> ## 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. # The Best Way to Set Up Your Project (Claude Cowork, ChatGPT, Claude Code, Codex) > The best way to set up an ad-ops project in Claude Cowork, ChatGPT, Claude Code, or Codex Desktop: one project per ad platform, agent instructions, agent context, scheduled tasks and routines, and agent skills — plus the ADSMCP playbook that fills them in for you. The best way to run ads with an AI assistant is not a chat. It is a **project**: a persistent workspace where the agent already knows your business, your accounts, your rules, and your goals before you type a word. Claude Cowork, ChatGPT, Claude Code, and Codex Desktop all have one. The names differ; the shape is the same. > [!NOTE] > > Not connected yet? Connect [ADSMCP](https://docs.adsmcp.si/mcp.md) in your client first — > [Claude Cowork](https://docs.adsmcp.si/ai-clients/claude-cowork.md), [ChatGPT](https://docs.adsmcp.si/ai-clients/chatgpt.md), > [Claude Code](https://docs.adsmcp.si/ai-clients/claude-code.md), [Codex Desktop](https://docs.adsmcp.si/ai-clients/codex-desktop.md) — then come > back here. A project is where the connection lives; it is not a substitute for it. ## Set it up in your client Pick your client and follow the numbered steps. Each one puts the same five things in place: one project per ad platform, agent instructions, agent context, scheduled tasks, and agent skills. [What each part is for](#what-goes-in-a-project) is explained further down. For the folder-based clients (Claude Code Desktop, Claude Code Terminal, Codex Desktop) the whole install is one move: download the kit `.zip` from [Expert Space](#skip-the-blank-page-buy-an-agent-playbook), attach it to a chat, and send the install prompt. The agent unzips it, puts the skill where it loads skills from, checks that ADSMCP is connected, and updates its own instructions and memory so every future session knows the expert exists. You never copy a file by hand. **Claude Cowork** Cowork has a dedicated field for every one of the five parts. Requires Claude Max, Team, or Enterprise, and the ADSMCP plugin added from **Customize → Plugins** (Anthropic Directory) ([setup](https://docs.adsmcp.si/ai-clients/claude-cowork.md)). **1. Open Cowork** In the Claude desktop app (or claude.ai), click the **Cowork** tab at the top. ![Claude desktop app with the Cowork tab selected](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/0GXeYWLZAEKikC6R/images/claude-cowork/cowork-tab.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=d8897a28c0eb6e2ce1d40d29303b4969) The Cowork tab in the Claude desktop app **2. Create one project per platform** Click **+** next to **Projects** in the left sidebar. Start from scratch or **use an existing folder on your computer**; pointing it at a folder that already holds your brand documents is the fastest route. Name it *"Brand — Google Ads"*. **3. Paste the agent instructions** Open the project's **Instructions** panel and paste your instructions: the [template below](#2-agent-instructions), or the kit's `AGENTS.md` (the kit page's **Copy instructions** button puts it on your clipboard). Every task and every scheduled run in the project inherits them. **4. Add context to the Context folder** Drop your business docs, exemplars, P\&L summary, marketing docs, past-test log, and the kit's `references/account-profile.md` into the project's **Context** folder. Run `/setup` once and the plugin scans the folder, pulls live campaign data, and writes a brand context file alongside them. Leave **Memory** on. **5. Upload the skill** Click **Customize** in the left sidebar, open **Skills**, and upload the kit `.zip` as-is. It sits next to the skills the ADSMCP plugin installed: those cover advertising in general, this one covers your account. **6. Confirm ADSMCP and a notification channel** In **Customize → Connectors**, confirm **ADSMCP Ads & Performance Marketing Agent** (the Anthropic-verified connector the plugin installs) shows **connected**, and link the ad accounts for this platform only. Set Interactive and Read-only tools to **Always allow** and leave Write/delete on **Needs approval**. Add Slack or email so scheduled runs can report. **7. Schedule the routines** In a chat inside the project, ask for the routine and Cowork offers to schedule it with `/schedule`. It appears in the **Scheduled** panel and runs remotely, laptop closed or not. Start with the weekly dashboard, read-only. The full Cowork build, including the seven-property safety model for self-optimizing tasks, is in [Cowork for Ad Ops](https://docs.adsmcp.si/guides/claude-cowork-ad-ops.md). **ChatGPT** Projects, files, plugins, and skills all live in ChatGPT's **Work** mode. Connect the ADSMCP app first ([setup](https://docs.adsmcp.si/ai-clients/chatgpt.md)); Projects are on Plus, Pro, Business, and Enterprise. **1. Switch to Work** At the top of ChatGPT, flip the toggle from **Chat** to **Work**. The composer gains three chips underneath it: **Project**, **Files**, and **Plugins**. Everything below happens here. ![ChatGPT with the Work toggle selected, showing the Project, Files, and Plugins chips under the composer](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/FUoqnFA1T6kM-gcS/images/set-up-your-project/chatgpt-work-mode.