My Journey With AI: How to Start, and How to Build the Right Data Repository for Better Results

If you ask ten people how to get better results from AI, nine of them will talk about prompts. Prompts matter, and I'll cover them. But the biggest jump in quality I have seen, in my own work and in the teams I advise, does not come from cleverer prompts. It comes from giving the AI better material to work with: a well-organized AI data repository that holds your voice, your offers, your audience, your proof and your rules.

I'm Luis A. Laguna Torres, a marketing and creative strategist and the founder of Lagoon of Randomness and CTRL+R Studio. AI has become part of my daily workflow across writing, research, image and video ideation, custom code and automation. This guide is two things at once: an honest look at how I approach AI, and a practical, template-heavy walkthrough of how to start and how to build the data repository that makes assistants like ChatGPT and Claude genuinely useful for your business.

A quick note on how I write about this. I keep the personal parts grounded in what I actually publish and do, and I frame everything else as a method you can adapt. AI tools change monthly. The principles in this guide are the part that lasts.

My journey with AI: from novelty to workflow

I did not start with a grand plan. Like most people, I started by trying things, and the way my use changed is a useful map for anyone beginning now. I describe it here as stages rather than dates, because your timeline will be different from mine.

Stage 1: Text as a thinking partner

The first useful thing AI did for me was help me think. Drafting outlines, pressure-testing an argument, summarizing a long document, generating ten headline options so I could argue with them. My résumé lists ChatGPT, Claude and Gemini side by side, because each is a little different in tone, strengths and limits, and I get better results when I match the tool to the job.

Stage 2: Images and creative direction

The next stage was visual: using AI image tools to explore looks, product renders and social layouts. The AI World page on this site shows real prompts and the images they produced, including 3D product renders, a character package and Instagram post and Story formats. The lesson I took from that page is one I repeat to clients: what you see is the working version, not the first draft. Each prompt went through several rounds of refinement, and some outputs still needed a Photoshop pass, the part AI cannot do for you. I also note there that prompt-based generation cannot reliably guarantee exact logo reproduction, so for packaging and brand work the best workflow is to generate the art first and place the real logo in a design editor.

Stage 3: Video and voice

Then came motion. Tools such as Luma AI, HeyGen and RunwayML, combined with a traditional editor like Adobe Premiere, made it possible to prototype animated characters, talking-head videos and product visuals far faster than a traditional production schedule. My studio's tool stack includes these, and the workflow that works is hybrid: AI for ideation and raw generation, humans for pacing, story, sound and finishing.

Stage 4: Code and interactive experiences

AI became a way to scaffold code: small interactive widgets, custom HTML sections and landing pages that I then test, debug and refine. The AI World page includes a self-contained interactive HTML example you can copy and test. AI can scaffold in seconds, but the real work is the troubleshooting afterward, which is why "prompting smarter, building better" is the framing I use.

Stage 5: Systems and automation

The latest stage is systems thinking: connecting AI to workflows through CRM and automation platforms such as GoHighLevel, HubSpot and Zapier, designing lead qualification logic, content pipelines and reporting. This is where AI stops being a novelty and starts saving real hours each week, and it is also where the quality of your underlying data decides everything. If your assistant gives a lead the wrong answer about your pricing, the problem is not the model. It is the source it was given.

What I learned

1. AI is a force multiplier, not a strategy. My essay on this, AI won't save a weak brand, it will expose one faster, makes the case that AI belongs in the middle of the process: humans define the question at the start and make the final call at the end.
2. Judgment stays in the loop. The most reliable rule I know is that a person reviews anything that goes out under your name.
3. Context beats cleverness. A mediocre prompt with excellent context outperforms an excellent prompt with none.
4. Process beats tools. Tools change every quarter. A documented process survives.

How to start with AI (without getting overwhelmed)

If you are new, follow this order.

Step 1: Pick one job, not "AI"

Do not start by "using AI for marketing." Start with one specific, repeated, low-risk task, such as drafting first versions of weekly social captions, summarizing meeting notes, writing product descriptions from a spec sheet or answering internal FAQs. A narrow job gives you something you can measure.

