Workflows

Opening a session

Ask for start_here. It is free and it answers with what is connected, the active accounts per platform with currency and timezone, the tasks left this period, and two or three prompts worth trying. If nothing is connected it hands over the link to Connections and stops.

When the right tool is not obvious, search_tools ranks all 231 against your own wording and returns the exact call line. Also free. See the tool surface.

Performance review

1. Fix the window. Thirty days by default; a week is not a trend.
2. The headline: google_get_campaign_performance, meta_get_campaign_performance, tiktok_get_campaign_performance, linkedin_get_campaign_performance or chatgpt_get_performance. Each returns the previous period of the same length beside the current one.
3. One level down where the answer is not yet clear: ad group, ad set, keyword, creative.
4. Explain the change rather than guessing at it: google_explain_performance_anomaly and meta_explain_anomaly rank the segments that account for it.

Judge every campaign on its objective’s own metric. Purchases and ROAS for sales, cost per lead for lead generation, CPC for traffic, CPM for awareness. No ROAS on an account that records no revenue.

Wasted spend, then negatives

1. google_analyze_wasted_spend for keywords and campaigns that spent with no conversions. The learning phase and trivial spend are already excluded.
2. google_analyze_search_terms for the queries behind them, sorted into converting, wasted and new.
3. Up to ten negatives at a time, each with what it blocks, then google_add_negative_keywords at campaign level after you approve.
4. Removing a negative later is the one reversal, and it needs confirm_delete: true.

Meta, TikTok and LinkedIn have the same first step: meta_analyze_wasted_spend, tiktok_analyze_wasted_spend, linkedin_analyze_wasted_spend.

Launching, on any platform

The order never changes, because each step removes a way to fail.

1. Spec. get_campaign_spec returns the required fields, text and asset limits, the objective matrix and the minimum budget for that campaign type.
2. Resolve. Free text never reaches a write: google_resolve_locations, meta_search_targeting, tiktok_search_targeting, linkedin_search_targeting, chatgpt_geo_lookup. Ambiguity comes back as candidates and stops the launch.
3. Check the creative. google_validate_ad_copy, meta_validate_creative_url, tiktok_validate_assets, linkedin_validate_assets, chatgpt_validate_chat_card. All free.
4. Dry run. validate_campaign_draft checks the whole draft field by field and returns every problem with a fix.
5. Plan. Objective, budget with currency, audience, placements, the copy in full, and what will exist afterwards.
6. Create. google_create_search_campaign, meta_create_image_campaign, tiktok_create_video_campaign, linkedin_create_image_campaign, chatgpt_launch_ad. Everything paused.
7. Verify. The create reads the objects back; verify_campaign_is_live answers whether a campaign exists, is switched on and has a budget it can spend.

Approvals

Under the inbox policy a write returns a proposal id, a preview and a diff. Nothing has happened yet. Approve it in the chat, or put your seal on it at Approvals. Under direct, your confirmation in the chat is the approval.

list_pending_proposals answers what is waiting. Proposals expire after 48 hours and then never run. Approving the same proposal twice is safe: the second call reports the first result and repeats nothing.

Monitors and briefs

1. test_monitor replays a rule over the last 7 days of synced data and says which days it would have alerted on. Free, so tune the threshold here.
2. create_monitor saves it. Evaluation is daily, costs no tasks, and reads Adako’s own copy of the account rather than the platform.
3. schedule_brief sends a daily or weekly summary by email; generate_report_now composes one immediately.
4. A firing monitor can propose an action. It never executes it: list_pending_actions and manage_action put it through the same approval as any other write.

Monitors, briefs and reports need the Pro plan or above. See Monitors.

When something fails

• explain_error turns a raw platform message into a cause, a fix and the next tool.
• why_did_this_fail reads your real call log and explains the calls that failed or were blocked.
• suggest_next_action looks at connections, approvals, the last ten calls and the quota left, and returns three concrete steps.
• get_connections_status says whether a token expired before you blame the tool.

Several accounts in one answer

meta_ads(action="execute", tool_name="meta_analyze_wasted_spend",
         arguments={"date_range": "last_30_days"}, accounts="all_active")

Reads only, one billed task per account, at most 20 accounts, one table per currency. Writes stay one account at a time.