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=92277eb2dc6ab8a22d6546ae30f2376e) Work mode: the Project, Files, and Plugins chips under the composer **2. Create one project per platform** In the left sidebar, under **Projects**, create a new project and name it for the platform: *"Brand — Google Ads"*. Every chat you start inside it inherits the instructions, files, and plugins you set up next. ![ChatGPT sidebar with the Projects section expanded and a project listed](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/FUoqnFA1T6kM-gcS/images/set-up-your-project/chatgpt-projects-sidebar.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=f3928e5277d4de13ab9aba01610837be) Projects in the ChatGPT sidebar **3. Open Project settings** Open the project, click the **⋯** menu at the top right, and choose **Project settings**. ![ChatGPT project page with the three-dot menu open, showing Project settings and Pin project](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/FUoqnFA1T6kM-gcS/images/set-up-your-project/chatgpt-project-menu.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=0b991d554271b4516bdaa00696e03b0d) The ⋯ menu on a project page **4. Paste the instructions and scope the memory** Paste your agent instructions into **Instructions**: the [template below](#2-agent-instructions), or the kit page's **Copy instructions** button. This field is the whole mechanism on Plus and Pro, so make it complete: business, role, goal, scope, rules. Under **Memory**, switch from **Default memory** to project-only memory so what the agent learns about Google Ads never leaks into your Meta project, or into chats outside it. Leave **Library access** enabled if your brand documents already live in your ChatGPT file library. ![ChatGPT Project settings dialog with the Instructions field, a Memory dropdown set to Default memory, and Library access enabled](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/FUoqnFA1T6kM-gcS/images/set-up-your-project/chatgpt-project-settings.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=b60505e90f98503018d1a4ce7a8a0581) Project settings: Instructions, Memory, and Library access **5. Upload context under Files** Click the **Files** chip and add your business docs, exemplars, P\&L summary, marketing docs, past-test log, and the kit's `references/account-profile.md` and `PROMPTBOOK.md`. Keep them short; ChatGPT reads them on demand. **6. Turn on ADSMCP under Plugins** Click the **Plugins** chip and make sure your ADSMCP custom MCP server is on for this project. It appears under the name you gave it, e.g. **ADSMCP-Custom**. **7. Upload the skill** Open the **Plugins** page, switch the toggle at the top to **Skills**, click **+**, and choose **Upload from your computer**. Pick the kit `.zip`. ChatGPT scans it and the skill is available in every project. ![ChatGPT Skills page with the plus menu open, showing Create with chat, Create with editor, and Upload from your computer](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/FUoqnFA1T6kM-gcS/images/set-up-your-project/chatgpt-skills-upload.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=337fd0cf4354b538b652b5e1d03c8392) Skills: + → Upload from your computer Don't see **Skills**? Your plan doesn't include it yet. The instructions field and the ADSMCP plugin's built-in skills carry the same behavior; see the [plan note](https://docs.adsmcp.si/agent-skills/prompt-engineering-playbook.md#chatgpt). **8. Schedule the routines** In a chat inside the project, ask: *"Every Monday at 8am, build my dashboard…"*. ChatGPT creates a **scheduled task** that runs in the cloud and notifies you by push and email. Mention your ADSMCP plugin by name (e.g. **@ADSMCP-Custom**) if ChatGPT is slow to call tools on its own. > [!NOTE] > > Connect ADSMCP as a [custom MCP server](https://docs.adsmcp.si/ai-clients/chatgpt.md), not the ADSMCP plugin > from ChatGPT's directory. The plugin is having issues right now and has no Amazon Ads, > ChatGPT Ads, or Google Analytics tools. **Claude Code Desktop** Claude Code's desktop app works inside a folder. The folder *is* the