Step 2: Choose one or two assistants and learn them

Pick one general assistant and stick with it for a few weeks. ChatGPT and Claude are both strong general-purpose tools, and they differ in style, features, limits and plan details that change often. Try the same task in each, compare, and choose based on the quality of the results for your work. Avoid entering sensitive or regulated information into any tool until you understand its data-handling terms.

Step 3: Learn a prompt structure that works everywhere

Anthropic's and OpenAI's own prompting guides converge on similar principles. In short: be clear and specific about the task, supply context, show examples of what good looks like, specify the format you want, and break big jobs into steps. A structure I recommend is:

Role: who the assistant should act as
Context: who this is for and what they need
Task: exactly what to produce
Constraints: length, tone, things to avoid, facts that must be included
Format: how the output should be structured
Examples: one or two samples of the style you want

Anthropic's guidance for Claude adds a few tips worth adopting. Structure prompts with clear tags around each type of content (for example instructions, context and input), give the model a role, include a handful of well-chosen examples, and, when working with long documents, place the documents at the top of the prompt and your question at the end. For long-document tasks it also suggests asking the model to quote the relevant passages first, so the answer stays grounded in your material.

Step 4: Iterate out loud

Treat the first output as a draft. Tell the assistant what is wrong ("too formal," "you invented a statistic," "make it about the customer, not us") and let it revise. Save the prompts that work in a document. That document is the seed of your repository.

Step 5: Add a review step

Decide in advance who checks the output, and what they check: facts, tone, claims, links and anything legal or regulated. AI can be fluent and wrong at the same time.

Why results depend on your data, not just your prompts

Assistants are trained on broad public information. They do not know your prices, your service area, your tone, your policies, your product specs, your past campaigns or what makes you different. Without that material, they fill gaps with plausible-sounding generalities. That is why the same tool can produce brilliant work for one person and bland, off-brand or inaccurate output for another.

A data repository fixes this at the source. It gives the assistant:

• Your voice, so drafts sound like you and not like every other brand.
• Your facts, so it stops guessing about offers, pricing, hours and policies.
• Your audience, so it writes for a specific person.
• Your proof, so claims are grounded in real results and real testimonials.
• Your rules, so it avoids topics, phrases and claims you can't make.
• Your examples, so it has models of "great" to imitate.

What is an AI data repository?

An AI data repository, sometimes called a knowledge base, context library or "brain," is a curated, well-organized collection of documents that describe your business and your way of working, written so both humans and AI can read them. It is not a data lake and it does not need a database. For most small teams it is a folder of plain-text or Markdown files, kept current and easy to upload or connect to your AI tool.

The qualities that make one good:

• Curated: only content that helps the AI do its job.
• Accurate and current: one source of truth, with dates and owners.
• Consistent: the same names, terms and structure across documents.
• Readable: short files with clear headings, not giant scanned PDFs.
• Safe: free of personal, confidential or regulated information that should not be shared with the tool.

Build it in seven steps

Step 1: Define the jobs the AI will do

List three to five jobs, such as writing social captions, drafting email newsletters, answering customer questions, preparing sales call briefs or summarizing reports. The jobs decide what goes in the repository. If a document doesn't help a job, leave it out.

Step 2: Choose a simple structure

Start with a folder layout like this:

/ai-repository
  /00-readme.md                  <- what this repository is, who owns it, how to update it
  /01-brand
    brand-voice.md               <- tone, vocabulary, do and don't
    positioning.md               <- who we serve, promise, differentiators
    messaging-pillars.md         <- 3-5 core messages with proof
  /02-audience
    ideal-customer-profiles.md   <- who they are, goals, fears, objections
    customer-language.md         <- real phrases from reviews, calls, DMs
  /03-offers
    services-and-pricing.md      <- what we sell, for whom, price signals
    faq.md                       <- verified answers to common questions
    policies.md                  <- refunds, shipping, booking, terms
  /04-proof
    case-studies.md              <- results with context
    testimonials.md              <- approved quotes with permission notes
  /05-content
    golden-examples.md           <- our best posts, emails, pages, annotated
    content-pillars.md           <- topics, formats, cadence
    banned-phrases-and-claims.md <- words and claims to avoid
  /06-operations
    processes.md                 <- how we do recurring work
    glossary.md                  <- terms, product names, spellings
  /99-archive                    <- retired versions, kept out of the AI's reach

Numbered folders keep the order stable. A short readme tells any human or assistant what the repository is for.

