> ## Documentation Index > Fetch the complete documentation index at: https://docs.adsmcp.si/llms.txt > Use this file to discover all available pages before exploring further. # Prompt Engineering Playbook: A Trained Agent for One Ad Account > The Prompt Engineering Playbook is a one-time $39 add-on that generates a custom agent skill, an account dossier, a prompt book, and agent instructions from your own ad account — then hands you the files to install in Claude, ChatGPT, Claude Code, Cowork, Cursor, or Codex. The **Prompt Engineering Playbook** is a set of agent files generated from **one of your ad accounts** — its campaign structure, spend history, and current state — and handed to you as a download you install in your AI client. It answers the question people ask after connecting ADSMCP: *"the tools work, but how do I actually get good at this?"* Instead of you learning prompt engineering, the playbook writes the prompts, the instructions, and the account briefing for you. > [!NOTE] > > **The single most important thing on this page: this is a one-time purchase, not a subscription.** > You pay **\$39 once** for a given ad account and nothing recurs — no monthly charge, no renewal, no > second bill. It is separate from your ADSMCP plan, which continues to work exactly as before. The > files are yours to keep, and rebuilding them later is free and unlimited. ## What it is — and what it is not ADSMCP ships **13 free [Agent Skills](https://docs.adsmcp.si/agent-skills/skills.md)** that teach any assistant how to run ads in general: how to structure a Google Search campaign, what Meta needs before a launch, the safety rules. Every plan gets those, at no cost. The Prompt Engineering Playbook is the layer above that. It is not general advertising knowledge — it is **your account**, written down. | | Free Agent Skills | Prompt Engineering Playbook | | - | - | - | | **Scope** | Advertising in general | **One specific ad account** | | **Knows your campaigns** | No | **Yes** — read from the live account | | **Knows your spend and currency** | No | **Yes** | | **Cost** | Free on every plan | **\$39 once, per ad account** | | **Who writes it** | ADSMCP, for everyone | Generated for you, from your data | You want both. The free skills are the grammar; the playbook is the briefing. ## What you get The download is a `.zip` laid out to the open [Agent Skills](https://agentskills.io) format, so the same folder works across clients. **One zip, one folder, flat inside** — only `references/` nests, because that's the spec's progressive-disclosure mechanism (those files load on demand instead of on every activation): ``` -expert-/ ├── SKILL.md # the operating doctrine, pinned to your account ├── AGENTS.md # your account's instructions, in the agents.md standard ├── PROMPTBOOK.md # the prompt book, organized by level and workflow ├── README.md # where each piece goes, per client └── references/ ├── account-profile.md # the dossier — your account's state and priorities └── working-with-adsmcp.md # how the agent should call ADSMCP's tools ``` | File | What it does | Where it goes | | - | - | - | | **`SKILL.md`** | The agent's operating doctrine for this account. Loads automatically when you ask about this account — you never invoke it by name. | Uploaded or dropped in as a skill | | **`AGENTS.md`** | The constitution — the account-pinned rules, in the [agents.md](https://agents.md/) standard. Read automatically by folder-based clients, and the text you paste into a chat client's instructions field. | Project root for coding agents; instructions field for chat clients | | **`PROMPTBOOK.md`** | The starting prompts — organized by level and workflow, written against *your* campaigns. | You read and copy these | | **`README.md`** | Says where each piece goes, per client, so the folder explains itself once unzipped. | You read this | | **`references/account-profile.md`** | A written dossier of the account: what state it's in, what to prioritize, what cadence to work on. | Read by the agent on demand | | **`references/working-with-adsmcp.md`** | The tool-calling contract and guardrails: discovery loop, read-before-write, pause-don't-delete. | Read by the agent on demand | > [!NOTE] > > **There is one instructions file, and it's `AGENTS.md`.