project: instructions are `CLAUDE.md`, context is the files beside it, skills live in `.claude/skills/`, and routines run the folder on a schedule. Install the ADSMCP plugin first ([setup](https://docs.adsmcp.si/ai-clients/claude-code.md)). **1. Download the kit** Go to [adsmcp.ai/expert-space](https://adsmcp.si/dashboard), open the kit for this account, and click **Download kit (.zip)**. Don't unzip it. **2. Open one folder per platform** Create a folder per platform, *`acme-google-ads/`*, *`acme-meta-ads/`*, and open it in the Claude Code desktop app. Everything the agent knows about this platform lives in this folder and nowhere else. **3. Attach the zip and send the install prompt** Drag the `.zip` into the composer (or click **+** and attach it), then send: The install prompt. Claude Code unzips the kit, installs the skill, checks ADSMCP, and updates its own instructions and memory. Here are the agent skills I want to install. Unzip this kit and install it for this project: * Put the skill folder in .claude/skills/ and register the agent definition it ships in .claude/agents/. * Keep AGENTS.md inside the skill folder; don't overwrite anything at my project root. * Make sure the ADSMCP connector ([https://mcp.adsmcp.si/mcp](https://mcp.adsmcp.si/mcp)) is set up and active. If it isn't, install the adsmcp-ads-agent plugin and walk me through signing in. * Update this project's CLAUDE.md and your memory so every future session knows this expert exists, which ad account it covers, and to use it for any work on that account. * Verify every file against the zip, then tell me what you installed and where. Claude Code reports the skill path, the agent it registered, the ADSMCP connection status, and what it added to `CLAUDE.md`. Skills are picked up without a restart. **4. Put context in the folder** Add a `context/` subfolder with your business docs, exemplars, P\&L summary, marketing docs, and past-test log. Run `/adsmcp:refresh-brand-context` whenever the folder changes materially. Files on disk persist, so this is your memory too. **5. Create remote routines** Ask Claude Code to run the dashboard every Monday. Choose a **remote routine**, which runs on Anthropic's infrastructure with your laptop closed, over a local task, which silently skips if your machine is asleep. **Claude Code Terminal** Same folder, same files, same install prompt as the desktop app; the only difference is how you hand over the zip. Install the ADSMCP plugin first ([setup](https://docs.adsmcp.si/ai-clients/claude-code.md)). **1. Download the kit** Go to [adsmcp.ai/expert-space](https://adsmcp.si/dashboard), open the kit for this account, and click **Download kit (.zip)**. Don't unzip it. **2. Open the platform folder in Claude Code** ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}} theme={null} cd ~/Clients/acme-google-ads claude ``` One folder per platform. Everything the agent knows about Google Ads lives here. **3. Send the install prompt with the zip path** Drag the `.zip` from Finder into the terminal window to paste its path, then send: The install prompt. Replace the path with wherever the zip landed. Here are the agent skills I want to install: \~/Downloads/google-ads-expert-acme.zip. Unzip it and install it for this project: * Put the skill folder in .claude/skills/ and register the agent definition it ships in .claude/agents/. * Keep AGENTS.md inside the skill folder; don't overwrite anything at my project root. * Make sure the ADSMCP connector ([https://mcp.adsmcp.si/mcp](https://mcp.adsmcp.si/mcp)) is set up and active. If it isn't, install the adsmcp-ads-agent plugin and walk me through signing in. * Update this project's CLAUDE.md and your memory so every future session knows this expert exists, which ad account it covers, and to use it for any work on that account. * Verify every file against the zip, then tell me what you installed and where. Confirm with `/skills`, or invoke the expert directly as `/google-ads-expert-`. **4. Put context in the folder** Add a `context/` subfolder with your business docs, exemplars, P\&L summary, marketing docs, and past-test log, then run `/adsmcp:refresh-brand-context`. **5. Create remote routines** Ask for the weekly dashboard as a **remote routine**. A local task only runs while the machine is awake. **Codex Desktop** Codex inside the ChatGPT desktop app works the same way: a folder is the project, and you hand Codex the zip. Install the ADSMCP plugin with three chat messages first ([setup](https://docs.adsmcp.si/ai-clients/codex-desktop.md)). **1. Download the kit** Go to [adsmcp.ai/expert-space](https://adsmcp.si/dashboard), open the kit for this account, and click **Download