Context-First AI Method · Stage 1 of 7

Capture: How to Gather the Content AI Needs

By Grove-White & Co.

Suppose you want AI to draft a routine business reply in your voice. What would it need to know?

Start with the material behind that reply: a few permitted examples, current facts and the rules you follow when you write. The same question applies when AI helps prepare a quote, summarise a meeting or research a question. Useful sources may be in documents, messages, files or recordings, or in the head of the person who does the work.

Capture produces a source map: a simple record of the material, its location and its purpose. The map also records ownership, freshness, authority, permission to use it and gaps. In Structure, the next stage, you organise the selected material for reuse.

What counts as content?

In this method, content is source material. It can include emails, text messages, notes, reports, policies, templates, recorded explanations, interviews, spreadsheets, approved website copy and finished work samples. It can also include negative examples: work that sounded polished but was wrong, incomplete or unsuitable.

Content and context are related but not identical. Content is what exists. Context is the relevant part of that information for the particular person, job or decision. A hundred emails may be available content. The two approved examples that show how you explain a difficult decision to a client may be the context a specific task needs.

Collecting more material can introduce noise, stale claims, conflicting instructions and privacy risk.

How does Capture relate to context engineering?

Context engineering involves selecting and managing the information available to an AI system for a task. Anthropic’s explanation of context engineering includes instructions, examples, tools and other information the model uses.

Capture addresses the source-selection part of that work. Identify the facts, examples and human knowledge needed for one job, then check their relevance, authority, freshness and permitted use. Structure, the next stage, organises the selected material for reuse.

Grove-White & Co.’s Context-First AI Method connects that preparation with configuration, direction, supervision, checking and improvement. Context engineering helps explain why the information supplied to AI matters; the seven-stage method gives people a practical sequence for preparing, using and reviewing AI-assisted work.

Worked example: capture the raw material behind a writing signature

Imagine an experienced consultant who wants help writing across several channels without sounding generic. They use a short, friendly style with family, a more deliberate style in client emails, a clear educational voice in articles and a spoken rhythm when scripting videos.

Their original content could be spread across sent emails, personal texts, published articles, presentation notes, transcripts and voice recordings. A single folder called “my voice” would conceal useful differences. At Capture, the consultant instead makes an inventory of those sources and asks what each might prove.

For example, three permitted client emails might show how they acknowledge a concern, explain an option and close with a next action. Several personal texts might show brevity and warmth. A published video transcript might show the use of questions, pauses and everyday examples. A revised article might show which exaggerated words the author deliberately removed.

These are illustrative examples, not quotations from a real client's private correspondence. The full messages do not belong on a public website or in an unapproved AI service merely because they are convenient training material.

Start with the job, not the archive

A first project needs a narrower goal than “I want AI to know everything about me.” Choose a real result instead: “Prepare a reply to a routine business enquiry in my normal tone, for me to review.”

For that first reply, you might only need two or three representative, cleared business email examples, your typical response structure, current facts about the enquiry and a description of what a satisfactory reply includes.

You probably do not need personal text history, tax documents, an entire inbox or years of meeting transcripts. Your first source map should be small enough to inspect. Expand only when a test reveals a material gap.

For a recurring job, capture the rules the experienced person follows without noticing. What do they always check? Which phrase would they never use? What would make them stop and ask a question? Tacit knowledge often matters as much as written source files.

Four questions to ask about each source

Once you have a short list, use these four questions to decide which sources belong in your Capture source map.

First, what job could this source help with? A good document can be irrelevant to the current task. Relevance matters more than volume.

Second, who owns or controls it, and are you allowed to use it? Access to a mailbox or connected drive is not automatic permission to process confidential customer material. Identify personal, internal, client-specific and public content separately. Use minimum-necessary excerpts or safe invented examples when that is all the task requires.

Third, how trustworthy and current is the information? An old quote, unapproved draft or casual recollection may be useful evidence of a process, but it may not be the source of today's price, policy or promise. Record the source date and its level of authority.

