges show position, length and self-preference biases. Have a person score a sample and compare before you trust automated scores.
How do I measure whether an assistant is worth it?
Track weekly users, questions still handled by hand, escalations, flagged answers and a small human review of answers. Compare with a baseline taken before launch, and present low, expected and high cases honestly. Jobulary's guide to if-then plans can help make that Friday review a habit.
Is there a way to check my own prompt-writing skill?
AssessAll offers a task specification and prompt writing work sample and an AI draft supervision assessment. Bodhih's Build Your Own AI Assistants Kit includes the design canvas, test sets and 12 blueprints used in this guide.