What a skill is

A skill is a named prompt pattern: a fixed instruction that tells the agent which MCP tools to call, in what order, and what to return. You invoke a skill by name or by description in any connected AI client.

Skills do not add scopes or bypass the draft-before-live gate. Every write a skill produces is a Draft object that waits for your approval in the SpendSignoff app. The agent itself cannot approve.

Skill: Audit

Reads the full account structure and performance snapshot, then summarizes wasted spend, pacing anomalies, and top movers.

• Invoke: "run an audit on my Google Ads account" or "audit [account name]".
• Tools called: list_ad_accounts, get_account_snapshot, query_entities (campaigns, ad sets, keywords).
• Returns: a plain-English breakdown of pacing status, campaigns burning budget with no conversions, top-performing ad sets by ROAS, and anomalies detected in the trailing 7 days.
• Produces no drafts. The audit is read-only.
• Typical completion time: under 30 seconds for accounts with fewer than 500 active campaigns.

Skill: Launch

Drafts a new campaign, ad set, and ad structure from a plain-English brief.

• Invoke: "draft a new retargeting campaign targeting users who visited the pricing page in the last 30 days, $50 daily budget, optimizing for purchases."
• Tools called: get_account_snapshot (to check existing structure), propose_change (typed payload for campaign create).
• Returns: a structured diff showing the proposed campaign, ad set, targeting parameters, budget, and bid strategy. The diff lands in the SpendSignoff approval queue.
• The agent does not upload creative assets in V1. If a new ad requires an image, it flags this in the draft and you add the creative in the ad platform UI before approving.
• Launch drafts that reference audience segments the account has not created will flag the gap rather than fail silently.

Skill: Optimize

Identifies high-impact changes across the account and drafts the top three by expected impact.

• Invoke: "optimize my Meta account" or "find budget inefficiencies and draft fixes."
• Tools called: get_account_snapshot, query_entities (performance by campaign, ad set, and ad), propose_change (up to three draft operations).
• The agent ranks candidate changes by estimated spend recovery or ROAS improvement and drafts the top three. It explains its reasoning in plain English before producing each draft.
• Each change is a separate Draft object in the approval queue. You can approve, reject, or defer them independently.
• The skill does not run the autonomy loop rule. It is a one-shot analysis triggered by your instruction, not a background process.

Skill: Report

Produces a normalized spend and performance summary across linked Google Ads and Meta accounts for a specified time range.

• Invoke: "give me a performance report for last week" or "compare this month to last month across all accounts."
• Tools called: query_entities with metric filters, cross-platform attribution substrate for unified ROAS computation.
• Returns: total spend, impressions, clicks, conversions, and ROAS per platform and per campaign, plus a plain-English narrative of the key movements.
• Reports are read-only. No drafts are produced.
• Time ranges use calendar weeks and months aligned to the account timezone. Specify a custom range with ISO 8601 dates: "report from 2026-05-01 to 2026-05-31."

Skill: Monitor

Runs a point-in-time anomaly check across linked accounts and returns any active anomalies with severity and recommended action.

• Invoke: "check for anomalies" or "is anything on fire in my accounts right now?"
• Tools called: get_account_snapshot, query_entities (trailing 24-hour metrics vs. 7-day baseline).
• Detects: spend spikes (more than 2x daily baseline), CTR drops (more than 40% below 7-day average), frequency above 5 for Meta ad sets, CPA drift above 30% above target, and budget exhaustion before 6pm local time.
• Returns each anomaly with: platform, entity type, entity name, metric name, observed value, baseline value, and a plain-English recommended action.
• The Monitor skill is read-only. If you want the agent to propose a fix for a detected anomaly, follow up with "draft a fix for the [anomaly name]."