---
spec_name: "AI Agents Builder — Implementation Spec for Claude Code"
spec_version: "1.0"
last_updated: 2026-05-07
target_audience: "Designers, creative directors, and freelancers shipping client work on Wix / WordPress / Webflow / Shopify who want to build autonomous AI agents in plain English — no engineering team required"
expected_outcome: "You ship your first working AI agent (client comms loop, weekly site QA, or Figma-to-embed code) on the right platform for your stack within one week"
ai_tools_supported: [claude_code, google_antigravity, cursor]
based_on: https://mikekwal.com/blog/ai-agents-builder
author: Mike Kwal
---

# AI Agents Builder — Implementation Spec for Claude Code

## Instructions for the AI

You are helping a designer or creative director build their first autonomous AI agent. Read this spec end-to-end, then walk the user through Steps 1–4 in order. Confirm each step before moving to the next.

The user is **not an engineer**. Default to plain-English explanations. If a step requires a code editor, walk them through opening one. If a step requires API keys, tell them exactly where to click. If you don't know what platform they're on (their team's daily tool), ASK before generating prompts.

For every change, explain what you're doing in one sentence so the user learns as they go.

---

## Step 1: Pick the right platform (90-second decision)

Wrong platform turns a 30-minute job into a 30-hour job. Answer two questions:

**Q1 — Where does your team already work every day?**

| Tool | Platform |
|---|---|
| ChatGPT Business/Enterprise | **OpenAI Workspace Agents** |
| Microsoft 365 | **Microsoft Agent 365** |
| Google Workspace | **Gemini Enterprise Agent Platform** |
| Notion/Slack/Linear/GitHub + Claude | **Claude Managed Agents** |
| Wix client site | **Wix Automations Builder** |

**Q2 — What's the agent doing?**

- Pre-built department work (finance, HR, legal) → **Anthropic Plug-and-Play**
- Coding (tickets → PRs, embeds) → **Cursor SDK**
- Cross-tool (email + Slack + CRM) → **OpenAI Workspace Agents**
- Custom business logic → **Claude Managed Agents**

If both answers agree, that's your starting point. If they conflict, **pick by where your team works** — adoption beats power.

---

## Step 2: Pick a designer use case to ship first

Don't build the perfect agent. Pick one of these three:

- **A — Client comms loop** — watches Gmail + Slack for project tag, drafts replies, posts Friday digest. Best on **OpenAI Workspace Agents**.
- **B — Weekly site QA** — Monday sweep across every client site for broken links, alt text, schema, AEO basics. One-line Slack health report. Best on **Claude Managed Agents**.
- **C — Figma-to-embed code agent** — reads a Figma annotation, writes a Webflow embed / Shopify Liquid / WP block, opens a PR. Best on **Cursor SDK**.

Pick one. Ship it. Then come back for the next.

---

## Step 3: Set up the agent (platform-specific)

### If you picked OpenAI Workspace Agents (Use case A — client comms loop)

1. Go to chat.openai.com → sidebar → **Agents** → **Create**.
2. Name it: `[Client Name] Comms Watcher`.
3. Connect tools: **Gmail**, **Slack**, **Google Drive**, **Calendar**.
4. Paste this brief into the agent description:
   > "You are a comms watcher for the [CLIENT NAME] project. Watch Gmail and Slack for anything tagged with [PROJECT TAG]. When a client message comes in: (1) draft a reply in my voice, (2) flag any unresolved question, (3) every Friday at 4pm, post a digest to #[CLIENT-CHANNEL] in Slack with the week's open items, decisions, and what's blocking. Never send anything without my approval. Always show me the draft first."
5. Set memory: ON. Set schedule: continuous + Friday 4pm digest.
6. Share with the team.

### If you picked Claude Managed Agents (Use case B — weekly site QA)

1. Go to console.anthropic.com → **Managed Agents** → **Create new agent**.
2. Name: `Weekly Site Health Sweep`.
3. Tools: web fetch, schema validator, link checker, Slack webhook.
4. Paste this brief:
   > "Every Monday at 7am, run a health sweep across this list of client sites: [LIST URLS]. For each site, check: (1) broken internal links, (2) missing alt text on images above the fold, (3) valid JSON-LD schema in <head>, (4) AEO basics (llms.txt reachable, FAQ blocks present on cornerstone pages, dateModified within 6 months). Output a single Slack message to #site-health with one line per site: green / yellow / red plus the top 1 issue if not green."
5. Schedule: weekly, Monday 7am.
6. Set guardrails: read-only, no writes to client sites.
7. Pricing baseline: ~$58/month for 24/7 runtime, less if you cap to weekly.

