Context-First AI Method · Stage 3 of 7
Configure: Give AI the Right Role, Voice and Boundaries
You have collected useful sources and organised them for reuse. AI still needs to know what job it is doing, which expectations apply and when to ask you.
Configure is stage three of the Context-First AI Method. It gives AI reusable instructions for working with those sources. Think of a short job description that a person can inspect and change.
What AI configuration should cover
AI configuration is a reusable set of working instructions. It describes the job, the intended reader, the sources to use, what a useful result looks like and when a person must decide.
Some instructions apply across many jobs: write clearly, distinguish facts from guesses and do not invent personal experience. Others belong to one task: use the approved business-email reference, keep the reply under 200 words and return a draft for review.
Keep the scope clear. A broad preference for casual language may suit a personal message. A client complaint may need a more formal response and a different review process. State the exception where it applies, and check that the chosen tool supports the setup.
Worked example: choose the right writing signature
Consider the constructed consultant example from Capture and Structure. The consultant has separate approved references for personal texts, business emails, longer articles and spoken video scripts. Each reference describes a writing signature: the patterns that help that type of communication sound right.
Capture selected the permitted examples. Structure made each reference findable and explained its purpose. Configure now tells AI how to use those references.
When drafting a client email, use the approved business-email reference. Use the personal-text reference for a personal message, the article reference for longer writing and the spoken-language reference for a video script. Ask before mixing styles.
A playful personal text may sound wrong in a client email. A detailed written explanation may be awkward when read aloud. One person can have several useful writing styles.
Now try the request, “Write a note to Jordan.” Without enough context, AI cannot reliably choose between friend, colleague and client. The instruction should tell it to ask which Jordan and what kind of note is needed.
Give AI a clear job and a checkable result
Choosing a voice is only part of the job. AI also needs to know what it may do. A useful role description for the consultant is:
You are a drafting assistant. Prepare options for human review. Do not speak for the person or commit them to an action.
For a business email, a useful result preserves the facts, answers the recipient’s question, sounds appropriate and makes the proposed next step clear. For a video script, it uses natural spoken language and a teaching sequence the listener can follow.
These standards give the reviewer something to inspect. “Act like the world’s best copywriter” leaves much more to guesswork.
Tell AI which source controls which decision
The right voice does not establish the right facts. A style example can show how to explain a price. It cannot tell you today’s price.
Tell AI which records provide current facts, which references guide style and which instructions govern the job. An old email may contain a useful opening and an outdated service promise. The opening does not make that promise current.
If sources disagree in a way that matters, the configuration should require AI to flag the conflict. It should not blend incompatible statements or treat the newest unapproved draft as the authority.
Define a fallback for a missing source. AI might return an outline with marked gaps or ask for the required input. New facts supplied for one task should not automatically overwrite the maintained reference library.
Set limits for reading, drafting and acting
A connector may let AI open a file or send a message. Being able to do that does not mean it is allowed for this task.
State what AI may read, what it may prepare and what needs human approval. The consultant’s assistant may draft a reply from approved references. Sending the reply or making a promise on the consultant’s behalf needs the appropriate authority.
In client or professional work, use systems approved for that information and task, with only the information needed. A paid AI account alone does not establish employer approval, client consent or permission to process confidential material.
Keep private sources separate from public-use material. An approved voice guide may support a public article without giving AI permission to quote private family messages or confidential client exchanges.
Written instructions are guidance, not a technical access control. Keep appropriate permissions, tool controls and human review in place.
Keep configuration, Skills and workflows distinct
Configuration states how AI should behave. A Skill explains how to perform a type of task. A workflow connects the stages, inputs, handoffs and finished result.
For the writing assistant, “draft only; ask before making a promise” is configuration. A repeatable method for preparing the email belongs in the Skill and workflow. A person still needs to check what happened.
More rules do not automatically improve the result. Start with expectations for one useful job. Add or change instructions when testing reveals a real gap, rather than building a global rule for every possible situation.
Test what AI does with the instructions
A statement that AI understands the rules is not evidence that it will follow them. Test the behaviour in an approved fresh session, with the sources and tools the real job would use.
- Try the normal job. Request a business email from a permitted scenario. Check the selected style reference, factual accuracy and draft status. Then request a personal text and inspect whether the style changes appropriately.
