Before You Build Your First AI Agent, Read This.

Oct 11, 2025

Your AI is dumb.

You're rushing. You see the potential. You want the leverage. So you throw a few documents at a custom GPT and expect a miracle.

You will not get one. You will get a high-speed intern who sounds confident while making things up.

Creating an AI agent isn't magic. It's engineering. A great engineer doesn't start with code. They start with a blueprint, a plan, and a pile of the right materials. In the realm of AI, those "materials" are your data: the stories you've written, the emails you've sent, the strategies you've drafted.

Without a solid foundation, your agent is a house of cards.

Why Data Curation Comes First

Imagine hiring a ghostwriter and handing them only your LinkedIn headline. Would you expect a masterpiece? Of course not. The same principle applies to AI. A content‑writing agent fed on thin gruel will produce thin gruel. Getting the right inputs is a discipline.

For content creation, start by gathering your best pieces of writing.

I pull three recent blog posts and convert them into Markdown using a simple prompt like this:

Transform "https://www.piotrmechlinski.com/blog/building-ai-strategy-part-3-run-to-win" into a markdown (.md) file with the same formatting for headlines, bold text, etc.

Output: blog-building-ai-strategy-part-3.md

Markdown is universal and lightweight. It preserves formatting across platforms and lets you edit with ease. It's like having your writing in a transparent container, rather than a locked PDF. Here's what Markdown gives you:

Portability: As a simplified form of HTML, Markdown can be read by both humans and machines. The files are small and easy to move or attach to prompts.
Editability: Tools like Obsidian allow you to open, annotate, and refine your samples without proprietary software. You could also use any plain‑text editor. I use the free version of Obsidian, and I find it great.
Flexibility: Because Markdown preserves headings, bold text, and other stylistic cues, an AI can learn your formatting quirks as well as your tone.

Once you have at least three strong samples, ask an AI to distill your unique style.

Here is a prompt that instructs the model to analyze voice, tone, sentence structure, word choice, rhythm, and point of view:

Task 1: Analyze the files I attached to uncover my writing style.

Information about me: I’m writing in the style of AI Expert and an Executive Coach. My target audience is Senior executives responsible for AI Strategy development who are overwhelmed by the pace and complexity of AI, seeking calm, clarity, and minimalism. (change to yours)

Task 2: Prepare my writing style description.

This description should include the sections below, each with clear, simple bullet points that are easy to understand and correct by a human reader. My writing style description should be also easy to understand for LLMs like you. For each section, you have a definition in the brackets, so you know what to focus on.

Voice:
1. Describe what stays consistent across all my writing.
2. Voice archetype mix (Storyteller, Opinionator, Fact Presenter, Frameworker, F-Bomber).
Tone (changes from piece to piece, list my most common tones).
Style:
1. Sentence structure: how sentences are constructed, length, and complexity.
2. Words: preferred, avoided, grade of English I use.
3. Rhythm: the flow and tempo of the writing.
4. Point of View: The perspective from which the story is told.
5. Figurative Language: Use of literary devices like metaphors.
6. Imagery: Descriptive language that creates visual images for the reader.

Remove duplicates and only output the unique style traits in a simple-to-read bullet point list

My writing style description should have a maximum of 20 bullet points in total.

Task 3:

Prepare a markdown file CURRENT-DATE-my-writing-style.md, where CURRENT-DATE in the file name should be replaced with the current date in format YYYY-MM-DD.

This file should include:

Information about me from Task 1.
My style of writing that has been prepared in Task 2.

Make sure the file is clean, simple and ready to use.

Remove any unnecessary descriptions and headlines that not deliver value for the writing style description.

The result is a concise description of your writing style—a style guide that can be reused across various tools and platforms. I would download the guide as a .md file, review it myself, and make corrections. This extra step matters: you're teaching the machine how to sound like you. Don't outsource that.

I would not recommend building an AI agent for writing at this stage, as this limits the use of modes like Deep Research or Agent in ChatGPT. So after drafting your post, attach a .md file with your writing style and tell your LLM to use it. Practice and refine this file until you're satisfied. I will come back to building agents for writing in future posts.

Gathering the Right Documents for a Strategic Advisor

Creating an AI content writer is simple. Creating a strategic advisor is not. A strategic advisor needs a deep understanding of your vision, constraints, and market.

Before you build, you must assemble its brain. This is a one-hour sprint to gather the essential documents that will ground your AI in reality.

Documents to Find and Export:

First, gather the existing artifacts that define your business. Create a cloud folder and export the following documents. Use lightweight formats, such as PDF or text files, where possible.

Company Overview: Your standard presentation covering mission, values, and overall organization.
Leadership Deck: The latest presentation delivered to the board or leadership team.
Organizational Chart: A current view of your team structure.
Financial Summary: A 12-month Profit & Loss (P&L) summary, including revenue broken down by product or region.
Sales & Customer Data: A current sales pipeline snapshot and a customer list detailing wins, losses, and churn.
Customer Voice: Raw, verbatim customer feedback or recent Net Promoter Score (NPS) comments.
Risk & Compliance: The latest version of your risk register or key compliance rules.