Step 3: Write documents in a format both people and AI can use

Plain text and Markdown are ideal. They are lightweight, portable and easy to version. Use headings, short paragraphs, bullet lists and tables. Avoid burying information in images, scans or screenshots, because text inside images may not be read reliably. If you must use a PDF, make sure it contains real, selectable text.

Give every file a small header so it is clear what it is and whether to trust it:

---
title: Brand voice
owner: Marketing lead
status: approved
last_reviewed: 2026-09-20
applies_to: social captions, emails, website copy
---

Step 4: Write the five documents that matter most

If you write only five, write these.

1. Brand voice.

# Brand voice
## In one sentence
We sound like a smart friend who knows the industry: warm, direct, never salesy.

## Do
- Use plain words and short sentences.
- Lead with the customer's problem.
- Use specific numbers and examples.

## Don't
- Use jargon such as "synergy" or "cutting-edge".
- Promise results we can't guarantee.
- Use exclamation marks more than once per piece.

## Vocabulary
Say "clients" not "users". Say "consultation" not "sales call".

2. Ideal customer profile.

# Ideal customer: [Name of segment]
- Who they are:
- What they want:
- What they fear or doubt:
- Where they look for answers:
- Words they use (quote real phrases):
- What makes them say yes:

3. Offers and pricing. A table with each offer, who it is for, what is included, what is excluded, price or price signals and the next step. Mark anything that changes often with a review date.

4. Verified FAQ. The 25 to 50 questions your team answers over and over, with approved answers. This is the single most valuable document for customer-facing assistants.

5. Golden examples. Five to ten of your best real pieces (a caption, an email, a landing page section) with a note on why each works. Examples teach style faster than descriptions.

Step 5: Decide what must stay out

A repository is only useful if it is safe. Leave out:

• Personal information about customers or patients, including names tied to health details.
• Credentials, passwords, API keys and internal security information.
• Confidential contracts, financials and anything covered by a non-disclosure agreement, unless your tool and agreement allow it.
• Regulated data (for example health or payment data) unless you are using a service configured and contracted for that purpose.
• Content you do not have the right to share.

If you work in healthcare or aesthetics, this deserves extra attention. My med spa marketing guide explains why consent and privacy must be designed in from the start. When in doubt, describe the pattern (for example, "clients often ask about downtime after treatment") without including personal data.

Step 6: Load it into your tools

Both major assistants offer ways to keep reusable context. Features and limits change often, so check current documentation, but the general options are:

• Project-style workspaces. Claude offers Projects, where you can add reference documents and custom instructions that apply to every conversation in that project. ChatGPT offers Projects and custom GPTs, which similarly let you attach files and instructions.
• Custom instructions. Short standing instructions such as your role, audience, tone and rules.
• Uploading files in a single chat for one-off tasks.
• Connectors and automations that let assistants read from tools such as your drive, CRM or help center, when appropriate and permitted.

Tips for using your repository well:

• Put reference material first and the request last when you paste it into a prompt.
• Wrap documents in clear labels (for example a tag or heading with the document name), so the assistant knows what each is.
• Ask the assistant to quote or cite the source document before answering factual questions, so you can verify.
• Add a rule to answer only from the provided material and to say "I don't know" when it is missing. This reduces invented answers.
• Keep instructions short and specific, and put the details in the documents.

Step 7: Test it and keep it alive

Create a small evaluation set: ten realistic tasks and the answer or qualities you expect. For example: "Write a caption announcing our new membership," "What is your cancellation policy?" and "Summarize our positioning for a new hire." Run them before and after you change the repository, and track whether the results improve.

Then set a maintenance rhythm:

• Monthly: review offers, pricing, hours and any changed policy.
• Quarterly: refresh your golden examples and customer language with new material.
• After every launch or policy change: update the source documents first.
• Annually: prune what is outdated and archive old versions.

Assign a named owner for each file. Repositories decay when everyone owns them and no one updates them.

A worked example: the same request, with and without a repository

To see why context matters so much, compare two versions of the same request. Imagine a small aesthetics practice that wants a caption announcing a membership.

Version A, no repository:

Write an Instagram caption announcing our new membership.