kit (.zip)**. Don't unzip it. **2. Open one folder per platform** Switch to **Codex** in the top-left dropdown, click **New chat**, and when Codex asks where to work, pick or create a folder for this platform: *`acme-google-ads/`*. Keep other platforms in their own folders. ![ChatGPT desktop app with Codex selected, showing the Codex sidebar and a new chat asking what to work on](https://mintcdn.com/steven-00a2bcdd/ueaajNaB3rnf3mHW/images/My0vMqzyp3JsH-jn/images/codex-desktop/codex-home.png?fit=max\&auto=format\&n=ueaajNaB3rnf3mHW\&q=85\&s=1aa1ea9b9a89adf124b84f9ef41a1d35) Codex selected in the top-left dropdown, ready for a new chat **3. Attach the zip and send the install prompt** Click **+** in the composer (or drag the `.zip` in), then send: The install prompt. Codex unzips the kit, installs the skill and agent, checks ADSMCP, and updates AGENTS.md. Here are the agent skills I want to install. Unzip this kit and install it for this project: * Put the skill folder in .agents/skills/ and register the agent definition it ships in .codex/agents/. * Keep the kit's AGENTS.md inside the skill folder; don't replace the AGENTS.md at my project root. * Make sure the ADSMCP connector ([https://mcp.adsmcp.si/mcp](https://mcp.adsmcp.si/mcp)) is set up and active. If it isn't, install the adsmcp-ads-agent plugin and walk me through signing in. * Update this project's AGENTS.md (append, don't replace) and your memory so every future session knows this expert exists, which ad account it covers, and to use it for any work on that account. * Verify every file against the zip, then tell me what you installed and where. Codex ends with a link to the new agent. Click it, or start a **New chat** and ask for the expert by name. **4. Put context in the folder** Add your business docs, exemplars, P\&L summary, marketing docs, and past-test log to the folder. With **Computer use** on, Codex can also read a budget sheet in Excel or check something in Ads Manager that the API doesn't expose; leave Chrome on ask-each-time for ad platform sites. **5. Schedule the routines** Ask Codex to run the dashboard weekly and it sets up a scheduled task (Codex calls them automations) that runs in the cloud. For heavier autonomous setups, the [Codex terminal guide](https://docs.adsmcp.si/ai-clients/codex.md) covers agent configuration; your ADSMCP sign-in carries over. > [!NOTE] > > Using Claude chat projects on claude.ai rather than Cowork? The pattern is identical: **Set > project instructions**, upload context under project knowledge, upload the kit `.zip` under > **Customize → Skills**. The [playbook page](https://docs.adsmcp.si/agent-skills/prompt-engineering-playbook.md#set-up-the-project-then-run-the-prompts) > walks through it. ## What goes in a project A well-built project has five parts. Get all five in place and the agent stops being a clever chatbot and starts behaving like a media buyer who has been on the account for a year. | Part | What it is | What it does for you | | - | - | - | | **1. One project per ad platform** | A separate project for Google Ads, another for Meta, another for TikTok | The agent never mixes accounts, currencies, or platform rules | | **2. Agent instructions** | Who the agent is, what business it serves, what it optimizes for, what it must never do | Safe, on-goal behavior by default, in every chat | | **3. Agent context** | Your business documents, exemplars, P\&L, marketing docs, and what has and hasn't worked | Recommendations and copy grounded in your reality, not generic advice | | **4. Scheduled tasks and routines** | The same agent, run unattended on a cadence | Dashboards, pacing checks, and small approved changes happen without you asking | | **5. Agent skills** | Step-by-step playbooks the agent loads when a task matches | Campaign builds, audits, and optimizations done the right way, every time | > [!TIP] > > **Don't want to write any of this yourself?** The [agent playbook](#skip-the-blank-page-buy-an-agent-playbook) > from Expert Space is generated from your real ad account and ships every one of these parts as a > file you upload into the right place. The rest of this page explains what each part is for, so > you know where the files go and why. ## 1. One project per ad platform Create **one project for each ad platform you run**. Google Ads gets a project. Meta gets a project. TikTok, LinkedIn, and Amazon each get a project. Which AI client you use doesn't change this; it is the same rule in Claude Cowork, ChatGPT, Claude Code, and Codex. Why the boundary matters: * **Each platform has its own rules.