- Leave something important unclear. Try “Tell Morgan we’ll take care of it.” AI should clarify the recipient, the issue and whether the promise is authorised before drafting a promise.
- Cross a permission boundary. Ask it to use private messages in a public marketing story. It should identify the missing permission and offer to work from approved material.
Also check the fallback when the required reference is unavailable. Record the actual response. If AI guesses where it should ask, revise the relevant instruction and repeat that test.
Decide whether the configuration is ready to use
For this drafting job, the configuration is ready for a limited trial when a fresh session uses the right reference, preserves the supplied facts, returns a draft and asks at the intended boundary. The person reviewing the work should be able to find and change the instructions.
If success depends on hints buried in the original conversation, make those expectations explicit before handing the job to someone else.
For a team, specify the review status that must travel with the draft. Fluent writing can still leave a colleague to discover missing facts or unsupported claims. Configuration helps state the standard; supervision and checking establish whether the work meets it.
A small successful test does not establish that every future case is ready for automatic execution.
Write and test one short job instruction
Choose one job you already do. Write the intended audience, the correct source reference, the desired output, two things AI must not do and the point where a person decides what happens next.
Try a normal request, an unclear request and one that crosses a boundary. Compare the results with your instruction. Change the rule that failed, then test again.
How Configure relates to context engineering
Context engineering concerns the information available to an AI system while it works. Anthropic’s explanation of context engineering includes instructions, tools, external information and conversation history.
Configure focuses on the working instructions within that broader picture. Structure organises and routes the reference sources; Configure states how AI should behave with them.
Grove-White & Co.’s Context-First AI Method connects this preparation to direction, supervision, checking and improvement. Its seven stages form a cycle. A failed test may send you back to the source, the structure or the instructions.
Frequently asked questions
AI for professional and client work
How do I keep AI writing in my voice without changing the facts?
Use an approved style reference for expression and the appropriate current records for facts. Ask AI to preserve the reasoning, important details and uncertainty. Review the draft against both sources; a familiar tone does not establish factual accuracy.
What should AI do when it cannot tell who a message is for?
Ask for the missing recipient or relationship before choosing a style or using sensitive information. A first name alone may not distinguish a friend from a client. Make the clarification rule part of the job instructions.
Can I let AI send a client email because it can access my inbox?
Inbox access does not establish sending authority. Define reading, drafting and sending separately. Start consequential correspondence with a draft and appropriate human review; any later sending authority needs its own process, controls and evidence.
What instructions should a team share before using AI?
Agree on permitted sources, the expected result, the named reviewer and the point where AI must ask or stop. State how unresolved questions and review status travel with a handoff. Keep individual client facts separate from reusable team guidance.
AI for running a business
Should every instruction go into global personalisation?
Keep broad, durable preferences there when appropriate. Put client, channel and job rules in the narrower place where they apply. The exact controls vary by tool. Check that the intended instructions are available in the session you test.
How do I stop AI inventing prices or promises?
Point to the current record that controls those facts and state which commitments require approval. Tell AI to mark gaps or ask when the source is missing. Then test an unclear commitment. Instructions alone do not replace permissions or review.
Do I need a custom GPT, agent or new platform?
No. Start with a clear job brief and saved instructions in an approved tool you already use. Add technology when a real limitation justifies the setup, maintenance and review effort.
Search and AI visibility for businesses
Can my configuration control how outside AI services describe my business?
No. It can guide your own content process to use approved public service descriptions, evidence and claim limits. It cannot command an outside search or AI service to describe or recommend your business in a particular way.
AI for everyday life and learning
What is the smallest useful configuration for a personal task?
Choose one task. State who the result is for, which sources to use, what you want back and when AI should ask. Use only information you are permitted to process. Test the instruction before adding more rules.
How can I tell whether saved instructions work in a fresh chat?
Start an approved fresh session with the intended instructions and accessible sources. Try a normal task, an unclear request and a permission boundary. Inspect the actual output and source use. Asking AI to repeat the rules is not enough.
What comes next?
Configure gives AI working expectations. Direct turns those expectations, approved inputs and human decision points into a repeatable Skill and workflow.
Explore Direct in the seven-stage overview. To revisit the source setup, read the Structure guide or read the Capture guide.
For further practical guides and examples, subscribe to Grove-White & Co. email updates.