### If you picked Cursor SDK (Use case C — Figma-to-embed)

1. `npm install @cursor/sdk` in a fresh TypeScript project.
2. Create `agent.ts` with the minimum-viable agent (under 15 lines):

```ts
import { Agent } from "@cursor/sdk";

const agent = new Agent({
  name: "figma-to-embed",
  instructions: `Read the Figma annotation passed in. Generate the matching
    Webflow custom embed (HTML+CSS+inline JS only). Output to ./embeds/[slug].html.
    Open a PR titled "Embed: [annotation title]" and tag @mikekwal for review.`,
  tools: ["filesystem", "git", "shell"],
});

await agent.run({ figmaNodeUrl: process.argv[2] });
```

3. Run: `npx ts-node agent.ts "https://figma.com/design/.../node-id=..."`.
4. Approve the PR. Done.

---

## Step 4: Verify the agent earns its keep

After the first run, check three things:

| Check | How |
|---|---|
| It ran | Confirm "completed" in the run log. |
| Output is usable | Would you ship it with light edits? If no, refine and rerun. |
| Time saved is real | After 7 days, count hours saved. Less than 1? Kill it or rewrite the brief. |

If output fails twice, the brief is too vague. Add a good example, a bad example, and a "show me the draft first" rule.

---

## Quick-start prompts for Claude Code / Antigravity / Cursor

Paste these into your AI builder. Replace bracketed placeholders.

### Prompt 1 — Pick my platform

```
Read the AI Agents Builder spec at this path. I'm a [designer / creative director / freelancer]
and my team mostly works in [TOOL]. The agent I want to build does [DESCRIBE THE JOB IN 1 SENTENCE].
Tell me which platform from Step 1 to use, and why.
```

### Prompt 2 — Generate the OpenAI Workspace Agent brief

```
Using Step 3 of this spec, generate a Workspace Agent brief for client comms on
the [CLIENT NAME] project. The project tag in Slack/email is [TAG]. Their shared
Slack channel is #[CHANNEL]. I want a Friday 4pm digest. Output the exact text
I should paste into the OpenAI agent description field.
```

### Prompt 3 — Generate the Claude Managed Agent brief

```
Using Step 3 of this spec, generate a Claude Managed Agent brief that runs every
Monday at 7am and sweeps these client sites for broken links, missing alt text,
busted schema, and AEO basics: [LIST CLIENT URLS]. Slack channel for output is
#site-health. Output the exact text I should paste into the agent description field.
```

### Prompt 4 — Scaffold the Cursor SDK Figma-to-embed agent

```
Using Step 3 of this spec, scaffold a fresh TypeScript project with @cursor/sdk
that reads a Figma annotation URL from argv, generates a Webflow custom embed,
saves it to ./embeds/[slug].html, and opens a PR. Include package.json, tsconfig,
and the agent.ts file. Don't run it yet — show me the diff first.
```

### Prompt 5 — Refine when output is weak

```
The agent's last output was off. Here's what it produced: [PASTE]. Here's what
I wanted: [PASTE]. Rewrite the brief — add a good example, a bad example, and
a rule about [WHAT WAS WRONG].
```

---

## OpenAI Assistants API — migration check

OpenAI's old Assistants API shuts off **Aug 26, 2026**. Run this on every OpenAI project:

```
Search this codebase for client.beta.assistants or /v1/assistants. List every
file and line. Propose migration to the Responses API per
https://platform.openai.com/docs/guides/migrate-to-responses. Show diff only.
```

Internal deadline: **July 2026**. Don't bet on the grace period.

---

## Verification checklist

- [ ] You picked one platform from Step 1 (not three)
- [ ] You picked one use case from Step 2 (not three)
- [ ] The agent ran end-to-end at least once
- [ ] You read the output and would ship it with light edits
- [ ] Time saved after 7 days is more than 1 hour
- [ ] If you use the OpenAI Assistants API anywhere, you've started migration

---

## Need help?

Four ways to go deeper:

- **Free** — Join the [Talk-to-Build community](https://www.skool.com/talktobuild). Builders shipping AI agents into client work, weekly office hours, prompt library, build-along sessions.
- **Done-for-you** — [MK-Way](https://mk-way.com) builds and deploys agents into client websites in 7 days.
- **Quick question** — DM Mike on [Instagram](https://www.instagram.com/mikekwal). He reads every message.
- **B2B / strategy** — Connect on [LinkedIn](https://www.linkedin.com/in/mikekwal/).

---

*This spec is the actionable implementation companion to the AI Agents Builder hub article: [https://mikekwal.com/blog/ai-agents-builder](https://mikekwal.com/blog/ai-agents-builder). The hub explains the why; this spec ships the what.*

*Last updated: 2026-05-07*