Meta Ads

Purpose

Use this skill for any Meta Ads read or write. The live catalogue covers accounts, campaigns, reporting, audiences, datasets, catalogs, experiments, policy, and help. A bare request to create or design an ad enters the complete campaign controller unless the advertiser explicitly asks only for a standalone image asset or organic post. The controller owns full-research versus explicit guided planning and every approval boundary; do not restate that flow here.

Ad spend belongs here. "How much did I spend", "what did it cost", "where did the budget go" and "break down my spend" are Meta Ads questions whenever an ad account, campaign, ad set or ad is in play — route them to this skill, not to a bank or card connector, which sees money leaving an account and knows nothing about delivery. Only reach for a financial connector when the user is asking about their bank, card, or business finances rather than their advertising. Discover availability with list-tools --names-only, inspect selected tools with describe-tool, and invoke them with call-tool, including ads_creative_upload_media for new media.

Spend is visible here; billing is not. amount_spent is delivery spend for the entity level and window you queried (last_28d when you pass neither date_preset nor time_range), so name both whenever you quote it, and call maximum "all time", never "since your last bill". It is not the figure on an invoice or on the account spending limit. No Ads tool returns the balance owed, prepaid funds, the account spending limit, payments, invoices, or ad credits: not when a credit was granted or applied, how much of it is used, or what is left. Never derive any of these from spend, a budget, or a figure the user gives you. Tell the user only that you can't see billing or credit details from here, never which tools or data you lack, and point them to Billing & Payments in Ads Manager. This is the one case where Ads Manager is the answer, because no tool can return it.

Policy questions: mandatory route

Treat any question about what a Meta, Facebook, or Instagram advertising policy means, allows, prohibits, or requires as a Meta Ads task, even when it names no account, campaign, or ad. Load references/policy.md, pass the user's policy question to ads_policy_tool, and call it before answering. Browser search, a public policy page, and model memory are not substitutes for the canonical tool result. If the tool is unavailable, still returns N/A after the one catalogue-title retry in references/policy.md, or returns a policy that does not cover the question, say the policy could not be confirmed instead of answering from another source.

Resolve the owning system before routing

Route on the object and its context, not on a generic noun. catalog, feed, product, and audience do not by themselves mean Meta Ads. They may belong to a storefront, a content feed, another commerce system, or Meta Ads.

Treat the request as Meta Ads when at least one of these establishes ownership:

• the user says Meta Ads, Facebook/Instagram advertising, Ads Manager, Commerce Manager, an ad account, campaign, ad set, ad, or dynamic ads;
• the object is Ads-specific, such as a Meta Pixel/dataset, Custom Audience, product set used for ads, or an Ads delivery setting; or
• the current conversation or a read-only Meta Ads lookup has already resolved the named object or id inside the user's Meta Ads account.

A specific catalog or product-feed name or id must load this skill first for read-only ownership resolution, even when the user did not say Meta Ads. The same is true for a specifically named feed when the request concerns its upload or refresh schedule, for example “change my Shopify nightly feed to refresh hourly.” Use the required Ads list/read chain to look for that exact object before searching cron jobs, hooks, tracking items, reminders, or local scripts; those surfaces are not substitutes for an Ads product-feed lookup. One unique match establishes Ads ownership; only then may an Ads write be proposed. If no object matches, or more than one object could be the target, ask one short disambiguating question and do not write. A platform-like word in a name, such as Shopify, is part of the supplied name and does not prove the feed belongs to that external platform.

When only a generic noun is present and ownership is still unclear, ask one short disambiguating question, such as whether they mean their Meta Ads product feed or another feed. Do not mutate either system while it is ambiguous. After an object is resolved to Meta Ads, do not hand it to a generic Feed or commerce tool, and do not silently fall back to another system when an Ads capability is unavailable.

References

Advertising is a domain where a fluent, confident, wrong answer costs the advertiser money. These files carry the rules that stop that. Read the ones the task touches before answering, not after drafting.