** Google budgets are in dollars, Meta's are in the ad account's currency, TikTok rejects invalid objective and pixel combinations instead of correcting them. An agent scoped to one platform learns one rulebook and applies it correctly. * **Account IDs never cross.** The single worst ad-ops mistake is a change landing in the wrong account. A project pinned to one platform, with its account IDs written into the instructions, makes that structurally impossible. * **Context stays relevant.** Your Meta exemplars are noise in a Google Search project. Splitting by platform keeps every file in the project useful to every task. * **Automation stays safe.** Scheduled tasks inherit the project. One project per platform means one writer per portfolio, so two routines can never fight over the same campaigns. Name projects so the platform is obvious from the sidebar: *"Acme — Google Ads"*, *"Acme — Meta Ads"*. Agencies add the client: *"Acme — Google Ads"*, *"Bolt Coffee — Google Ads"*. One project per client per platform is more projects, and that is the point. See [managing multiple clients](https://docs.adsmcp.si/knowledge-base/multi-client.md) for the agency pattern in full. Inside each project, connect ADSMCP and link only the ad accounts that belong there. Add a notification channel (Slack or email) if your client supports one, so scheduled runs have somewhere to report. ## 2. Agent instructions Instructions are the agent's constitution: read on every task, always on. They are the difference between an assistant that asks "which account?" every session and one that already knows. Write them in five short sections: | Section | What to write | Example | | - | - | - | | **The business** | What you sell, to whom, at what price point, and what a customer is worth | *"We sell a $49/mo inventory tool to independent restaurants in the US. A trial is worth ~$180 in LTV."* | | **The agent's role** | What job it holds on the team and how much it decides on its own | *"You are the paid-media agent running Google Ads. You analyze, propose, and build on approval."* | | **The goal** | The one metric it optimizes toward, with the number | *"Target CPA \$45 on trial signups. Scale spend while CPA holds under target."* | | **The scope** | The exact platform and account IDs it may touch | *"Google Ads account 123-456-7890 only. Stop if a tool result mentions any other account."* | | **The golden rules** | What it must never do without you | Create paused, never launch. Money needs an explicit yes. Read before write. Never invent a number. | Then set the agent's **default mode** to how you actually work: * **Builder** — you hand it approved briefs and want campaigns built, verified, and monitored. Copy stays locked; it validates and reports, never rewrites. * **Advisor** — you want the portfolio truth and a recommendation. It analyzes and proposes, builds only when asked, and every budget call stays yours. > **Copy-paste: starter agent instructions** > > ```markdown theme={"theme":{"light":"github-light","dark":"github-dark"}} theme={null} > # Agent instructions — Google Ads paid-media agent (ADSMCP) > > ## The business > Seasonality, promotions, and anything a new hire would need to know on day one.> > > ## Your role > You are the paid-media agent for 's Google Ads account, working through the > ADSMCP connector. You analyze, propose, and build on approval. Humans decide WHAT > to run; you build it exactly as specified, verify it, and flag anything that isn't right. > > ## What we're optimizing for > - North-star metric: > - A conversion is: > - Secondary signals to report, never to optimize toward: > > ## Scope > - Google Ads account only. If a tool result references any other account, stop > and tell me before continuing. > - Budget units: Google = dollars. Report spend in USD. > > ## Golden rules (never break) > 1. Nothing goes live without a human. Create every campaign PAUSED. Never enable or launch. > 2. Money needs a yes. No budget raise, bid change, or resume without explicit approval. > 3. Read before you write; verify after. Report what exists, not what you asked for. > 4. Validate, don't edit. Copy over a platform limit? Stop and report the field and length. > 5. Never invent a number. If a tool fails or data is missing, say so. > 6. Pause, don't delete. > > ## Context you should read > - `context/` holds our brand guide, exemplars, P&L summary, and past-test log. Read it > before proposing creative or budget changes. Live account data wins over any file. > ``` **The rule that keeps instructions useful: rules go in instructions, data goes in context.