** Earlier kits also shipped a separate > `AGENT_INSTRUCTIONS.md` with identical text; it was dropped as a duplicate. Coding agents read > `AGENTS.md` from a project root automatically, and for chat clients the kit page has a one-click > **Copy instructions** button that puts the same text on your clipboard. Kits you downloaded before > this change still work — keep using the file you have. > [!TIP] > > `PROMPTBOOK.md` is a **menu, not a script.** Your agent is told to open it to suggest a sensible > next step or to find the right approach for something you asked for — then run only that one. It > will not work through it start to finish, because every prompt in it spends real tool calls. ## How it's built When you buy, ADSMCP runs a five-stage build against the live account. It usually finishes in a few minutes, and the page shows each stage as it completes: **1. Reading account structure** Pulls the campaigns, ad groups, budgets, and settings that exist right now. **2. Classifying the account's state** Works out what kind of account this is — scaling, stalled, newly built, seasonal. **3. Writing the account dossier** Writes `account-profile.md`: the state read, the priorities, the working cadence. **4. Writing your prompt book** Writes `PROMPTBOOK.md` against your actual campaigns and currency. **5. Assembling the skill package** Builds the folder, validates it against the Agent Skills spec, and packages the `.zip`. If a build fails, the kit page shows it and offers a retry. A failed build costs you nothing extra. ## Price | | Price | Billing | Covers | | - | - | - | - | | **Prompt Engineering Playbook** | **\$39** | **One time — nothing recurs** | **One ad account** | * Priced **per ad account**, because it is generated from that account's data. Two accounts means two playbooks. * **Select every eligible account and 20% comes off the whole cart.** Tick them all in Expert Space — the cart bar has a **Select all** shortcut — and the bundle discount applies automatically at checkout. It needs **at least two** accounts, and it has to be *all* of the ones you can still buy; a partial selection pays full price per kit. The cart always shows the live discount before you pay. * Buying several is a **single checkout and a single charge**, not one payment per account. * **Annual plans include one playbook free.** If you're on an annual plan, claim it from Expert Space — there's no checkout step and no charge. * The add-on also appears as an optional **"Add to your order"** on plan checkouts. Adding it there charges the \$39 immediately, on the same invoice as your plan — which is why the line item reads **"One-time Agent Setup — 1 Ad Account"**. > [!WARNING] > > \*\*Seeing $39 next to your plan charge does not mean your plan went up.** If you took the add-on > during checkout, the $39 sits on the same invoice as your subscription, as a separate one-time line. > Your recurring plan price is unchanged, and the \$39 will not appear on next month's invoice. Your > [billing page](https://adsmcp.si) shows the recurring amount on its own. Because each playbook is generated on demand from your account, it's **non-refundable once the build has completed and the files are available to download**. If a build failed, or you were charged for an account you can't access, [contact support](https://docs.adsmcp.si/knowledge-base/support.md) — that's a different situation and we'll sort it out. ## How to get it **1. Open Expert Space** Go to [adsmcp.ai/expert-space](https://adsmcp.si/dashboard). You'll see one card per connected ad account, grouped by platform. **2. Pick the account** Cards show what the playbook will be built from — campaign count, months of history, currency. Accounts that aren't active show **Inactive**; activate them in [Connections](https://adsmcp.si/dashboard/connections) first. **3. Preview, then buy** **Preview Expert** shows sample prompts before you pay. Buy one account, or tick several and check out together. **4. Wait for the build** A few minutes. You can close the tab — the kit keeps building. **5. Download the .zip** From the kit page: **Download**. You can also hit **Copy instructions** to put the agent instructions straight on your clipboard, read the prompt book in the browser, or download the prompt book on its own as a `.md` file — handy for pasting into a doc or dropping beside a project without unzipping the whole kit. > [!TIP] > > Lost the link? Everything you've bought is always at > [adsmcp.ai/expert-space](https://adsmcp.si/dashboard) — owned accounts show an **Open Kit** > button. The purchase email also carries a direct link, but you never need it. ## Install it