Read this	When
references/safety.md	Always. Ten HARD boundaries: audience guidance, PII, proposal vs execution, Special Ad Category, policy sourcing.
references/account-scope.md	Any question about an account, catalog, audience, feed, or experiment — which is nearly all of them. Especially when the question names none: deciding whether to ask, or to carry one forward from an earlier turn, is what this file governs.
references/evidence.md	Any answer that reports a number, or that has to describe something missing.
references/analysis.md	Why, trend, ranking, comparison, or recommendations based on existing delivery performance.
references/response-style.md	Reporting, analysis, or metric-definition answers. Campaign-stage references own campaign presentation.
references/tool-routing.md	Reads not already routed by the active campaign-stage reference, especially catalogs, datasets, audiences, experiments, and analysis levels.
references/policy.md	Before any ads_policy_tool call, any question about what Meta's ad policies allow or prohibit, and any general advertising how-to.
references/campaign-creation.md	Start here for a complete new campaign. This short controller identifies the first incomplete stage; do not preload its later-stage references.
references/campaign-planning.md	Only while a complete campaign still needs identity resolution, a product brief, research, recommendations, hierarchy, full-plan approval, or an optional strategy artifact.
references/campaign-guided.md	Only when the advertiser explicitly requests step-by-step planning.
references/campaign-delivery-compatibility.md	While settling objective/optimization after destination and tracking are known, again on exact arguments before final review, and before adding an ad set to an existing campaign.
references/campaign-targeting.md	Only when campaign planning must resolve a non-country place, interest, or language into a canonical targeting object.
references/campaign-budget.md	Only when a proposed campaign is ready for budget pricing or its pricing inputs changed; full research loads it from planning after those inputs stabilize.
references/campaign-creative.md	Only when preparing or approving any campaign creative, including image, video, carousel, boosted-post, and partnership-ad formats.
references/campaign-execution.md	Only when preparing the final review, collecting its exact approval, creating the paused hierarchy, presenting its immediate handoff, or recovering partial creation.
references/campaign-handoff.md	Only after the advertiser selects a post-create delivery or editing action.
references/campaign-manual-setup.md	When an account is listed with is_ads_mcp_write_disabled: true, when a write was rejected as not available for this ad account — by a tool result, or quoted by the advertiser from an earlier attempt — or the advertiser asks to set the campaign up themselves in Ads Manager. Read it before explaining that rejection.
references/writes.md	Before executing a standalone create, update, activate, pause, delete, connect, or upload, and before saying whether a pixel, dataset, or Conversions API setup step can be done here. Complete campaigns load it only when their staged references direct.

Tooling

Use exec to run the installed binary directly:

/opt/hatch/bin/meta-ads-cli <subcommand> [options]

HARD invocation boundary: run every Meta Ads command as its own exec tool call. Never combine Meta Ads commands with a newline, &&, ;, a pipe, or a wrapper such as cd, env, timeout, or bash. That includes | head and | python3: results over 128 KiB already come back as an output_file (below), and describe-tool --input-only returns just the schema, so there is nothing to truncate or parse inline. In particular, the complete command string for the first protected account lookup must be exactly:

/opt/hatch/bin/meta-ads-cli call-tool --name ads_get_ad_accounts

The runtime directly spawns and attests that installed binary. A combined or wrapped command requires a shell and intentionally cannot open the trusted professional-consent presentation.

Connection management and diagnostics

/opt/hatch/bin/meta-ads-cli status
/opt/hatch/bin/meta-ads-cli disconnect-url

status currently includes the full tool catalogue. Reserve it for connection diagnostics; do not use it for normal capability discovery.

When the user asks to disconnect, disable, or turn off Meta Business access, run disconnect-url as its own exec call. When the response contains disconnect_url, share exactly [Disconnect Meta Business Manager](<disconnect_url>), without also pasting the raw URL. The link opens Hatch's standard connector disconnect confirmation; do not claim access was removed until the user confirms there. Never use the retired disconnect command, which belonged to connector OAuth.

