atromxintelligence/adcopilot-claude-plugin/adcopilot-docs

AdCopilot docs

Questions about AdCopilot itself, answered from what the connector reports — never from memory. The product changes; this file does not move with it, so every number in your answer comes from a live read or from the documentation site, and nothing comes from recall.

Call get_org_context (with source: "claude-plugin/0.2.16") first. Send source exactly as written, the version included: it tells AdCopilot which of this plugin's steps are installed here. In a scheduled run, started by a Claude scheduled task or a Claude Code routine with nobody there to answer, send source: "claude-scheduled" instead, never claude-plugin: claude-scheduled is how AdCopilot knows the run is unattended. A question about AdCopilot itself is answered from that one call, plus the documentation site where this file says so. Make no Google Ads, Analytics, Search Console or Tag Manager read for it, and start no other AdCopilot skill. The answer's next_step, playbook (its sections on the installed skill and on a returning customer included), cross_reads and question_map describe the account, not this question: do not follow them on this turn, not even their offer of a skill's steps. Keep from them only a limit next_step names (look-ups left today, a teammate's login) and the facts below. If the customer says yes to the read you offer, that is a new request, and the skill that answers it runs then. The answer carries the facts most of these questions want:

• org.plan, org.ops_today_used, org.ops_today_cap, org.accounts_cap,

org.trial_days_left, org.upgrade_url

• products — which of the customer's Google products are connected, and what

each one unlocks if it is not

• app_urls — where the customer goes to change any of it

What AdCopilot can do

Never recite a list of products or a count of tools from memory: both move, and a stale list is a promise the connector cannot keep. Ask get_org_context and answer from products — what is connected, and for anything that is not, what the connector itself says connecting it would unlock. Invent nothing.

Answer it once, as one list, in this order: what is connected and what each lets them do; what connecting each of the others would unlock; the routines installed here (the daily check-in, the weekly summary, the audit, the negative-keyword clean-up, the first-campaign launch, conversion tracking); what AdCopilot may change, with its condition (below); and the plan's look-ups. Say what each routine reads and proposes, and leave approval to the part on what AdCopilot may change, where it is said once with its condition: no routine's line adds that nothing happens until they say yes. That is the whole answer: never a second list in another grouping, and never come back later in the same reply to answer it again. One fact from get_org_context's 30-day summary (an account's last_30d in ads_probe, with the dates it names, start_date to end_date) may be the reason for the offer that ends the reply. Anything that needs a read is that offer, not part of the answer.

If a tool seems missing, that proves nothing on its own: tool lists are cached by the client and can be out of date. get_org_context is the only thing that says what is connected — believe it over the tool list, and say so plainly rather than telling a customer a permission is missing.

Look-ups, and running out

A workspace gets a number of look-ups a day. ops_today_used against ops_today_cap is the live answer — give those two numbers, not a remembered allowance. A null cap means no daily limit on that plan.

When a workspace runs out, say what it just bought before saying what it costs. The honest version is short: name what the reads established — the spend, the waste found, the change proposed — then say the allowance is used up for today, when it resets, and that a paid plan raises it. Send them to org.upgrade_url — but only to someone who can act on it. Where the server says the workspace is reached through a teammate's login, or tells a member to ask their owner, pass that on instead of sending them to a billing page they cannot use. Never dress the limit up as an error, and never imply the work was lost: what was found is still true tomorrow.

Do not quote prices from memory. Plans and prices live at https://adcopilot.cloud/pricing; the customer's own plan and caps come from get_org_context.

What AdCopilot is allowed to change

Worth stating plainly, because it is the question behind most hesitation:

• Nothing is deleted except a negative keyword, and the server enforces it.

The campaign, ad group, ad and keyword tools accept only ENABLED or PAUSED and refuse a REMOVED status outright; the asset tools refuse a remove. Pausing is as far as it goes. This is the strongest guarantee the product has, because it does not depend on the assistant behaving — say it first. (A negative keyword is the exception, and it is deliberate: each of the three ways to add one has a removal behind it, so a negative can be taken back out. Setting a campaign's languages also replaces the list it had.)

• The approval lives in your AI client, not in AdCopilot. Reads run freely,

and AdCopilot proposes each change first. Whether the client then asks before the change runs is the client's setting, so the hard stop is the client set to ask before AdCopilot's write tools: Needs approval in claude.ai, Cowork and the desktop app; in Claude Code, tools that are not allowlisted, in a session that is not in auto or bypassPermissions mode; in Cursor, the write tools kept off the allowlist in its Allowlist run mode (Settings, then Agents, then Approvals & Execution, at the time of writing). Cursor has no Needs approval setting: its Auto-review mode, Cursor's recommended default, sends a tool that is not allowlisted to a classifier that can let a change through without asking, Run Everything asks nothing, and a Cursor cloud agent never asks. Take that question away and changes to things that already exist, a budget or a switched-on campaign among them, apply without being asked. The one exception is set_campaign_status, the tool that pauses a campaign or switches one on: Claude Code (version 2.1.214 or later) asks before every use of it whatever its settings, and a run with nobody there to answer cannot use it; the Claude desktop app may also ask every time for it. This plugin pauses or switches on a campaign only with that tool, never with a status in a change to the campaign's settings (update_campaign), which follows the client's settings like any other change. Only a new campaign is still built paused regardless. Never state the approval as unconditional: name the hard stop, because a customer who has turned it off deserves to know what they turned off. That holds wherever approval comes up — a heading, a parenthesis, a list item: never promise a yes before every change without the hard stop beside it.

• A campaign AdCopilot builds is created paused. Switching it on is the

customer's click, in Google Ads.

• It reads only the ad accounts switched on in the workspace, and never sees a

Google account that was not connected.

When something is not working

The steps are at https://adcopilot.cloud/docs/troubleshooting, and that page can be opened and read here — every AdCopilot documentation page has a plain-text copy. Read it and give the customer the step, rather than sending them away with a link alone.

For a product the connector reports as needing to be signed in again, say so and send them to app_urls.accounts. For one it flags as unreachable, the connector has to be removed and re-added in the customer's AI client — say that plainly; no amount of retrying fixes it.

Where the answers live

For anything about privacy, data handling, security or what AdCopilot is allowed to change, the written answer is on the docs site and the customer can be sent to it: https://adcopilot.cloud/docs — the concepts pages cover approval-gated writes, what can never be deleted, read versus write, account scope and where data travels. Every docs page has a plain-text copy you can read, so read it and quote it rather than answering from memory. get_org_context's help field carries the troubleshooting address for the same reason.

What is next

Answer what they asked first, whole. Then, when the question pointed at something a read of their own account would answer — what AdCopilot can do for them, whether a product is worth connecting, why a tool seemed missing — end the reply with exactly one question, offering one read: what it would show them about their account, tied to the 30-day fact that makes it worth doing. Not a fix, not a change, and not a second offer. When ops_today_cap is set and the hints say few look-ups are left today, say that the read uses one of them.

Skip it after a question about look-ups and running out, about the workspace's allowance or billing, or about privacy or safety: there the answer is the whole reply, and an offer to spend a look-up would read as a nudge. That holds even when plugin.skill names other steps.

What breaks

Answering a pricing or limits question from memory: the numbers drift, and a customer who is told the wrong allowance stops trusting the rest of the answer. Treating a missing tool as a missing permission: it is almost always a cached tool list, and telling a customer they lack access they have is the worst version of this mistake.