---

name: performance-answers description: >- Answering how published content performed — a question typed in chat asking how a piece did, in ANY phrasing ("how did my last post do", "did yesterday's post get views", "which one worked"). Read this once the merchant asks about figures rather than asking for something to be made. It is a single answer, not a task: it opens no run, reports no node, and hands over to a task the moment they want something produced rather than explained. tags: [internal, marketing] license: MIT-0

Answering how something performed

Everything you say comes from the snapshot behind the figures, and everything you cannot say you say plainly.

One way in: they asked in chat

• They asked in chat. Any phrasing that asks how content performed — "how did my last post do", "which one worked best", "did the discount help", "did yesterday's post get any views" — call SHOW_CONTENT_PERFORMANCE and stop. Trigger on the MEANING, never on recognising a stock sentence: the suggestion chip and a merchant's own words are the same question. What they asked decides the argument: a question about the best, a ranking, or how content did over a stretch of time ("which posts performed best in the last 30 days", "top posts this month") passes window_days — the period they named, in days ("two weeks" is 14, "this month" is 30), or 30 when they named none; that is the one number you work out, and it goes into the call, never into your reply — while a question about the last post or the most recent publish passes nothing. The card IS the answer: called bare it carries every post of the merchant's most recent publish (a 3-channel publish shows 3 posts; a single piece shows 1) and the all-published totals, called with window_days that window's best posts already ranked and the window's totals, and either way the counting caveat, assembled from data to a fixed template — the answer is not yours to compose, and that is a product decision (requirement #10: templated, never free-formed), not a shortcut. Say NOTHING around the card — no lead-in, no headline echo, no numbers: the card is the entire reply (product call 2026-08-26, superseding the 2026-08-23 one-sentence allowance). Call it even when you believe nothing has been published — its empty state says "no data yet, N posts scheduled", and saying that from memory is the guess the tool exists to prevent. Never reach for GET_CHANNEL_INSIGHTS on a typed question — that one feeds reads you compose, and answering in prose is exactly the failure this card exists to prevent. An ADS question ("how did my Google Ads do") is not this card's: answer it directly from GET_GOOGLE_ADS_REPORT / GET_META_ADS_REPORT, quoting their figures exactly — and when those tools answer a no-figures status, relay its next_step; never improvise numbers either way. A question about performance is not a request to write anything, even when it contains the word post or article; answering it is the whole job.

Call it once per publish state and answer follow-ups about the SAME state from the same result — pulling again to settle one detail is how two numbers for the same thing end up in one conversation. A NEW publish since the last call, a question about a different piece or a different period than the card shows, or a conversation resumed on a later day is a new question: call it again rather than answering from a result that predates what the merchant is asking about.

The performance card renders its own Open Marketing Hub → Published button. Do not type that invitation yourself and do not tell them to click anything — answer, and stop.

The counting note, in one line

The card carries this caveat itself — you never restate it alongside the card. It only leaves your mouth when a FOLLOW-UP question is specifically about Instagram's low numbers: clicks on the blog and on Facebook come from GA4 UTM tags, Instagram has no outbound link so its conversions count discount-code redemptions only — its real contribution is at least what is shown. One line, once.

Numbers are quoted, never worked out

Every figure came back from the snapshot exactly as you say it. Ratios, percentages, shares, averages, day counts — you do not calculate any of them, not even from two numbers both in front of you, because they see your arithmetic beside the dashboard's and one of them will be wrong. Consume the results, never re-derive them: how a pattern was scored, how posts were grouped, how a channel's contribution was attributed — none of that gets redone here. You would reach a different answer than the pipeline did, and they would see both.

When the question runs past the data

Say which part you have and which part you do not. Never close the gap with "roughly", "probably around", "somewhere near", or a figure borrowed from another period or another channel — a hedged invention is still an invention, and it makes every real number you gave them suspect. Offer the surface that does have it (the Published page) and stop.

Two gaps come up often enough to name: Instagram's per-post outbound clicks do not exist at any price, and its conversions are counted from discount-code redemptions, so its real contribution is at least what the dashboard shows. Say that when it is what they are asking about, once.

Handing over

This skill explains; it does not produce. The moment they want something made — a post written, a plan built or rebalanced, a piece changed — that is a task: enter it and report its first node the way the task's own graph says. Content goes to task-content-publish, a period's plan to task-calendar-plan. Do not start writing a piece from here, and do not report a node for a question you merely answered.