MCP operations

/opt/hatch/bin/meta-ads-cli list-tools --names-only
/opt/hatch/bin/meta-ads-cli describe-tool --name <tool>
/opt/hatch/bin/meta-ads-cli describe-tool --name <tool> --input-only
/opt/hatch/bin/meta-ads-cli call-tool --name <tool> --agent-output --arguments-json '<json-object>'
/opt/hatch/bin/meta-ads-cli estimate-budget <typed plan flags>
/opt/hatch/bin/meta-ads-cli render-campaign-summary --summary-json '<json-object>'
/opt/hatch/bin/meta-ads-cli render-campaign-success --success-json '<json-object>'
/opt/hatch/bin/meta-ads-cli render-chart --chart-json '<json-object>'

Before any estimate-budget invocation, read references/campaign-budget.md; its typed flag grammar is complete. Never substitute --help or raw MCP-schema arguments.

The three render commands are local and credential-free. Pass each returned widget.kind and widget.data to widget.create unchanged; never replace the payload with hand-authored HTML or a placeholder. render-chart returns an html_file card plus a full_view.path, not a link to include in prose. --chart-json accepts exactly these keys and rejects any other: {"type":"line"|"bar","title":"…","metric":"…","unit":"currency"|"percent"|"number","currency":"USD","x_labels":["…"],"series":[{"name":"…","values":[…]}],"reference":{"label":"…","value":…}}; currency (the account's ISO code) is required with unit: "currency" and rejected with any other unit; reference is optional, and name is required when there are two or more series.

Status, discovery, and tool-call responses over 128 KiB return output_file, output_bytes, and output_format: "json" instead of the inline response. output_file contains the complete original JSON. Read or query that file in a separate tool call, selecting only the fields or array slice needed; do not print the entire file back through exec. Small responses keep their usual shape. The file is temporary and need not survive a restart.

ok still describes the Ads operation. If output_available: false is returned, output delivery failed even if the operation succeeded. Do not repeat an Ads write solely to recover its result; check the resulting Ads state first.

Placement-specific media generation

For campaign images, follow references/campaign-creative.md and run the Ads-owned wrapper as a standalone command:

/opt/hatch/bin/meta-ads-cli creative generate-image --placement <feed|story|reel> --prompt '<text>' --output-dir workspace/your_files

The wrapper owns placement ratios and output validation. Repeat --prompt for ordered constraints; add --source-image <path> only to adapt that accepted image. Use local_path for review and retain media_handle when returned. After approval, upload with that handle, or the exact local_path as file when no handle exists, then use only the returned Ads reference.

Campaign video comes from media.generate_video; it carries no placement ratio, so references/campaign-creative.md covers fitting its shape.

New media reaches Ads only through ads_creative_upload_media. Asset selection, source precedence, returned references, and approval ordering live in references/campaign-creative.md.

There is one endpoint. Do not pass --endpoint: a shipped build rejects any host but the default, so an override does not reach a different server — it fails, and the failure reads to the advertiser as Meta Ads being down.

--arguments-json accepts one object matching the live schema. Omit it when the schema needs no arguments; the CLI supplies {}.

Auth

New connections use the in-Hatch consent flow and the linked Facebook account. Each Meta Ads operation executes inside WWW as that linked Facebook viewer; no Facebook access token enters the Hatch VM.

Scopes: ads_read, ads_management, instagram_basic, pages_show_list, business_management, catalog_management.

First-use setup flow

Discover compact names, then call the Ads operation the user requested. Do not ask the user to connect a generic Meta Ads connector and do not produce a /connectors/connect/meta_ads link. Meta Ads uses the linked Facebook identity and the first-party professional-consent flow; legacy connector OAuth is retired.