The result is typically upbeat and generic: something about "glowing skin," an exclamation-heavy opening and a vague invitation to learn more. It could belong to any practice in any city. Worse, it may invent benefits you don't offer.

Version B, with a repository:

Role: You are the social media writer for [Practice name].
Context: Use the attached brand voice, membership details and FAQ.
Task: Write three Instagram caption options announcing our Glow Membership.
Constraints: Under 120 words. Warm, direct, no exclamation marks beyond one.
Only state benefits listed in the membership document. Do not make medical
claims or promise results. Include the booking link placeholder.
Format: Caption, then a suggested call to action, then three hashtag ideas.
If anything you need is missing from the documents, say so instead of guessing.

Because the assistant has the actual membership terms, the voice guidelines and explicit rules, the output uses your vocabulary, names the real benefits and stays inside your compliance boundaries. It also tells you when it lacks information, which is the behavior you want.

The prompt in Version B is not especially clever. It is just specific, grounded and rule-aware, and every part of it is possible only because the material exists in a repository. That is the whole idea.

What to check every time

• Did it use only information from our documents?
• Does it sound like us? Compare against a golden example.
• Are there claims we cannot support? Remove them.
• Would a specific customer understand and act on this?
• Has a human approved it before it goes public?

If the answer to any of these is no, fix the source document, not just the output. Improving the repository makes every future answer better.

Common mistakes and how to avoid them

A seven-day starter plan

• Day 1: Choose one job and write down what "good" looks like.
• Day 2: Write your brand voice document and a five-sentence positioning statement.
• Day 3: Write one ideal customer profile using real customer language.
• Day 4: Build your offers table and start your FAQ with ten questions.
• Day 5: Collect five golden examples and annotate them.
• Day 6: Load the documents into a project or custom assistant, and add instructions to answer only from the material.
• Day 7: Run your ten-task evaluation, note what failed and fix the documents.

By the end of a week you will have something small, useful and measurable, which is worth more than an ambitious repository you never finish.

How this connects to the rest of your marketing

A good repository improves everything downstream. Social content becomes more consistent, which supports the trends I cover in social media trends for 2026. Content creation gets faster without losing originality, which matters given how platforms treat reposted or generic material, as I explain in 10 things to keep in mind when starting your social media content creation. And the same documents can power website copy, email flows and customer support answers, so your voice stays consistent everywhere.

Frequently asked questions

What is an AI data repository? It is an organized set of documents about your business, such as brand voice, audience, offers, FAQs, proof and examples, kept in a clean format and used to give AI assistants accurate context so their outputs are on-brand and factual.

How do I start using AI for my business? Pick one narrow, low-risk job, choose one assistant, use a clear prompt structure, review every output and save what works. Then turn your best prompts and facts into a simple repository.

Should I use ChatGPT or Claude? Both are strong. They differ in style, features and plans, and those details change often. Test the same real task in each and choose the one that produces the best results for your work.

Can I put customer data in my AI tools? Be careful. Avoid entering personal, confidential or regulated information unless the tool and your agreement are designed for it and you have the right to use the data. When in doubt, anonymize or leave it out.

How big should my repository be? Small enough to maintain. Many teams get most of the value from five to ten well-written documents. Add more only when a specific job needs it.

How do I stop AI from making things up? Give it verified source material, instruct it to answer only from that material and to say when it doesn't know, ask it to quote its sources and always review before publishing.

Key takeaways

1. AI is a force multiplier, not a strategy. Keep human judgment in the loop.
2. Start with one narrow job and one assistant.
3. Use a clear prompt structure, and iterate out loud.
4. Build a small, curated AI data repository in plain text: voice, audience, offers, FAQ, proof and examples.
5. Keep sensitive and regulated data out.
6. Test with a small evaluation set and maintain it monthly.

If you want help designing an AI-assisted workflow, from the repository to the automation, CTRL+R Studio does this work across marketing, content and operations. You can contact Luis A. Laguna Torres or book a strategy call, and you can see live prompts and outputs on the AI World page.

Sources and further reading

• Anthropic, Prompting best practices, Claude documentation.
• OpenAI, prompt engineering guidance, platform.openai.com/docs.
• Luis A. Laguna Torres, AI World: real prompts and code, Lagoon of Randomness.
• Luis A. Laguna Torres, AI won't save a weak brand, Lagoon of Randomness.