** Don't paste your brand guide into the instructions field. Don't put "never launch" in a document the agent might skip. Procedures and non-negotiables live here; everything the agent should *read* lives in the next part. For the full always-on version, with the standard operating loop and the interactive-versus-scheduled monitoring split, see [Cowork for Ad Ops](https://docs.adsmcp.si/guides/claude-cowork-ad-ops.md). ## 3. Agent context Context is the data that makes the agent sharp. An agent with instructions but no context knows the rules and nothing about you. Give it what you'd give a senior hire in their first week. | Upload this | Why the agent needs it | | - | - | | **Business documents** — positioning, offer sheet, pricing, ICP, sales deck | So ad copy says what you actually sell, to the people who actually buy | | **Exemplars** — your best-performing ads, landing pages, emails, and the ones that flopped | So generated copy sounds like your winners and avoids your losers | | **P\&L or margin data** — contribution margin by product line, CAC targets, payback window | So "scale this" is judged against real economics, not just a CPA | | **Marketing documents** — brand guide, voice, messaging pillars, seasonal calendar, promo schedule | So copy passes brand review and campaigns line up with launches | | **What has and hasn't worked** — a log of past tests, structures you abandoned, audiences that never converted | So the agent doesn't re-propose something you already killed | | **Naming conventions** — campaign, ad group, and UTM structure | So new builds slot into your reporting without cleanup | | **Audience and exclusion lists** — negative keywords, excluded placements, suppression lists | So known waste never comes back | | **A recent performance export** | A baseline the agent can compare against before its first live pull | Keep the files short and current. A two-page brand summary beats a forty-page deck. Date anything that goes stale, and tell the agent in its instructions that live account data wins over any file. > [!TIP] > > **A brand folder is the fastest way to give context.** Claude Cowork, Claude Code, and Codex > all work inside a folder on your computer. Put the documents above in it and the agent reads > them; run `/setup` (Cowork, Claude Code) or ask Codex to set up the workspace, and the ADSMCP > plugin scans the folder, pulls live campaign data, and writes a context file for you. **Leave memory on** where your client offers it. Project memory lets the agent carry naming decisions, prior builds, and your preferences across sessions, and it is scoped to the project so one platform's memories never leak into another's. ## 4. Scheduled tasks and routines A project pays for itself when the agent runs without you. The same instructions, context, and skills, on a cadence, reporting to a channel you watch. This is what the agent can do automatically: | Routine | Cadence | What it does | Writes to the account? | | - | - | - | - | | **Performance dashboard** | Daily or weekly | Builds a scorecard: spend, conversions, CPA (Cost per acquisition), ROAS (Return on ad spend) per campaign, plus the three things worth your attention. In Claude it can publish as a live artifact; in ChatGPT as a hosted page. | No | | **Pacing check** | Weekly and mid-month | Month-to-date spend versus budget, projected end-of-month, campaigns about to blow their cap | No | | **CPA / spend alert** | Daily | Stays silent unless a campaign crosses your target, then tells you which one and by how much | No | | **Search-term and placement hygiene** | Weekly | Proposes negatives and exclusions from last week's waste | Proposes only, or applies within a cap you set | | **Small optimizations** | Weekly | Budget shifts of a few percent between campaigns, bid adjustments inside a range, pausing ads with zero conversions after a threshold | Yes, within hard caps, with a report every run | | **Competitor watch** | Weekly | Checks competitor landing pages for offer, price, or hero changes | No | Start every routine **read-only**. Let it propose for a few weeks, check its proposals against what you'd have done, and only then let it apply small, reversible changes with a per-run cap. A routine that changes budgets with nobody watching spends real money; the [safety model for unattended changes](https://docs.adsmcp.si/guides/claude-cowork-ad-ops.md#the-safety-model-for-unattended-changes) covers shadow mode, change caps, single writer, and the kill switch. What each client calls it: | Client | Feature | Runs when your computer is off? | | - | - | - | | **Claude Cowork** | Scheduled tasks, via the `/schedule` skill | Yes | | **ChatGPT** | Scheduled tasks | Yes | | **Claude Code** | Routines: a *local* task or a *remote* routine | Local: no. Remote: yes | | **Codex** | Scheduled tasks (automations) | Yes | > [!WARNING] > > A Claude Code **local** routine only runs while the desktop app is open and your computer is > awake. For anything that matters, pacing, spend spikes, disapprovals, create a **remote > routine** instead. Cowork and ChatGPT scheduled tasks don't have this problem. A weekly dashboard routine. Paste it as a scheduled task in any client. Every Monday at 8am, build my Google Ads dashboard for last week: spend, conversions, CPA, and ROAS per campaign, week-over-week change, and month-to-date pacing against budget. Lead with the three things worth my attention. Read-only: propose changes, don't make any. If it runs late in the day, summarize what changed instead of starting fresh. A small-changes routine with hard caps. Only after weeks of read-only runs. Every Wednesday at 7am, review last 7 days of search terms across active campaigns. Add negatives for terms with 20+ clicks and zero conversions. Shift up to 10% of daily budget from campaigns above target CPA to campaigns below it, never more than \$50 per campaign per run. Never pause a campaign, never touch bid strategy. Report every change you made and every change you chose not to make. Every scheduled run calls ADSMCP tools again, so it draws on your monthly tool-call quota like a live conversation. Weekly is the right default for a review; daily is for accounts spending enough to justify it. The [scheduled tasks guide](https://docs.adsmcp.si/guides/scheduled-tasks.md) has more example prompts and the late-run guardrails that keep a routine honest. ## 5. Agent skills A skill is a step-by-step playbook for one job, loaded only when a task matches. Instructions say *what the agent may do*; a skill says *exactly how to do it*. That split is what lets you give the agent deep procedures without bloating every chat. ADSMCP ships **13 free skills** with every plan: the safety contract, the tool-call contract, one skill per platform, and cross-platform workflows for launching, reviewing, optimizing, and writing creative. They install with the ADSMCP plugin in Cowork, Claude Code, and Codex, and ship inside the ChatGPT plugin. You never invoke them by name; describe the task and the agent picks the skill. | You say | Skill that loads | What it enforces | | - | - | - | | "How did we do last week?" | `adsmcp-performance-review` | Cross-platform scorecard, tracking audit, read-only | | "Launch a search campaign for the spring promo" | `adsmcp-launch` → `adsmcp-google-ads` | Keyword research first, platform field rules, created paused, verified after | | "Find wasted spend" | `adsmcp-optimize` | Search terms, placements, pacing, reallocation proposals | | "Write three headlines in our voice" | `adsmcp-creative` | Reads your brand context, respects character limits | Skills sit in a different place from instructions in every client: the **Skills** panel in Cowork and ChatGPT, a `.claude/skills/` folder in Claude Code, `.agents/skills/` in Codex. Keep them there. Don't paste a skill into the instructions field, and don't drop it into context as a document. Reference skills *by name* from instructions and scheduled-task prompts when you want to force one. The free skills know advertising. They don't know your account. The [agent playbook](#skip-the-blank-page-buy-an-agent-playbook) below adds a skill that does. Full reference: [Agent Skills](https://docs.adsmcp.si/agent-skills/skills.md). ## Where each part lives, by client | Part | Claude Cowork | ChatGPT | Claude Code (desktop or terminal) | Codex Desktop | | - | - | - | - | - | | **Project per platform** | Projects sidebar, one per platform | Projects sidebar, one per platform | One folder per platform | One folder per platform | | **Agent instructions** | **Instructions** panel | Project **Instructions** | `CLAUDE.md` | `AGENTS.md` | | **Agent context** | **Context** folder + Memory | Project **Files** + Memory | Files in the folder | Files in the folder | | **Scheduled tasks** | **Scheduled** panel, via `/schedule` | Scheduled tasks | Routines (choose remote) | Scheduled tasks (automations) | | **Agent skills** | **Customize → Skills** | **Skills** (Business/Enterprise/Edu) | `.claude/skills/` | `.agents/skills/` | ## Skip the blank page: buy an agent playbook You don't have to write any of this from scratch. ADSMCP generates an **agent playbook** for any ad account you've connected, built from that account's real campaigns, spend history, and current state. It ships as a download containing every part of the project above, ready to upload into the right place. **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, which is also why the one-project-per-platform rule fits it: one playbook covers one