- Library
- Edition 2.0
- AI Media Buying
Edition 2.0
AI media buying with the controls left in.
How to use Claude for research, diagnosis, reporting and creative operations on Meta, and how to let it propose changes without ever handing it the budget.
For media buyers, performance marketers and agency teams. About 180 minutes across 14 chapters. Facts checked .
Replaces the AI Media Buying Framework (V1), which is kept online for people arriving from older videos.
Get the full guide
Free. 14 chapters, about 180 minutes. One sign-up opens all three Edition 2.0 guides on this device.
Chapters
14 chapters. Take them in order; each one ends with a task and a checklist.
- 01
What changed since V1
Meta automated the campaign structure, removed two attribution windows and shipped its own AI connectors. Find the V1 habits that now break, and see which chapter teaches each job you want Claude to do.
10 minDone - 02
Read, propose, approve, execute
The operating model behind every workflow in this guide. Five separate authorities, what each boundary must enforce, three working modes, labelled evidence, and the decision-rights card that fixes who may approve what.
13 minDone - 03
Set up a safe Claude workspace
Pick a Claude surface, install and verify Claude Code, learn the five terminal commands you need, then build a media-buying project folder with secrets blocked, project memory in CLAUDE.md and a first task on synthetic data.
12 minDone - 04
Data, attribution and evidence
Get data to Claude in a form it can trust. Four ways in, a safe export procedure, the attribution facts that changed in 2025 and 2026, a data contract, tracking health, what never goes to a model, and how to keep hypotheses separate from causes.
14 minDone - 05
Daily diagnosis, monitoring and reports
A morning routine in a fixed order. Signal health, the morning diagnosis, four monitor watches, alerts that run inside Meta, anomaly triage, and daily, weekly and monthly report templates, all without giving Claude write access.
15 minDone - 06
Budgets, tests and scaling proposals
Replace fixed ROAS rules with break-even and target numbers from your own margins, then run the weekly budget review, draft a monthly allocation, plan tests, choose a scaling move and read market signals, all as proposals with a rollback rule.
15 minDone - 07
Creative operations
Creative is the input you still control. Turn customer language into themes and store fixes, keep a claims library, write copy by angle, stage and placement, brief winner variations, take assets from a folder to a named and mapped plan, and stay inside Meta's AI disclosure rules.
14 minDone - 08
Launches, destinations and stock
Take a brief to a paused draft that passes QA, launch a new product with alerts set in advance, check every destination that receives paid traffic, and keep ads away from products that cannot sell.
14 minDone - 09
Skills and SOPs that hold
Turn your SOPs into Claude skills that make good work repeatable. Where skill files live and how they load, a full template, V1's budget optimiser rebuilt as a skill, an evaluation set to test it, chaining without chaining trust, and the controls a skill cannot replace.
15 minDone - 10
APIs, docs and version changes
What an API is and is not, how to tell a read from a write, how to read a Meta Marketing API reference page, V1's campaign example corrected, five error categories with the right response to each, and the protocol to follow when Meta changes the API.
11 minDone - 11
Development versus production credentials
Set up a Meta development app and test account that cannot spend real money, choose the right credential for development, staging and production, decide where each one may live, read store data safely, and check every token before you trust it.
15 minDone - 12
The approval machine
The controlled agentic pattern end to end. Claude returns a schema-checked proposal, a person approves its exact hash, and a small executor applies only that, within caps, paused by default, with a journal and a reconciliation pass.
15 minDone - 13
Security, policy and incidents
The threats that actually hit AI media-buying setups and the control that stops each one, prompt injection, connectors, browser agents, client separation, the Meta and Anthropic policies you operate under, and a step-by-step incident playbook.
10 minDone - 14
Rollout, operations and readiness
A staged path from synthetic data to production, the evidence each stage must produce, a first-month plan, the metrics that show whether the system itself is healthy, the re-verification calendar, and the final readiness check.
9 minDone
What changed since V1
- Meta now runs most delivery decisions itself. The AI’s job moved from tweaking bids to diagnosing, briefing creative and proposing changes.