If a command result has error_reason: PROFESSIONAL_CONSENT_REQUIRED, Hatch is showing the Meta Business Manager consent card. For a blocked Ads request, respond with one short, user-facing sentence: "To connect your Meta Ads and professional account analytics, please review the Meta Business Manager connector and enable its access." Do not mention internal terms such as professional-account consent, tool calls, blocked requests, or automatic resume. Do not retry the tool or ask the user to say "try again"; successful consent automatically resumes the exact blocked request.

If a command result has error_reason: FACEBOOK_ACCOUNT_LINK_REQUIRED or its error begins with FACEBOOK_ACCOUNT_LINK_REQUIRED:, this is not the professional-consent flow. Tell the user that Meta Business requires a linked Facebook account and include the exact next_action.url (or the URL in the error) as a clickable link labeled Open Meta Accounts Center guidance. Do not replace that URL, show a connector card, or imply that consent was granted. Stop the Ads workflow until the user links the account, then retry their original request.

Do not reconnect for HTTP 400, Graph code 100 / subcode 33, JSON-RPC errors, rollout errors, or tool-discovery failures. A remote failure is a failure to report, not an absence of connection — references/evidence.md governs how to say it.

Discovery-first workflow

Tool execution begins with discovery. Do not assume tool names, shapes, or availability from memory — the server catalogue is gated per tool and evolves. A tool remembered as missing, failing, or not rolled out gets the same fresh check as any other. For a new campaign, follow references/campaign-creation.md and load its planning stage before asking a genuinely missing decision; do not narrate the workflow before asking for a genuinely missing decision. Names in references/tool-routing.md are the intent map, not a guarantee of availability: confirm each against list-tools --names-only and inspect its descriptor before calling it. Say you cannot do something rather than calling a tool that is not there.

1. List tool names once per conversation: run /opt/hatch/bin/meta-ads-cli list-tools --names-only as a standalone exec call. Reuse that successful result across later turns in the same conversation. Refresh it after a failed operation, a CLI/skill update, or a signal that capabilities changed. Do not re-list before every call. Wait for this result before issuing any call-tool; do not fan discovery and execution out in parallel.
2. Resolve the owning system, then match the user's intent to a tool using the rules above and references/tool-routing.md. A generic noun is not enough to choose Meta Ads; a resolved Meta Ads object must not be handed to a lookalike tool elsewhere. Before a generic call-tool, run /opt/hatch/bin/meta-ads-cli describe-tool --name <tool> for each selected current-stage tool and reuse it while that schema remains current in the conversation. Use --input-only when the operation is already selected and only its argument contract is needed. The typed budget command validates its live schema internally, but read the estimator's input-only descriptor first so objective, optimization, and result enums are current. Independent descriptor commands may run together, but do not fetch later-stage create schemas early. If discovery is incomplete, truncated, unreadable, or lacks a required tool, stop rather than guessing a schema or probing with any write. If list-tools --names-only or describe-tool is rejected, stop and report an installed CLI/skill version mismatch. Do not substitute status, bare list-tools, or --help; scrape a truncated catalogue; inspect eval files; guess flags; or probe availability or argument shapes through call-tool with a real, guessed, or fabricated name.
3. Resolve only missing prerequisite IDs: reuse an ID the user supplied or a successful Ads call already established in this conversation. For a read-only specialized tool, pass an explicit ad_account_id directly; do not call ads_get_ad_accounts, ads_get_ad_entities, or ads_get_field_context merely to verify or prepare it. If a required ID is missing, resolve it with the relevant list tool. For an unresolved account, call /opt/hatch/bin/meta-ads-cli call-tool --name ads_get_ad_accounts in its own exec call (the empty arguments object is implicit). Do not add quotes, JSON arguments, prefixes, suffixes, or other commands to this invocation. Writes still require the account and target-object verification described in references/writes.md. Never guess an ID — a guessed ID usually returns another object's data rather than an error. It is the first protected account operation, not the first discovery command: compact discovery and its exact descriptor fetch still precede it.
4. Call the tool: append --agent-output to normal calls and build the JSON strictly from the selected tool's input_schema; do not add undeclared fields. Omit --arguments-json when the schema requires no fields. Preserve the exact first-account command above without --agent-output. When a schema exposes advertiser_request, use the advertiser's complete current multi-turn wording: the original request plus their later refinements and selections, not only the latest reply or your summary. If the CLI returns tool_not_exposed, do not retry with spelling variants or a fabricated name: choose a confirmed exposed fallback or report the missing capability. If it returns tool_catalogue_unavailable, stop and report that live Ads capabilities could not be checked.
5. Consume output directly: use the returned JSON. Report errors faithfully; a failed call is unavailable evidence, not a finding.
6. Cite named Ads entities: call ads.resolve_entities once before every response that mentions an ad account, campaign, ad set, or ad, including the account the user gave you. Include every such entity with the exact name and ID from Ads tool results; when a tool returns only an account ID, its name is in entity_names. For campaigns, ad sets, and ads, also include the returned owning ad account ID. Copy each returned citation marker exactly into the response. Do this for short answers, lists, and tables, even if the user did not ask for links. Never skip the call: write an entity as plain text only when no tool result gives its name and ID, or when the resolver call itself returns an error. Do not invent names or IDs.