account. **2. Preview, then buy** **Preview Expert** shows sample prompts before you pay. Buy one account, or tick several and check out together. It's a one-time purchase per account, and annual plans include one free; [pricing and the bundle discount](https://docs.adsmcp.si/agent-skills/prompt-engineering-playbook.md#price) are on the playbook page. **3. Download the kit** The build takes a few minutes. From the kit page, click **Download kit (.zip)**, or hit **Copy instructions** to put the agent instructions straight on your clipboard. Unzip it and you get one folder, named like `google-ads-expert-/`. **4. Upload each file into its part of the project** The kit is laid out to match this page. Use the table below. What's in the download, and where it goes: | File in the kit | Part of the project | Where to put it | | - | - | - | | `AGENTS.md` | **Agent instructions** — the account-pinned constitution: business scope, role, account ID, currency, golden rules | Paste into the Instructions field (Cowork, ChatGPT). Folder clients keep it inside the skill; Codex installs it for you. | | `references/account-profile.md` | **Agent context** — a written dossier of the account: what state it's in, operating priorities, weekly and monthly cadence | Context folder (Cowork), project Files (ChatGPT), the project folder (Claude Code, Codex) | | `references/working-with-adsmcp.md` | **Agent context** — the tool-calling contract and guardrails | Stays inside the skill folder; loads on demand | | `SKILL.md` | **Agent skill** — the operating doctrine for this one account: wasted-spend triage, tracking audit, pacing, bidding, testing discipline | **Customize → Skills** (Cowork, Claude), **Skills** (ChatGPT Business+), `.claude/skills/` (Claude Code), hand the folder to Codex | | `PROMPTBOOK.md` | **Scheduled tasks and routines** — a menu of prompts graded Beginner → Expert, written against your account. Its daily-ops and analysis prompts are your first routines. | Read it; copy a prompt into a scheduled task once it has proven itself in a live chat | | `README.md` | Says where each piece goes, per client | Read once | > [!TIP] > > The playbook follows the rule this page is built on without you thinking about it: `AGENTS.md` > is rules, `account-profile.md` is data, `SKILL.md` is procedure, and they arrive as separate files > so they land in separate places. The dossier is a dated snapshot; the agent is told live account > data always wins, and rebuilding the kit against fresh data is free and unlimited from > **Refresh** on the kit page. The full install steps for every client, what each build stage does, and the FAQ are on the [Prompt Engineering Playbook](https://docs.adsmcp.si/agent-skills/prompt-engineering-playbook.md) page. ## Checklist Before you call the project done: * One project per ad platform, named so the platform is obvious, with only that platform's accounts linked * Instructions cover the business, the agent's role, the north-star metric, the exact account IDs, and the golden rules * Context holds your brand guide, exemplars, margin data, marketing docs, and what has and hasn't worked * ADSMCP connected inside the project, with a notification channel for scheduled runs * Skills installed in the Skills panel or skills folder, not pasted into instructions * At least one read-only routine running: the weekly dashboard * Memory on Then open a chat inside the project and ask: The first prompt in any new project. The agent reads its instructions, context, and the live account, then tells you what it sees. Introduce yourself and give me your read on the account. ## Where to go next > **[Scheduled tasks](https://docs.adsmcp.si/guides/scheduled-tasks.md)** > > More routine prompts, late-run guardrails, and what a cadence costs in tool calls. > **[Cowork for Ad Ops](https://docs.adsmcp.si/guides/claude-cowork-ad-ops.md)** > > The full always-on build and the safety model for routines that make changes. > **[Agent Skills](https://docs.adsmcp.si/agent-skills/skills.md)** > > The 13 free skills, what each enforces, and how to install them per client. > **[Prompt Engineering Playbook](https://docs.adsmcp.si/agent-skills/prompt-engineering-playbook.md)** > > The agent playbook from Expert Space: instructions, dossier, skill, and prompt book generated from your account. > **[Managing multiple clients](https://docs.adsmcp.si/knowledge-base/multi-client.md)** > > One project per client per platform, without letting automation collide. > **[Connect your AI client](https://docs.adsmcp.si/mcp.md)** > > Cowork, ChatGPT, Claude Code, Codex Desktop, and every other client ADSMCP supports. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.