in your AI client There are three install shapes. Which one you use depends on the client, not on the playbook. | Shape | What you do | Clients | | - | - | - | | **Upload** | Upload the kit `.zip` into the client's Skills UI | Claude, Claude Cowork, ChatGPT *(Business/Enterprise/Edu only)* | | **Paste** | Paste the agent instructions into the project's instructions field, and attach the rest as project files | ChatGPT *(Free, Plus, Pro)*, Windsurf | | **Folder** | Copy the unzipped folder into the client's skills directory | Claude Code, Cursor, Codex, OpenClaw, Antigravity | The kit page's setup wizard shows the **literal install path** for whichever client you pick — `~/.claude/skills/` for Claude Code, `.agents/skills/` for Codex, Cursor, OpenClaw, and Antigravity — so you don't have to go looking it up. > [!NOTE] > > The playbook assumes ADSMCP is already connected in that client. It teaches the agent how to work > *your account*, not how to authenticate — connect first via [ChatGPT](https://docs.adsmcp.si/ai-clients/chatgpt.md), > [Claude](https://docs.adsmcp.si/ai-clients/claude.md), [Cowork](https://docs.adsmcp.si/ai-clients/claude-cowork.md), > [Claude Code](https://docs.adsmcp.si/ai-clients/claude-code.md), [Codex](https://docs.adsmcp.si/ai-clients/codex.md), or [Codex Desktop](https://docs.adsmcp.si/ai-clients/codex-desktop.md). ### Claude (web and desktop) **1. Turn on code execution** Skills need it. In Claude, go to **Settings → Capabilities** and turn on code execution. **2. Open Customize → Skills** Go to [Customize → Skills](https://claude.ai/customize/connectors) in Claude and open the **Skills** tab. **3. Upload the .zip as-is** Click **+ Create skill** and upload the kit `.zip` you downloaded — no unzipping first. The zip has the skill folder as its single root entry, which is exactly the shape Claude's uploader expects. Claude reads `SKILL.md` and the `references/` files together. **4. That's it — the skill is on your account** Skills are added at the **account** level, not per project, so your expert is available in every chat and every project, and activates on its own when you work on this account. Optionally, create a Project for this account and paste the agent instructions into its instructions field for extra focus — the kit page has a one-click **Copy instructions** button. Ask a normal question — *"how did we do last week?"* — and the skill loads on its own. You do not name it. See the [Claude setup guide](https://docs.adsmcp.si/ai-clients/claude.md) for connector details. ### Claude Cowork Cowork supports skills the same way, from the **Customize** panel. **1. Open Customize** Click **Customize** in the left sidebar. You'll see **Skills**, **Connectors**, and **Plugins**. **2. Add the playbook under Skills** Add the unzipped folder as a skill. It sits alongside the 15 skills the ADSMCP plugin installs — the free ones cover advertising in general, the playbook covers this account. **3. Use a brand folder** Open the brand folder for this account and drop `AGENTS.md` in it. Cowork picks it up as standing context for everything you do in that folder. See the [Cowork setup guide](https://docs.adsmcp.si/ai-clients/claude-cowork.md) for the plugin install. ### ChatGPT Which path you take depends on your **ChatGPT plan**, because ChatGPT's Skills feature is not on every tier. > [!WARNING] > > **ChatGPT Skills are available on Business, Enterprise, Healthcare, and Edu — not on Free, Plus, or > Pro.** If you're on Plus or Pro (the tiers the [ADSMCP ChatGPT plugin](https://docs.adsmcp.si/ai-clients/chatgpt.md) requires), > use the **Project instructions** path below. It gives you the same behavior without the Skills > feature. On Enterprise and Edu, skills may also be switched off until a workspace admin enables them. **Plus and Pro — paste into Project instructions** **1. Create a Project for this ad account** One project per ad account keeps the agent from mixing them up. **2. Paste the agent instructions** Hit **Copy instructions** on the kit page — one click puts the whole thing on your clipboard — and paste it into the project's custom-instructions field. (The same text ships in the kit as `AGENTS.md` if you'd rather open the file.) **3. Upload the rest as project files** Add `PROMPTBOOK.md` and `references/account-profile.md` to the project's files so ChatGPT can read the dossier and the prompts on demand. **Business, Enterprise, Healthcare, Edu — upload it as a skill** **1. Open