Answer Ads Questions

In one line: answer "how are my ads doing?" from a tool you called this turn, say which days the numbers cover and when they were last synced, and say out loud when the data is incomplete.

• Numbers come from tools, never from memory or the docs. The ads/ docs explain why Goose did something. They are not a source of metrics.
• Every number carries its window and sync time. "You spent $212 from Sep 17 to Sep 23 (data last synced Sep 24, 09:00 UTC)."
• Partial, stale or disconnected data is said straight after the answer, never buried or left out.
• This skill only reads. It never pauses, changes or writes anything.

Where this skill's files are

Paths are relative to this skill's own folder, not your working directory (in GooseWorks: agent-config/skills/answer-ads-questions/). The harness docs contract ships with the meta-ad-manager skill, installed beside this one.

Choose one mode

Use the first mode that is available:

1. GooseWorks when the GooseWorks Meta read tools are callable. Read references/gooseworks-adapter.md.
2. Direct Meta when there is no GooseWorks, but a Meta token with ads_read and the ad account id are available. Read references/direct-meta-adapter.md.
3. Planning only when neither is available. Read references/planning-only.md.

"The adapter" below means the file for the mode you chose.

Purpose

Give a user a true, dated answer about their Meta ads that they could act on. The usual way this goes wrong is that the agent repeats a number from an earlier turn or a report, and presents it as current. Or it misses that half the account did not sync. This skill fixes the source (a tool read this turn) and the framing (window, sync time, connection state).

Reach for it on any question about spend, results, delivery, which ad is best or worst, whether a push is live or paused, what Goose has noticed, or why Goose made a decision.

Inputs

• The user's question (required), in their words.
• The brand's ads/ folder (optional). Per the harness contract (the meta-ad-manager skill's contract/RULES.md): ads/README.md first, then only the campaign in play. Read ads/brand.md for one field only: familiarity. A brand with no ads/ folder has never run the harness. That is fine for metric questions. For "why" questions it means there is no recorded decision.

<!-- shared:answer-ads-questions start --> <!-- This block is maintained in ONE place and copied byte-for-byte into both the maintainers' source copy and the published goose-skills copy. Edit one, copy it to the other, and run the parity script before shipping either. -->

Workflow

1. Pick the read from the question

The adapter maps each read below to a real tool or API call, and says where each read's data window and sync time come from.

Only claim a read you made. If the host has no read for something (alerts, a watch, a push record), never say "I checked" or "there are no alerts". Say what you did read, and that the rest is not visible from here.

2. Check the connection before you say any number

For any question that ends in a Meta number, read the connection state once in the turn, in addition to the numbers. Some reads report "ok" while the connection is only partly synced, so the numbers' own status is not enough. The adapter says which read carries the connection state and how each class below shows up.

3. Answer by the six rules

1. Only say a number a read returned this turn. Not one from memory, an earlier turn, a report, or an Observed: line in state.md or decisions.md. You may quote a docs number only as history, with its own date: "the Sep 23 report showed $41 per subscriber".

2. Every number carries its window and sync time, in plain words, with money in the ad account's currency. Report rates as the adapter says the source returns them (a fraction or a percent): 0.021 as a fraction is "2.1%". Null means unknown. Never say it is zero. Missing days are unknown too.

3. Say the connection state when it is not clean (step 2), in the sentence right after the answer.

4. The docs are for decisions and reasoning, never for metrics. For "why", quote the decisions.md entry: its date, what was recommended, who decided, and its Evidence: pointer, written out (for example "evidence: push 15b8dfb0"). If no entry covers it, say "I don't have a recorded decision for that". Never reconstruct a reason. If the push record shows the ad was changed outside Goose (in Ads Manager), say someone changed it outside Goose. Do not say Goose did it.

5. An ad Goose did not publish is read-only. An ad is Goose's only when its context says Goose published it (a Goose push is its source), or when a push listed in the campaign's state.md meta_push_ids has an item with that ad id, or when the adapter's own launch record includes it. Only then do Goose's pause and revert apply (the adapter says which of them exist). Everything else:

   • "No Goose link recorded" alone does not prove Goose had nothing to do with it. A push the read cannot see (another brand's record, a creative uploaded by hand) looks the same. So before you say "Goose didn't make this", check the pushes in state.md, and whether the ad name carries a [gw:…] tag. With a tag but no push item, say it uses a Goose creative but you can't confirm Goose published it.
   • A link that says the ad uses a Goose creative, but was not pushed by Goose: say exactly that. It was not published through Goose.
   • With no Goose pushes at all (the adapter says when that is the case), every ad is the user's own.

   For every ad that is not Goose's, answer from the reads and do not offer to pause, edit, replace or scale it. If the user asks what to do, the most you offer is new creatives.

6. Match the wording to familiarity from ads/brand.md:

   • novice: no jargon. Give a one-line meaning for any metric you use ("CTR, the share of people who saw the ad and clicked").
   • intermediate: use the terms with a short gloss.
   • expert: terse. Metric names, no glosses.
   • Missing or unknown: intermediate.

Rules for reading any source:

• A clipped or truncated result: narrow the window, the entities or the metrics, and read again. Never answer from a clipped result.
• A result that withholds numbers (for example, the ad's creative was replaced): say that. Don't estimate.
• Never recompute a metric the source returned, and never compare against outside benchmarks.
• Name a metric exactly as the source defines it. Cost per click over all clicks is not cost per link click. A return on ad spend of 0 with no purchases tracked means no purchases were tracked, not a result.
• Report what a number is, not a business verdict. A return on ad spend of 0.7 is "about 70¢ of Meta-reported purchase value per $1 spent". It is not "losing money": Meta's purchase value is its own attributed figure, not revenue, margin or repeat purchases. The same goes for "profitable" and "wasting money".
• Never call data "complete" or vouch for a total ("the total holds"). A clean connection means the last sync finished, not that every row is proven. Say which parts synced and which failed, and let the numbers stand as reported.
• A same-length comparison period can be shown side by side. Do not say the account is "up" or "down".
• Performance of Goose's own creatives is not attribution. Never call it that, and when there are no Goose creatives with data, say so rather than falling back to account totals.

4. Shape the reply

1. One sentence that answers the question, with its window and sync time.
2. The caveat, if there is one, immediately after.
3. Any other numbers, each with its window and sync time.
4. At most one next step. None when the data is stale. None that changes an ad Goose did not publish.

Output

A chat reply in the shape above. Nothing is written: not the ads/ docs, not Meta, not memory.

Example (intermediate, partial connection):

You spent $212 on Meta from Sep 17 to Sep 23 (data last synced Sep 24, 09:00 UTC). One caveat: the last sync only partly finished, because ad creatives didn't import. Spend and clicks came through, but per-ad breakdowns may be missing some ads.

<!-- shared:answer-ads-questions end -->

Quality Checks

• Every number in the reply appears in a read made this turn.
• Every number has a date range and a sync time next to it.
• The connection state was read this turn, and its class decided what was said.
• A "why" answer cites a decisions.md entry date and evidence pointer, or says there is no record.
• No offer to change an ad that Goose did not publish.
• The words "currently", "right now" and "live numbers" never describe a synced figure.
• No business verdicts ("losing money", "profitable") and no "complete" claims.

Failure Modes