Fourth, what might the example fail to show? A polished final email tells you what was sent, not which parts were corrected before sending. Where available and permitted, a before-and-after pair can reveal more about the author's standards than a final sample alone.

A practical Capture process

Now turn those source decisions into a record someone else can use. Build your Capture source map in five steps.

Define one outcome

Write a one-sentence job description with the audience, the intended result and the person who reviews it. For example: “Draft a polite, accurate first reply to a prospective client; do not send it.”

List likely sources

Note where relevant material lives, such as an approved template folder, a selected email set or publicly available writing samples. You can record locations without copying the full files.

Classify each source

Label it as approved for use, needs permission, restricted or unsuitable. Add the date, owner and a short explanation of its value. Separate facts and business rules from style examples. If a source contains both, it may need redaction or a more limited use.

Capture exceptions and counterexamples

Include one example where the usual approach would fail: a complaint requiring a careful response, a personal text that should never be used as business style, or a message with sensitive information. These boundaries help prevent the later system from treating every sample as interchangeable.

Identify what is missing

Mark unknown facts, unclear permissions and unwritten judgement rules. Get clarification or narrow the task; do not ask AI to guess.

What happens if you skip Capture?

AI may produce generic language because it has no representative examples. It may pick up outdated claims because no one identified the current source. It may imitate the wrong audience because examples were mixed together. Worse, a person may supply private or client information when a harmless example would have been enough.

Skipping Capture can also hide that a job is not ready for automation. If no one can explain which source is authoritative or what a good outcome looks like, a sophisticated workflow will not fix the underlying ambiguity.

Collecting every document before running one small test delays learning and increases the amount you must maintain.

Capture for different kinds of work

An insurance or wealth professional may begin with an approved, non-confidential example, a current firm guideline and a reviewer-defined checklist. Identify what may be used and what the licensed professional must still decide.

A small-business owner might capture one quote template, a parts list and a spoken explanation of how estimates are checked. Their source map should respect an existing process that already works.

A person learning AI for the first time may need nothing more than their own paragraph, a desired reader and a clear editing rule. A researcher or community volunteer may need cited public sources and a record of what remains unverified.

For a business improving public search visibility, the capture task changes: gather verified services, locations, qualifications, examples and fit limitations. Private operational material should not be mixed into information intended for public discovery.

How do you know Capture is done well enough?

Before moving to Structure, check whether a competent person could read your Capture source map and answer five questions:

• What job is AI helping with?
• Which sources matter, and where can they be found?
• Who owns the material, and is its use permitted?
• Which sources contain current authoritative facts, and which are illustrative samples?
• What important unknown or exception still needs attention?

Capture is ready for Structure when those answers make clear what AI should use, ignore or ask about.

A five-minute exercise

Choose one routine writing task. Record its audience and intended result. Find two examples you own and may safely use, one current factual source and one example of something you would reject. Write down one information gap. Do not upload anything yet.

Now ask: could another person understand why these sources were chosen? If not, improve the source map.

Frequently asked questions

Do I need to upload every email or transcript?

Capture does not require every email or transcript. A small permitted set with clear reasons for inclusion is a better starting point. Add material when testing reveals a need.

Can AI infer my style from a few samples?

AI may identify useful writing-style patterns from a few samples, but its interpretation can be wrong or overly general. A person must review the proposed patterns and decide which are worth retaining.

What if the best example is confidential?

A confidential example belongs in an environment approved for that information and use. Alternatively, create an appropriately de-identified or fictional example. Do not assume that a paid AI subscription makes client-data processing permissible.

What comes next?

With your source map in place, you can move to Structure. This second stage turns selected sources into Context Infrastructure: material you can find and reuse for the right job. Explore Structure in the seven-stage overview.

For further practical guides and examples, subscribe to Grove-White & Co. email updates.