Skills** Go to **Skills** in ChatGPT and choose **Create → Upload from your computer**. **2. Upload the playbook folder** ChatGPT scans the upload before it becomes available — most are ready straight after the scan. The playbook follows the open Agent Skills standard, so it's the same folder you'd install anywhere else. **3. Still worth making a Project** Paste the agent instructions into a project for this account too. The skill carries the expertise; the project keeps every chat scoped to one ad account. See [Skills in ChatGPT](https://help.openai.com/en/articles/20001066-skills-in-chatgpt), [ChatGPT custom instructions](https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions), and the [ChatGPT setup guide](https://docs.adsmcp.si/ai-clients/chatgpt.md). ### Claude Code, Cursor, Codex, OpenClaw, Antigravity Folder-based clients read the files off disk. Unzip the playbook into the directory you work in: ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}} theme={null} unzip google-ads-expert-acme-retail.zip # the folder contains SKILL.md, PROMPTBOOK.md, README.md, references/, # and the agent definition for your client ``` > [!TIP] > > **On Codex Desktop, skip the copying.** Attach the unzipped folder in a Codex chat and ask it to install the kit — Codex places the skill and agent definition itself. [Steps →](https://docs.adsmcp.si/ai-clients/codex-desktop.md#install-your-expert-kit) Copy the whole folder into your client's skills directory, then put the agent definition in its agents directory: | Client | Skills directory | Agent definition | | - | - | - | | **Claude Code** | `.claude/skills/` | `.claude/agents/` (markdown) | | **Codex** | `.agents/skills/` | `.codex/agents/` (TOML) | | **Cursor** | `.agents/skills/` | `.cursor/agents/` (markdown) | | **Antigravity** | `.agents/skills/` | `.agents/agents/` (markdown) | | **OpenClaw** | `/.agents/skills/` | — | > [!WARNING] > > These paths are **project-scoped on purpose**. The `~/` home-directory equivalents load one > account's ads expert into every unrelated project you open — including work that has nothing to do > with advertising. * **`SKILL.md`** is what the skills directory picks up. See [Claude Code skills](https://code.claude.com/docs/en/skills), [Cursor skills](https://cursor.com/docs/context/skills), [Codex skills](https://learn.chatgpt.com/docs/build-skills), and [OpenClaw skills](https://docs.openclaw.ai/tools/skills). * **The agent definition** carries your account instructions as the agent's system prompt and points at the skill. It's a uniquely named new file, so it can't overwrite anything you already have — and uninstalling the kit is deleting it. * **On Claude Code, Codex and Antigravity the agent file declares `mcp.adsmcp.si` itself**, so there's no connector step. **Cursor is the exception:** its subagents inherit tools from the parent and can't declare their own server, so add `https://mcp.adsmcp.si/mcp` to `.cursor/mcp.json` yourself. > [!NOTE] > > Earlier versions of the kit told you to copy an `AGENTS.md` to your project root. **Don't** — that > is a single well-known filename, and doing so replaces whatever instructions were already there > with "work only on this ad account". The agent definition above does the same job without > touching your existing setup. If you followed the old instructions, check your project root. **Claude Code** watches these directories, so it picks the skill up **without a restart**. Claude loads it when a task matches, or you can invoke it directly as `/google-ads-expert-acme-retail`. ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}} theme={null} mkdir -p .claude/skills .claude/agents cp -r google-ads-expert-acme-retail .claude/skills/ ``` **Codex** discovers skills by their `SKILL.md` frontmatter and loads the full file only when a task matches, so you don't invoke it by name — though you can with `$` to force it. Run `/skills` to confirm it loaded. ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}} theme={null} mkdir -p .agents/skills .codex/agents cp -r google-ads-expert-acme-retail .agents/skills/ ``` > [!TIP] > > The same skill folder works across all of these unchanged — one Agent Skills bundle, several > clients. Only the directory it goes in, and the format of the agent definition, differ. > [!NOTE] > > **Windsurf is no longer a supported client for the Expert Kit.