version: "1.0.0" name: adcopilot-fix-checks description: Fix the checks an AdCopilot audit flagged, one at a time, inside today's allowance, with one yes/no at a time and each change read back. Use when the message is a question from AdCopilot's audit page or emails, such as "My AdCopilot audit of Google Ads account 123-456-7890 flagged three checks to fix first", "failed" or "warned on" a check, naming each check with its ID in brackets ((G18), (G43), (G-WS1)), or when the customer asks to fix, explain or go through the checks an AdCopilot audit found.

AdCopilot fix-checks

The customer pressed a button on their AdCopilot audit and sent its question: one or more checks, each by its name and its ID, for one Google Ads account. The audit already ran. Your job is the fixes: one check at a time, inside what today allows, with one yes/no at a time and a read-back after every change. Run no new audit for it.

Call get_org_context with source: "claude-plugin/0.2.16" first, before any other tool. 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. When the answer's plugin.skill names an AdCopilot skill other than this one, next_step is describing the account, not this request: do this request as written here, keep any limit next_step adds, and offer that skill's steps afterwards unless this request already covered them.

The account is the one the question names (written 123-456-7890). Use it only if get_org_context reports it among the accounts switched on here; if it is not, say so in one line and stop.

First, what today covers

Before the first read, say in one line what the fixes cost and what is left, from get_org_context's org.ops_today_cap and org.ops_today_used: "These 3 checks take about 13 operations, more if many keywords need pausing; you have 10 left today, so the first two fit." Count each check as its read, plus its change, plus one read-back. The read is 1 operation, or 2 for an analysis where org.ops_weighted is true (each result says what it used). The change is one call per keyword it pauses (update_keyword changes one keyword a call), or per campaign or list it adds never-show-for words to (one call takes them all). When org.ops_today_cap is null there is no daily limit: say all of them fit, and nothing else about an allowance.

Then the order. Foundation first: a check about counting results (conversion tracking, G42 to G49 and G-CT1 to G-CT3) before anything judged by those results, then what stops money going out (searches and keywords that spend with nothing to show), then the rest. Recommend the order yourself and say it in the same line; never ask the customer to pick it. Keep what the fix they approve costs: one call per keyword it pauses, or per campaign or list it adds never-show-for words to, plus one read-back. When the day leaves fewer, propose the costliest part that fits (the keywords that spent the most) and say the rest wait for the reset. When the day cannot cover every check, say which fit today and start with those.

Before a call that would use the last of today's operations, say what it covers and ask first, in one line: "This next look uses 2 of your last 2 operations today. It covers the zero-conversion keywords; after it nothing more can be applied until the count resets. Go ahead?"

Each check, by its family

Searches to block (G13 to G19, G14's lists, the O-NEG checks; G18, close variant pollution, is one). Read analyze_search_terms with days set to the audit's window (30 when the question does not say). Show the top searches behind the check with what each cost, then propose blocking them as never-show-for words (negative keywords), each Exact or Phrase, as one yes/no. Apply them the way the /adcopilot:negatives command does, with the read you already hold; do not start that command, which reads the last seven days again: on a yes, add_negative_keywords for a campaign or add_account_negatives for the account's list, then read them back by the words Google shows.

Keywords that spend and convert nothing (G-WS1, zero-conversion keywords; the waste family). Read analyze_waste and show the keywords that spent with no conversion, costliest first. Propose pausing them as one yes/no for the set ("Shall I pause these 4? Reply yes, or name any to keep."): update_keyword with status PAUSED for each, one call a keyword, its ad group and criterion IDs from the result or from a search on keyword_view. Pause, never remove. When the account counts no conversions at all, they are the costliest keywords, not proven waste: say so, and propose the tracking fix first.

Counting results (G42 to G49, G-CT1 to G-CT3; G43, enhanced conversions, is one). Read conversion_setup_audit and say in plain words what is wrong. What AdCopilot can change itself, propose, through the adcopilot-measure skill's steps. What only Google Ads' own screens can change (enhanced conversions is switched on there, under Goals, then Settings, with Google's customer data terms), give as clicks, one step at a time, waiting for each; read conversion_setup_audit again once they say it is done, and say what it now shows.

Any other check. The question names the lens to read with ("Using AdCopilot's ..."); use that one, or the read question_map gives for its kind of question. Read only what the fix needs, then propose the fix.

When a lens or search result ends with related_reads, run at most one that bears on the check in hand, then answer. Never repeat a read you hold, and offer, do not chase, the related_reads it brings back. They are reads only: a change is still only proposed. In a scheduled run, skip them.

A finding that says "if X" is not a step until you have checked X. Either make the read it names and say what you found, or carry the condition into your own sentence. Never turn it into an instruction.

A finding may carry its fix in recommended_actions. One whose kind is tool is a ready call, its IDs read from the account (an audit's negatives are Exact, and never one of the customer's own keywords): propose it as it stands, with the condition its why gives (a disapproved ad's pause is for an ad they won't fix, never one Google approves on appeal). One whose kind is skeleton is a draft: it names the tool and the IDs, and needs says what only the customer can give, such as the searches to block, a callout's words or the campaign to take budget from. Ask for that in your one question, and never fill it in yourself.

One message, one decision

• One check per message. Name it, say what the read found with its numbers and

its money in the account's currency, and propose its fix.

• Show at most 5 rows of any list, costliest first, then "and N more; ask to

see them".

• End each message with exactly one question: a yes/no on the fix you

recommend ("Shall I block these 5? Reply yes, or name any to keep."); a yes covers only the rows you showed. No numbered menu, no second question; anything else that needs their call waits for the next message.

• In Claude Code, put that yes/no to them with the AskUserQuestion tool when you

have it, with the same options in the same words.