- Credentials are split by environment: short-lived tokens for testing, system users for production, and no token ever inside the model’s context.
- Official connectors and Agent Skills replace copy-pasted scripts and prompt-only guardrails.
- Every write goes through a proposal, a human approval and a deterministic executor that creates objects paused.
Review status
Beta Available for testing; behavior and interface may change. Next review due .
- Owner review of every chapter before the status moves from beta to active
- Re-verify Meta API version, token and access-tier facts at each quarterly Marketing API release
- Replace synthetic examples with account-calibrated thresholds before any implementation
Primary sources used in this guide
Official documentation and changelogs only. When a source and this guide disagree, the source wins.
- Marketing API versions Meta
- Introducing Graph API v26 and Marketing API v26 Meta
- Marketing API v25.0 changelog Meta
- Advantage+ campaigns Meta
- Out-of-cycle changes 2025 Meta
- Marketing API v26.0 changelog Meta
- Meta ads AI connectors Meta
- Advertising Week New York 2026 Meta
- Install apps and generate system user tokens Meta
- Claude connectors directory Anthropic
- Claude Code security Anthropic
- Marketing API overview Meta
- Structured outputs Anthropic
- Claude Code advanced setup Anthropic
- Terminal guide (Claude reads images in your folder) Anthropic
- Claude Code overview (scheduled tasks) Anthropic
- How Claude remembers your project Anthropic
- Configure permissions Anthropic
- Claude Code sandboxing Anthropic
- Claude Code permission modes Anthropic
- Claude Developer Platform workspaces Anthropic
- Ad account insights reference Meta
- Insights limits and best practices Meta
- Out-of-cycle changes 2026 Meta
- Ad set reference (learning_stage_info) Meta
- Split testing Meta
- Conversions API Meta
- Deduplicate Pixel and server events Meta
- Meta Business Tools Terms Meta
- Custom conversion reference Meta
- Meta Platform Terms and Developer Policies Meta
- Ad Rules Engine Meta
- Ad rules execution spec Meta
- Marketing API rate limiting Meta
- Pricing (Batch API discount) Anthropic
- Budgets Meta
- Advantage campaign budget Meta
- About daily budgets Meta Business Help Center
- Currencies and offsets Meta
- Ads MCP server A/B tests and conversion lift studies Meta
- Ad Library API (ads_archive) Meta
- Automated Data Collection Terms Meta
- URL tags for tracking Meta
- Generative AI transparency in Meta's ads products Meta
- Meta prepares for the 2026 US midterms Meta
- Ads MCP server overview Meta
- Anthropic Usage Policy Anthropic
- Ad campaign reference Meta
- Ads MCP server ad creation and management tools Meta
- Product sets reference Meta
- Advantage+ catalog ads Meta
- Shopify access scopes Shopify
- Playwright MCP Microsoft
- Extend Claude with skills Anthropic
- Hooks reference Anthropic
- Claude Code MCP configuration and scopes Anthropic
- Define success criteria and build evaluations Anthropic
- Ads MCP server rules best practices Meta
- Marketing API error reference Meta
- Graph API error handling Meta
- Marketing API versioning Meta
- Meta Node.js Business SDK releases Meta
- Marketing API get started Meta
- Access tokens Meta
- Get long-lived tokens Meta
- System users Meta
- Securing requests (appsecret_proof) Meta
- debug_token reference Meta
- Marketing API authorization and access tiers Meta
- Marketing API changelog (access tier rename) Meta
- Ads MCP server, get started Meta
- Authorizing access for AI tools Shopify
- Shopify API versioning Shopify
- Shopify client credentials grant Shopify
- Models overview Anthropic
- Using Claude in Chrome safely Anthropic
- Custom connectors (unlisted) Anthropic
- MCP authorization specification 2025-11-25 Model Context Protocol
- Meta Terms of Service Meta
Education, not legal, security or financial advice. Never paste a live access token, API key or customer data into an AI chat, a prompt, a screenshot or a public repository.