Why Custom Instructions Are Not Your AI Business System

Reviewed and updated 7 August 2026.

In January 2026, I called my first strategic AI setup my “Business Bestie”. It was a useful starting idea: give an AI clear instructions so it challenges vague thinking, uses Australian English and focuses on business outcomes.

What I would not tell a business owner now is that one custom-instruction script creates a trained AI assistant in five minutes.

Custom instructions can shape behaviour. They are not a business knowledge system, a workflow, a permission model or proof that the output is correct.

What custom instructions are good for

OpenAI describes ChatGPT custom instructions as information you want ChatGPT to consider in its responses. Claude provides profile instructions, project instructions and styles for different kinds of personalisation.

These features are useful for stable preferences such as:

• using Australian and UK English;
• keeping answers concise and scannable;
• asking for evidence when a claim is uncertain;
• challenging vague ideas with clear reasoning;
• following a preferred response structure; and
• respecting named approval boundaries.

That is useful behaviour design. It is only one layer.

What one instruction cannot do

OpenAI also states that custom GPTs do not use saved memory, custom instructions or previous conversations, and that each conversation starts fresh. Product features and boundaries differ, which is another reason not to treat one portable script as a complete system across every AI platform.

From Business Bestie to AI Brain for Business

Business Bestie was an early name for the behaviour I wanted from AI: direct, useful, commercially aware and willing to challenge me.

The idea evolved because behaviour without business context still produces confident guesswork.

I created the AI Brain for Business to combine the information, rules and workflows an AI needs to support the way a business actually works.

Its foundation includes:

• Business DNA: strategy, audience, offers, customer knowledge, goals and the Revenue Roadmap;
• Brand DNA: positioning, messaging, voice and communication rules;
• Content DNA when needed: channels, formats, publishing practice and performance context;
• project instructions: the purpose, constraints and requirements for a specific body of work;
• skills and workflows: repeatable steps for completing named tasks;
• permissions and approval boundaries: what the system may read, draft, change or send; and
• testing and maintenance: real examples, quality checks, change records and regular source updates.

For the client, the outcome is more relevant work, less repeated briefing and a clearer way for a person to review how an answer was produced.

The right place for instructions in Claude

Claude’s official guidance separates personalisation into different layers:

• Profile instructions for account-wide preferences;
• Project instructions for context and requirements that apply only within a particular project;
• Styles for how Claude formats and delivers a response; and
• Skills for repeatable behaviours or specialised tasks.

Claude Projects can also include a project knowledge base containing relevant documents, text, code or other files. Claude uses those sources to understand the background for chats inside that project.

The practical lesson is simple: put each instruction at the narrowest useful level. A company-wide rule should not be buried in one chat. A client-specific fact should not be placed in an account-wide profile. Sensitive information should not be copied into a system before the business has approved the tool and the use.

A better starter setup

1. Set only stable account preferences. Language, response shape and broad working style belong here.
2. Create one project for one purpose. Name the outcome, audience, scope and boundaries.
3. Add current source files. Include only permissioned information needed for the work.
4. Write a task workflow. Define the trigger, inputs, steps, quality check, human approval and destination.
5. Test real examples. Include good cases, bad cases and edge cases.
6. Record what changes. Update source information and instructions when the business changes.

A simple instruction pattern

A useful project instruction should be specific enough to test. For example:

Purpose: Help prepare first drafts of Australian small-business content from the approved source files in this project.

Rules: Use Australian English. Do not invent facts, testimonials, prices or links. Distinguish verified facts from recommendations. Flag conflicts between sources.

Quality check: Before finishing, confirm the audience, outcome, evidence, current offer and next action.

Approval boundary: Draft only. Do not publish, send, upload or change external systems.

That instruction still needs current business information, a defined workflow and human review. The strength comes from the system around it, not from pretending the prompt is magic.

Official product references

• OpenAI Help Centre: ChatGPT custom instructions
• OpenAI Help Centre: GPTs in ChatGPT
• Claude Help Centre: Claude personalisation features
• Claude Help Centre: Claude Projects

Your practical next step

If you want to move beyond generic chat and set Claude up around real business work, start with my Claude setup and training guide for Australian businesses.

If your business knowledge is scattered across your head, files and team, explore the AI Brain for Business. If you want help diagnosing the right starting point first, book a 1:1 AI Business Power Hour.

About Deb Szabo

Deb Szabo brings 30 years in marketing to her work as an AI Business Strategist and Claude Specialist. She helps Australian founders and small teams turn business knowledge into governed AI context, practical workflows and useful day-to-day capability.

Deb passed Anthropic’s official proctored Claude Certified Associate - Foundations exam on 6 August 2026. She is registered in the Claude Partner Network and is based in Pokolbin in the Hunter Valley, NSW.