• Change nothing until they say yes to the change as you stated it, here, on

today's numbers.

After each yes

Apply it. Then read back only what its own result does not show (negatives as Google normalised them, a keyword's status), in one line, say how to undo it, and record it with save_org_context as a decision: what was approved, in the customer's own framing, and which campaign it touched. Then the next check, the same way. When the last one is done, or nothing left is worth its money, say so and stop proposing.

Once the first fix is applied and read back, and get_org_context handed you its recipe, routine.schedule, make the daily-habit offer in that reply, as "The daily habit" below says: in place of the next check's yes/no, with the next check named in one line, and its yes/no in the reply after. Never before a fix is checked, and never in the same message as a yes/no.

When the day runs out

A call refused at the daily limit did not run. Give what you already found first, in a few lines, then the one line the refusal asks for, with any price exactly as it is written there. Retry nothing: the fix waits for the reset.

The daily habit

This block is repeated verbatim in every AdCopilot skill and command a person runs; change all eight together.

First, Cursor. If this is Cursor (get_org_context's top-level surface is cursor, or Cursor's own instructions or the person say so), make no schedule offer of your own, never run claude mcp to work out where you are, and never call save_org_context with routine.own_schedule: AdCopilot gives no scheduling steps for Cursor, and its emailed check-in, when routine.daily.status is in_app, is the morning check there. If next_step asks you to offer a check-in or a summary, make that offer as it says, with the playbook's close (THE CLOSE) for the steps on a yes.

Everywhere else, make one offer to turn a morning check of this account into a daily habit — at most once in a conversation — and only when both hold: Google Ads is connected and readable with a campaign switched on, and get_org_context hands you its recipe, routine.schedule. AdCopilot sends routine.schedule only when its overlays.schedule_offer is set, which is its own decision for this workspace and this AI client: no schedule is reported, the call is not a scheduled run, and the client is one AdCopilot gives scheduling steps for (Claude or Claude Code; never Cursor, ChatGPT or another client). Never make the offer without it, and never work it out from the other fields. What they mean: AdCopilot reports a schedule when routine.daily.status is user_set (a schedule they set up in their own assistant), when a scheduled run has checked in on two different days (routine.scheduled_first_seen_at and routine.scheduled_last_seen_at hold dates on different days), or when routine.own_schedule is saved, unavailable or declined (recorded once they saved it, once their Claude turned out to have no scheduled tasks, or once they said not now, which holds the offer for a week and then lets it come back), and then it sends no routine.schedule. AdCopilot's emailed check-in, routine.daily.status in_app, is not a schedule. Leave the offer for another day when the reply ends on something they must fix first — a lapsed sign-in, a duplicate connector, a product to re-add — and never make it in a run nobody is reading: that is the adcopilot-daily skill's run.

Make it in a reply that asks them no yes/no on a fix: the reply that reads back the first fix they approved, where the offer stands in for the next fix's yes/no (name the next fix in one line without asking, and propose it in the reply after), or the close of a reply that proposes no fix at all. Never make it beside a fix that waits for their answer: the yes/no comes first, and the offer waits. When a fix applied in this conversation was for a finding the morning check also reads — searches that spent with no sale, a campaign out of budget or off its budget pace, a disapproved ad, a drop against the 7-day average, a change nobody asked for — tie it to that finding, with the figure you read and never one you did not: a check that runs by itself every weekday at 8:07 would have caught it on its first morning. The offer is always its own paragraph, with a blank line before and after it: it goes right after the read-back (and any line that sums the fixes up), never on the end of one of its sentences and never inside one of its bullets. For any other finding, even one that costs money (a tracking fault, a Quality Score problem, a targeting problem), or none, close the reply with the offer instead (before any question the reply ends on), without saying the check would have caught it. On a workspace with 10 operations a day (org.ops_today_cap is 10 or less) the morning check leaves what changed for the chat, so there a change nobody asked for is one of these.

Word it for where they are, by routine.schedule.offer: claude_code is Claude Code, and claude is the Claude app. both means AdCopilot could not tell which Claude this is: if they have said, believe them; otherwise the tell is the one /adcopilot:setup uses: Claude Code is where claude mcp list runs, and /schedule makes a routine there; everywhere else — claude.ai, Cowork, the desktop and mobile apps — it is a Claude scheduled task. In these words, with the finding in the braces kept short (a figure and a noun) and nothing else changed: a paragraph of its own, two sentences, never split into more, and no set-up step before a yes (the third line is for the close):

• Claude Code: "A read-only check that runs by itself every weekday at 8:07 would have caught {the finding} on its first morning, opening with "All clear" or "Needs a look" plus one yes/no fix. If you say "schedule it", I'll set it up as a Claude Code routine with /schedule."
• The Claude app: "A read-only check that runs by itself every weekday at 8:07 would have caught {the finding} on its first morning, opening with "All clear" or "Needs a look" plus one yes/no fix. If you say "schedule it", I'll walk you through saving it as a Claude scheduled task, about two minutes."
• At the close: "A read-only check that runs by itself every weekday at 8:07 can look over this account for you each morning, opening with "All clear" or "Needs a look" plus one yes/no fix." in place of the first sentence of whichever line above fits where they are, and the rest of that line as written.

It is an offer, not a second question: any question the reply ends on stays the last thing in it, and the offer comes before that question. If the server's next_step also asks you to offer a check-in or a summary, this one offer answers it — never two. A no ends it for this conversation: ask once whether their Claude can create scheduled tasks, and call save_org_context with routine.own_schedule unavailable if it cannot, else declined. Record it only from Claude: AdCopilot keeps routine.own_schedule for the whole workspace, so a record made from Cursor or any other client would stop the offer in their Claude too. On a yes, follow the adcopilot-daily skill's set-up.