Weekly Creative Readout Ad Creative

Weekly Creative Readout

Shared quality bar: ../references/output-standard.md. All numbers cited here live in ../references/thresholds.md.

This is the only composing skill in the pack. It runs no new arithmetic and issues no diagnosis of its own; it collects verdicts, resolves conflicts between them and writes the week in plain words.

Use this skill when

The weekly reporting slot is due, or a decision meeting needs one page. It is not the place to diagnose a specific ad mid-week; route that to the skill that owns it and bring the verdict back here.

It is also not the place to decide what gets built next. creative-brief-from-account-data-ad-creative runs on the same weekly ad-level export as the degraded path below, and reports the same three figures from it, spend, delivery share and the collapsed angle count, so the two are genuinely hard to tell apart at the point of routing. The deliverable separates them. If a designer, editor or freelancer is waiting for instructions, it is that skill. If a client or a meeting is waiting for the week, it is this one. This skill names production as a priority where coverage is short, and stops there; it never writes the brief items.

Required input

Four upstream outputs, each optional but each named when absent:

• Asset coverage check, for angle, format, ratio and library age.
• Creative fatigue read, one per creative proposed for cutting.
• Creative test readout, one per finished test.
• Losing ad diagnosis, one per ad proposed for rewriting.

Plus the week as dates, spend by ad set, any change a human already made inside the week, and which Advantage+ creative enhancements changed inside it (test.creative_enhancements_rule). An enhancement switched on mid-week changes the served asset, so any week-on-week creative comparison crossing that date is comparing two things.

On a normal Monday none of the four exist. What exists is a raw ad-level export, and the readout still has to happen, so the degraded path below is the default case rather than the exception. The minimum this skill will run on is that export: ad name or ID, ad set, spend, impressions, link clicks and conversions for the week, with the attribution setting stated.

Analysis workflow

1. List which of the four upstream outputs exist for this week and which do not. An absent input is written into the readout as missing, never silently skipped and never replaced by this skill's own judgement.
2. For each absent input, state the one decision it would have unblocked and what it costs to leave it unresolved for another week.
3. Collect every ship, cut, test, rewrite, investigate and monitor decision from the supplied outputs, carrying each one's evidence label and confidence through unchanged. Never upgrade a confidence on aggregation.
4. Resolve conflicts. Where a test readout says ship and a fatigue read on the same creative says cut, the later window wins and the conflict is stated in the readout rather than quietly dropped.
5. Drop any recommendation whose source row is low confidence out of the ship and cut lists, into monitor, with the date it becomes readable.
6. Check this week's proposed tests against test.concurrent_variant_cap and test.one_change_rule before they are committed, so the readout does not commission a test that cannot be read.
7. Check the coverage input against coverage.stale_spend_share, reading the share of each ad set's spend that rode old creatives rather than the launch date of the newest file, and where the share is above the key, put production ahead of optimisation in the priorities.
8. Say something about the money, because a readout that names what to ship and cut while issuing no number about spend has a hole in the middle of it. Judge every proposed shift against money.relevance_floor: a difference that does not move cost per acquisition by that much is not a reason to move budget, and the readout says so instead of recommending a shuffle. Where a move is justified, state it as a single step inside money.budget_step_cap, with the date it may next be stepped again, and hand it to a person. This pack never moves a budget. Where cost per acquisition was not supplied, name the decision the readout therefore cannot make rather than substituting a rate.
9. Run the degraded path when upstream outputs are missing, which is most weeks. From a raw ad-level export alone this skill may state: spend by ad set and how it moved; which creatives actually carried the delivery, which sat under delivery.never_delivered_floor and therefore did not run at all, and which were flagged by delivery.starved_flag, which is a different line in the readout because those ones did run and the platform kept declining to spend on them; which creatives clear delivery.ranking_floor and may therefore be compared with each other at all; the count of live creatives and, where angle sentences were supplied, the collapsed angle count against coverage.angles_per_ad_set; the share of spend on old creatives against coverage.stale_spend_share; and any human-made change already logged for the week. It may not state, and writes in as missing instead: any fatigue verdict, any winner or tie between variants, and any funnel-stage diagnosis of a weak ad. Those three sections carry the name of the skill that owns them and the decision each would have unblocked. The degraded readout is a shorter document that is true, not the same document with the gaps filled from the export.
10. Rewrite every line in client-readable language: what happens, to which ad, and what it is expected to change. No threshold keys and no internal labels in the client-facing paragraphs; those stay in the table.