Documents to Create:

Next, add your own context. This is the qualitative insight that the AI cannot find in formal reports. Create these as simple text files (.md or .txt).

Competitive Landscape: A brief note listing your top competitors and a frank assessment of why you win or lose against them (10 sentences is enough).
Hard Constraints: A simple list of the top three constraints (financial, operational, market-based) that your business currently faces (again, a few sentences is a good start).
Strategic Directive: A single text file that contains:
- A 10-line summary of your current strategy.
- The three most important strategic questions you want the AI to help you answer.
- A statement clarifying which of the uploaded data sources the advisor must trust most.

With these files collected, run a fast data audit. Is everything current? Are there apparent contradictions? Rename messy files and remove duplicates.

This isn't busywork. It's the scaffolding that enables your AI advisor to deliver a genuine signal, not noise.

Building a Strategic Advisor

Once your documents are in order, the next step is crafting a custom instruction that sets the AI's role and guidelines.

Drawing on best practices and my own experience, here's a template you can adapt for a custom GPT:

Role & Goal

You are [Company Name]’s Strategic Advisor. Your role is to provide expert-level insights and recommendations tailored to our specific business context. Your ultimate purpose is to serve as a skeptical, unbiased sparring partner to challenge my assumptions and improve strategic decision-making.

Context & Knowledge

You have been trained on our company's mission and values, customer insights, marketing and sales strategies, financial and operational data, competitive landscape, strategic plans, long-term goals, risk assessments, and market research.

Core Principles & Persona

Analytical Approach: Analyze all situations through multiple lenses: strategic, financial, operational, and psychological.
Root Cause Focus: Your primary function is to identify systemic root causes rather than surface-level symptoms.
Tone: Maintain brutal honesty without being destructive. Avoid generic advice. Your responses must always be structured, specific, and contextual.

Interaction Workflow

For any challenge I present, you must follow this exact five-step process:

Start With the Truth: Begin your response with a direct, unvarnished truth about the situation.
Ask Clarifying Questions: Ask me 2-3 insightful, clarifying questions designed to expose potential blind spots. Do not proceed to the next step until I have answered.
Analyze Root Causes: Based on my answers, analyze the systemic root causes of the issue.
Propose Strategies: Provide 2-3 actionable, high-leverage strategies with clear, sequential steps. Where applicable, reference relevant strategic frameworks (e.g., Blue Ocean Strategy, Porter's Five Forces) to support your recommendations.
Assign a Challenge: Conclude with a specific, tangible challenge or assignment for me to complete.

Response Guidelines

Your final output must adhere to the following rules:

Data-Backed: Base all answers strictly on the provided documents. When suggesting actions, reference specific data points.
Action-Oriented: Offer at least three concrete, prioritized recommendations. If a question is too broad, ask clarifying questions before providing an answer.
Concise and Professional: Keep responses focused and directly relevant to our business goals.
Ethically Aware: Flag any potential compliance, privacy, or bias issues you identify. Your role is to provide analysis and options, not to make final decisions.
Scoped: If the information required to answer is missing from your knowledge base, state that clearly. Do not invent or "hallucinate" facts.

Upload your curated documents and this prompt into your chosen platform for creating a custom agent (ChatGPT, Gemini, Claude, or Copilot).

Test the advisor with real strategic questions.

Fine-tune the custom instructions (prompt above) until the answers feel precise and grounded.

Guarding Your Data

You wouldn't hand your confidential strategy deck to a stranger.

Treat AI the same way. Use paid versions of large‑language models that allow you to disable data collection for model training.

OpenAI notes that ChatGPT Business, Enterprise, and Edu plans do not use inputs to train models by default. For the Plus and Free plans, you can disable model improvement. Microsoft's Copilot and Google's Gemini offer similar toggles, which are switched off by design. Always check. Disabling training helps protect sensitive information.

If your organization prohibits the use of external AI tools, consider exploring on-premise options or waiting until policies evolve. Use rational guardrails, easier for content creation than for strategic advisory.

Final Thoughts: Garbage In, Garbage Out

Building an AI agent without careful preparation is like asking a junior analyst to brief the board without ever reading your strategy. The results will be generic at best and dangerous at worst. Spend the time to collect and curate your data. Convert your best writing into Markdown, distill your style, and craft an explicit instruction. For strategic advisors, gather a 360-degree view of your business—its mission, goals, competitive environment, risks, and financials—and protect it with the proper privacy settings.

AI can amplify your thinking, but it cannot replace it.

Use the agent as a sounding board, not a decision maker.

Draft your own thoughts first, then invite the agent to challenge them.

Your move.