3.2: Construct¶

Part of: Build Workflows

What This Is¶

In Design (3.1), you worked with the model in two phases — first collaborating on architecture decisions and execution pattern (normal conversation), then activating plan mode so the model could plan the full AI Building Block Spec. You reviewed and approved that spec.

Now the model shifts from planning to execution. Exit plan mode — the model takes the approved spec and builds the actual building blocks your workflow needs, in whatever format your platform requires.

This step is facilitated by the building-workflows Business-First AI Framework Skill — the same skill that ran Design. After you approve the AI Building Block Spec, the model transitions into Construct and begins generating building blocks.

If you're starting Construct separately (e.g., in a new conversation), provide the AI Building Block Spec:

Build the workflow from my AI Building Block Spec.

Upload or paste your [workflow-name]-building-block-spec.md file when prompted.

What Construct Produces¶

By the end of this step, you'll have two things:

Your building blocks — the prompts, skills, agents, context files, and configurations your workflow needs, generated in whatever format your platform requires. The model only builds the building blocks your execution pattern calls for — a Prompt pattern doesn't get agent files, and a Single Agent pattern doesn't get multi-agent orchestration.

Run Guide ([name]-run-guide.md) — a plain-language walkthrough tailored to your platform and technical comfort level:

1. What was built — every building block listed with what it does and where it lives
2. Setup steps — numbered instructions for getting each artifact into the right place on your platform
3. First run — a guided test with sample input, expected behavior, and common first-run issues
4. What to do next — how to run it again, share with teammates, and when to revisit

The Run Guide is your bridge from Construct to Run (3.3) — follow it to get your workflow running.

What Gets Built¶

The model constructs from the nine AI building blocks — but only the ones your workflow needs:

How the Model Builds for Your Platform¶

The model generates building blocks in whatever format your platform requires — but it doesn't guess. It follows a two-step research process:

1. Start with curated platform documentation. The cookbook maintains platform reference pages for Claude, OpenAI, Google Gemini, and M365 Copilot — each with links to official documentation, SDK references, and setup guides. The model reads the reference page for your platform first.
2. Verify and supplement via web search. The model then uses web search to confirm the documentation is still current and to find any newer resources. This catches renamed APIs, deprecated SDKs, and newly released tools.

This hybrid approach anchors the model in authoritative sources while ensuring it stays current as platforms evolve. The model provides the implementation (how to build it on your platform); the AI Building Block Spec provides the specs (what each building block should do).

Pattern-Specific Build Paths¶

Your execution pattern (chosen in Design) determines which building blocks get built. The model works through only the steps that apply:

Code-First Build Paths¶

When your architecture approach is code-first, the same four execution patterns apply — but the implementation uses APIs and SDKs instead of platform UIs. The model generates code artifacts instead of platform configurations.

The Run Guide for code-first workflows includes additional setup details:

• Dependency installation — pip install, npm install, or equivalent for your SDK
• Environment variable setup — API keys, configuration values, and how to set them
• How to run locally — python main.py, npx tsx agent.ts, or equivalent
• Deployment options — Docker, cloud services, CI/CD pipelines (when applicable)

Your Role During Construct¶

The model does the heavy lifting — reading your spec, researching your platform, and generating artifacts. But a few things require your direct involvement because the model can't do them on its own.

Context — gather your business-specific materials¶

The model generates prompts, skills, and agents, but it can't create your business knowledge. You need to provide the context artifacts that make the workflow specific to your organization.

1. Open the Context Inventory from your AI Building Block Spec
2. Focus on "Needs Creation" items — these are gaps you need to fill
3. Start with rules and criteria — scoring rubrics, qualification criteria, and decision logic are often required for the workflow to function correctly. Without them, the model guesses instead of following your standards.
4. Use the "Key Contents" column as your drafting checklist — the spec already identified what each artifact should contain
5. Create each artifact as a standalone Markdown file or data export (e.g., buyer-persona.md, scoring-rubric.md)
6. The model will tell you what it needs and when — if you're using the building-workflows skill, it prompts you for each context artifact at the right point in the build

MCP — configure external tool connections¶

If your Building Block Spec lists external tools in the Tools and Connectors Required section, you'll need to set up those connections. Most workflows don't need this — only when live data or external actions are required.

Connecting an external tool works like connecting an app to your phone — set it up once, authorize access, then it's available whenever you need it:

1. Get access credentials — Most tools require an API key. Create this in the external tool's settings, usually under "Integrations" or "Developer."
2. Add the connection to your AI tool — Tell your AI tool where to find the external service and how to authenticate.
3. Test it — Ask the AI to do something simple with the connected tool to verify the connection works.

How different platforms handle external connections:

Agents — build on your platform¶

When the model generates agent blueprints (for Single Agent and Multi-Agent patterns), how you operationalize them depends on your platform. Agent platforms fall into two categories:

Builder/GUI platforms — You take the agent details from your AI Building Block Spec (name, instructions, tools, model) and manually configure them through the platform's visual interface. This is point-and-click — paste the instructions, select the tools, and configure settings.

Code-based platforms — Your AI Building Block Spec gives you the what (agent role, instructions, tools, goals), but you'll need to translate that into code using the platform's SDK or framework. Refer to the platform's documentation for how to operationalize the agent configuration — defining tools, setting up orchestration, and deploying.