Decision rules

• Ship only what arrived as ship at high or medium confidence from a test readout.
• Cut only what arrived as cut from a fatigue read that satisfied fatigue.verdict_rule. A creative proposed for cutting without that verdict goes to monitor with its re-read date.
• An Advantage+ enhancement changed inside the week, or nobody can say: no week-on-week creative comparison is presented as one. Under test.creative_enhancements_rule the readout says the served asset changed mid-week and names that as the reason a comparison is absent, rather than printing the numbers and leaving the client to draw the line.
• Test only pairs that satisfy test.one_change_rule and fit inside test.concurrent_variant_cap.
• Where the coverage check is missing, do not commission production. State that the brief is blocked on a count nobody has run.
• No upstream skill output exists at all: produce the degraded readout. State the export-only sections, write the fatigue, test and diagnosis sections as missing with the skill that owns each, and do not let the export's own numbers stand in for a verdict. A ranked list of last week's ads is not a test result and is never presented as one.
• A budget move is proposed and the difference behind it is inside money.relevance_floor: do not recommend the move. Say the difference is real and not worth spending against, which is a finding.
• A budget move survives that floor: recommend exactly one step within money.budget_step_cap and name the earliest date for the next step. Larger steps destabilise delivery and make next week's readout unreadable.
• Cost per acquisition not supplied for the week: name the ship-and-cut decisions that therefore cannot be made, rather than making them on rates.
• Anything touching claims.restricted_topics is approval_needed before it reaches the client page.

Output format

Open with one line: what changed in the creative programme this week.

Then three short client-readable paragraphs, ship, cut, test this week, then What this could not see, Missing data, Approval gates.

Practical example

Illustrative made-up account data, not a real advertiser.

Inputs supplied: a coverage check reporting 9 live creatives across 2 angles, below coverage.angles_per_ad_set, with 71 per cent of the week's spend riding creatives older than the age in coverage.stale_spend_share; two fatigue reads, one satisfying fatigue.verdict_rule and one where only the CTR signal moved. Spend for the week 18,400 across two ad sets, cost per acquisition 47 in ad set A against 44 in ad set B. No test readout was supplied, and no losing ad diagnosis.

Reading: one cut is supported, the creative whose fatigue read satisfied fatigue.verdict_rule. The second creative moves to monitor, because one signal is not fatigue and this skill does not re-decide that. Coverage puts production ahead of optimisation, since the angle count is below coverage.angles_per_ad_set and the spend share on old creatives is above coverage.stale_spend_share, so next week's numbers will again be read on ageing supply.

Money: the request to move budget from ad set A to ad set B is declined. 47 against 44 is a difference of 3, which is 6.4 per cent of 47, inside money.relevance_floor, so it is not worth spending against and the readout says which key stopped it. Had it cleared, the recommendation would have been a single step inside money.budget_step_cap with the date it could next be stepped, handed to a person to make.

Declined: nothing is shipped or scaled this week. The test readout is missing, so no variant has a comparison against a control, and naming a winner from raw weekly figures would invent a result the pack refuses to produce. That absence is written into the readout with its cost: the hook test that finished on Friday stays unread, so the rollout decision slips a week.

Output: one cut, one monitor, one approval_needed on production budget, and missing against the test readout row.

Guardrails

• Never issue a fatigue, test or funnel verdict in this skill.
• Never upgrade a confidence label while aggregating.
• Never present an absent input as a clean week.
• Never put threshold keys or internal labels into the client-facing paragraphs.
• Do not action anything. This readout is a decision put in front of a human.