** Cognition renamed it Devin > Desktop, it isn't on the Agent Skills client list, and nothing documents the desktop app loading a > local skill. Using Windsurf as an [MCP client](https://docs.adsmcp.si/ai-clients/windsurf.md) is unaffected. Managing several accounts from one machine? Keep one folder per account and open the one you're working on — see [set up your project](https://docs.adsmcp.si/guides/set-up-your-project.md) for the project-per-account and project-per-channel patterns. ## Set up the project, then run the prompts Installing the files is half of it. The playbook pays off when it lives inside a **project** — a persistent workspace scoped to that one ad account — so every chat starts already knowing the account. The pattern is the same in all three chat clients: **create a project → paste the agent instructions → add the dossier as knowledge → work through the prompt book.** **Claude** **1. Create the project** Go to [claude.ai/projects](https://claude.ai/projects) and click **+ New Project**. Name it for the ad account — *"Acme Retail — Google Ads"* — so you never wonder which account a chat is about. Projects are available on every plan, including free (free is capped at 5 projects). **2. Paste the agent instructions** On the right-hand side of the project, click **Set project instructions**. Paste the whole of the agent instructions — the kit page has a one-click **Copy instructions** button, and the same text is in the kit as `AGENTS.md` — then click **Save instructions**. This is the constitution: it pins the account ID, the currency, read-before-write, and create-paused. Every chat in the project inherits it. **3. Add the dossier as project knowledge** Still on the right, click **+** and upload `references/account-profile.md` and `PROMPTBOOK.md`. Claude uses them as context in every chat in the project. **4. Upload the skill (optional, but better)** Separately, upload the skill folder under [Customize → Skills](https://claude.ai/customize/connectors). Instructions are the rules, the skill is the expertise — keeping them in their own layers is what stops one from drowning out the other. **5. Run a prompt** Open a new chat **inside the project** and paste the first prompt from `PROMPTBOOK.md`. Work down the levels rather than cherry-picking — they're ordered so the agent builds context before it proposes changes. **Claude Cowork** Cowork projects are **desktop-only and stored locally**, which suits the playbook well — the folder you downloaded can *be* the project. **1. Create the project** Click **+** next to **Projects** in the left navigation. You get three options: start from scratch, import an existing Claude project, or **use an existing folder on your computer**. Pointing it at the unzipped playbook folder is the fastest route — the files are already laid out the way a project wants them. **2. Paste the agent instructions** Open the project's **Instructions** section — *"Add tone, formatting, or rules to help guide how Claude works on all tasks in the project"* — and paste the agent instructions (`AGENTS.md`, or the kit page's **Copy instructions** button) into it. **3. Add the account dossier to Context** Use the project's **Context** area to add the local playbook folder (or link the Claude chat project you already set up). That's where `account-profile.md` and `PROMPTBOOK.md` belong — data, not rules. **4. Add the skill** In **Customize → Skills**, add the playbook folder as a skill so it sits alongside the two skills the ADSMCP plugin installs. **5. Run a prompt, then schedule one** Work through `PROMPTBOOK.md` in the project. Once a prompt is proving itself by hand, move it to the **Scheduled** panel — see [scheduled tasks](https://docs.adsmcp.si/guides/scheduled-tasks.md) and [Cowork for Ad Ops](https://docs.adsmcp.si/guides/claude-cowork-ad-ops.md) for the unattended safety model. **ChatGPT** **1. Create the project** Create a Project for this ad account. One project per account — the agent should never have two accounts in scope at once. **2. Paste the agent instructions** Paste the agent instructions into the project's custom-instructions field. On Business, Enterprise, Healthcare, or Edu you can *also* upload the skill under **Skills → Create → Upload from your computer**. On Free, Plus, and Pro the instructions field is the whole mechanism — see the plan note above. **3. Upload the dossier and prompt book as project files** Add `references/account-profile.md` and `PROMPTBOOK.md` to the project's files so ChatGPT can read them on demand. **4. Run a