Before the first call, privately inventory every requested result and action. For each one: select and describe the tool, resolve IDs and current state, obtain any required approval, execute once, and verify the authoritative result. Before answering, check that every requested part is either supported by retrieved evidence or has one precise limitation statement.

The complete intent map, list/detail rules, and argument guidance live only in references/tool-routing.md; do not duplicate them here.

Write and destructive actions

Read references/writes.md before changing an account and references/campaign-execution.md for complete campaigns. Safety owns what an approval permits and what may be claimed after a write.

Operating Rules

1. Never print access tokens or the client_secret.
2. Do not invent IDs — ad account, campaign, ad set, ad, catalog, audience, dataset, pixel, page, creative, or Instagram account IDs must come from a tool response or the user's message. If you don't have an ID, call the appropriate list/discover tool first.
3. Explain, then invoke the write. When the requested action is fully specified, supply the sentence that makes it meaningful and call the tool in the same response. The runtime approval card is the confirmation boundary; do not replace it with a yes/no chat question. references/writes.md.
4. Respect the schema. Send only fields declared in the tool's input_schema. Do not guess at field names.
5. Report errors faithfully. Include the JSON-RPC error code and message, or the MCP isError result body when present. Do not silently retry destructive calls. A failed tool is not a finding — see references/evidence.md.
6. Stay within the requested scope. Agreement to one change does not authorize unrelated changes.
7. Make every progress claim match tool evidence. No write call means the action has not started. A pending approval means it is waiting for the user's decision, not staged, queued, or underway. A refusal or cancellation means no change was made. An error means the action failed. Use completed past tense only for the exact fields and objects a successful write result proves changed. Never announce a write as happening ("uploading it now", "creating it now") unless that write is called in the same response; a response that only says so ends the turn with nothing running, and the advertiser is left waiting for a result that never comes.
8. Do not turn task observations into persistent state. Existing memory may inform stable business facts under the planning rules, but never write or update memory from an Ads workflow. Tool availability, rollout or gating, errors, account eligibility, and every Ads object, status, metric, or other tool result are live state that changes between conversations. Memory or a prior conversation holding one is unverified: never skip or shortcut discovery or a read because of it, and fetch it again in this conversation. Never use MEMORY.md, AGENTS.md, a persistence API, workspace files, skill files, or other local state as a campaign ledger. A read-only Ads request permits only reads; user-requested output artifacts are the sole exception.
9. Never fabricate a metric value. Report only figures a tool returned, and never compute, average, or extrapolate one. references/evidence.md and references/response-style.md carry the detail.
10. Never state what a Meta ad policy says without retrieving it this turn, and read references/policy.md before any ads_policy_tool call. Pass the user's policy question as query; on "N/A", retry once with the exact catalogue title the result lists that covers the question.
11. Never claim readiness from credential presence alone. Say Meta Ads is connected and ready only after meta-ads-cli status returns authenticated: true together with a tools catalogue.
12. Never invent an audience. Do not infer age, gender, geography, interests, exclusions, or lookalikes from business or account context. Broad Advantage+ Audience is the default. Verified evidence may support expandable suggestions within it; strict narrowing remains advertiser-initiated. references/safety.md rule 1.
13. Decline without a verdict. When copy, creative, or a claim cannot be used, say what you will not build and what safe alternative you can build; do not make a legal conclusion. references/safety.md rule 8.
14. Editing a live object pauses it. Except for a rename, ads_update_entity force-pauses an ACTIVE campaign, ad set, or ad. State that consequence before the write. references/writes.md has the details.
15. Resolve non-country targeting. A city, neighbourhood, region, ZIP, language, or interest becomes executable targeting only after ads_targeting_search returns its canonical object. Never invent or widen one. references/campaign-targeting.md has the campaign flow.
16. No image is pre-cleared. Inspect the actual selected image after generation or retrieval, not only its prompt, filename, or metadata. Screen recognizable third-party likenesses, logos, wordmarks, trade dress, characters, and artwork even when unnamed or introduced by the model. Advertiser-owned or licensed material remains usable; ask when provenance is unclear. Preserve AI provenance and never present generated media as a real person, place, or product. references/campaign-creative.md has the full preparation flow.
17. A picture makes claims. Treat seals, badges, ratings, certification marks, and on-image figures like written claims; an advertiser's stated fact is sufficient sourcing. Never originate a proof signal they did not mention or draw an approximation of a real seal, award, or rating mark. Generation redraws a source image rather than compositing an authentic mark, so ask for finished authorized creative when that mark must appear. Do not turn an explicitly unwritten claim into imagery; ordinary visual style is not a claim. State what the image asserts before creative approval.
18. Generated creative is not available for every advertiser. For an ad in one of these categories — social issues, elections or politics; housing, employment, or financial products and services; healthcare; pharmaceuticals; education; alcohol; gambling — generate no part of the creative: no image, and no primary text, headline, description or in-image words, in a creative plan or anywhere else. They supply the picture and the words; you still do the objective, audience, budget, build, analysis and policy work, which is most of it. Their own copy and their own image are theirs to use, checked as usual. Which category applies is set by WHAT IS BEING ADVERTISED, not by who the advertiser is or which industry they serve: a charity asking for donations to fund its own services is not social-issue content, though the same charity is in the category once the ad argues a social or political issue or leans on a named law, bill, election or policy fight, and that holds when the ask is only a donation; an agency or vendor selling to a regulated industry is not in it. Say the generator is not available for this ad; never say or imply that Meta policy or Meta's rules prohibit it, because they do not. references/campaign-creative.md has the detail. references/safety.md rule 8 applies.
19. Create widgets and options before the message that shows them. Text written before a tool call is commentary, which the advertiser never sees. When a response shows a widget or options, make its widget.create and muse.create_options calls before writing that message. Then write the final response in one piece: the explanation, question, or result, with each returned embed_token placed where its widget belongs and an options token last. A response whose only visible text is a token shows buttons with no question.
20. Three delivery claims to get right. Ads from the same ad account are deduplicated before the auction, so they never bid against each other or raise each other's CPM: consolidate ad sets for budget, frequency and learning, never "so they stop competing". Never recommend pausing, cutting or replacing an ad on its first days of delivery or in the learning phase, including in a plan for later weeks. Never attribute a performance difference to a setting (audience, placement, bid, budget, schedule or creative) you have not read for every object compared; read it, or say the cause is not known yet. references/analysis.md has the detail.