prompt** Start a chat in the project and paste the first prompt from `PROMPTBOOK.md`. > [!TIP] > > **The split that makes this work:** instructions are *rules*, the dossier is *data*. Paste > `AGENTS.md` into the instructions field and upload `account-profile.md` as knowledge — > never the other way round. [Set up your project](https://docs.adsmcp.si/guides/set-up-your-project.md) explains why, and > covers memory, connectors, and scheduling around it. ## Keeping it current Ad accounts change. The playbook is a dated snapshot, and the agent is explicitly told that live data wins over anything written in the dossier. When the account has moved on, hit **Refresh** on the kit page. ADSMCP rebuilds against current data and gives you a new version to download. > [!TIP] > > **Refreshes are free and unlimited.** Rebuilding is included in the original \$39 — there is no > charge for a new version, ever. ## FAQ > **Is this a subscription? Will I be charged \$39 every month?** > > No. It is a single one-time charge for one ad account, and nothing recurs. If you took it as an > add-on during plan checkout, it appears on the same invoice as your subscription as a separate > one-time line called **"One-time Agent Setup — 1 Ad Account"** — that is why the total looked > higher that month. Your recurring plan price is unchanged and the \$39 will not appear again. > **Does it cover all my ad accounts?** > > No — one playbook covers one ad account, because it is generated from that account's own > campaigns and history. If you run Google Ads and Meta for the same brand, those are two > accounts and two playbooks. Buying several is one checkout and one charge, and **selecting > every eligible account takes 20% off the whole cart** — minimum two accounts, and it must be > all of the ones you can still buy. > **Where do I download it?** > > [adsmcp.ai/expert-space](https://adsmcp.si/dashboard) — accounts you own show **Open > Kit**, and the kit page has the **Download** button, a one-click copy for the agent > instructions, and the prompt book in a reader. The purchase email also links straight to it. > **Do I keep it if I cancel or downgrade my ADSMCP plan?** > > The downloaded files are yours — they're on your machine. What changes is that the tools the > playbook drives run on your ADSMCP plan, so the agent needs an active plan with quota to > actually do the work. > **Can I get a refund?** > > Each playbook is generated on demand from your account, so it is non-refundable once the build > has completed and the files are available to download. If the build failed, or you were charged > for an account you can't access, that's different — [contact support](https://docs.adsmcp.si/knowledge-base/support.md). > **Does it set up the ADSMCP connector for me?** > > No. Connecting ADSMCP to your AI client is free and covered in the > [client setup guides](https://docs.adsmcp.si/ai-clients/claude.md). The playbook starts after that: it teaches the agent > how to work *your* account well. > **I'm on an annual plan — do I still pay?** > > Annual plans include one playbook free. Claim it from > [Expert Space](https://adsmcp.si/dashboard); there's no checkout and no charge. Additional > accounts beyond the free one are \$39 each. > **What if my account data changes?** > > Hit **Refresh** on the kit page for a rebuild against current data. Refreshes are free and > unlimited. The agent is also instructed to always read live campaign data rather than trusting > the dossier's snapshot, so a slightly stale playbook still behaves correctly. > **Which AI clients does it work in?** > > All of them. The files follow the open Agent Skills and agents.md standards — Claude, Claude > Cowork, ChatGPT, Claude Code, Cursor, Codex, Windsurf, OpenClaw, and Antigravity. Only the > install shape differs: upload, paste, or drop the folder in. ## Related documentation > **[Agent Skills](https://docs.adsmcp.si/agent-skills/skills.md)** > > The 13 free skills every plan gets. The playbook builds on top of these. > **[Set Up Your Project](https://docs.adsmcp.si/guides/set-up-your-project.md)** > > Where the playbook's instructions and context belong in a project. > **[Pricing](https://docs.adsmcp.si/knowledge-base/pricing.md)** > > Plans, tool calls, and where the one-time add-on sits. > **[Support](https://docs.adsmcp.si/knowledge-base/support.md)** > > Billing questions, failed builds, and anything else. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.