# Angle Inventory Ad Creative > Angle Inventory - Skill: `mardab96/angle-inventory-ad-creative` (Agent Skill) - Install (CLI): `npx skillmds@latest add mardab96/angle-inventory-ad-creative` - Raw SKILL.md: https://api.skillmd.com/api/skills/mardab96/angle-inventory-ad-creative/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: mardab96 (https://skillmd.com/u/mardab96) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/mardab96/angle-inventory-ad-creative --- # Angle Inventory Shared quality bar: `../references/output-standard.md`. All numbers cited here live in `../references/thresholds.md`. The deliverable is a coverage map: which arguments exist, which are duplicates wearing new art, and which gaps are worth filling. ## Use this skill when The library has grown but the arguments have not, or an account claims dozens of tests and cannot name what any of them established. Use it before a quarterly plan, when onboarding an account whose history is unknown, or when every ad in an ad set turns out to say the same thing. It counts and de-duplicates arguments across the whole served library, live and paused, back through the period under review, and its question is historical: what has this account actually argued. Do not use it to ask whether enough variety is live right now to learn from this week's spend, which is `asset-coverage-check-ad-creative` and reads only what is serving today. It does not rank creatives or issue production instructions (`creative-brief-from-account-data-ad-creative`) and it does not decompose one ad (`winning-ad-teardown-ad-creative`). ## Required input - The full list of ads served in the period under review, live and paused, with their headline, primary text and creative. - Ad set membership for each ad, since `coverage.angles_per_ad_set` is read per ad set, not per account. - First and last served date per ad. - Format and aspect ratio per ad, to read against `coverage.formats_per_angle` and `coverage.ratio_set`. - Optional: spend and conversions per ad, used only to mark which angles were ever given a readable chance. ## Analysis workflow 1. Write each ad's angle as one sentence in the buyer's terms, using the definition of angle fixed in `thresholds.md`. Format, offer and visual are not angles and are recorded in separate columns. 2. Collapse duplicates. Two angle sentences are one angle under `coverage.angle_distinctness` when they share their main claim AND their reason for it, which is a test on the argument, not on the wording. "Save time" and "stop wasting hours" share a word in neither position and are still one angle. Do not apply the noun and verb test here; that one belongs to `coverage.rsa_headline_floor` and is about headline slots. Record the collapsed count alongside the raw ad count; the gap between them is usually the finding. 3. Per ad set, count the surviving distinct angles against `coverage.angles_per_ad_set` and note where the recalibration to two applies because the ad set sits under `test.volume_floor`. 4. Per angle, count formats against `coverage.formats_per_angle` and ratios against `coverage.ratio_set`. An angle that only ever ran as one static in one ratio has not been tested, it has been mentioned. 5. Mark each angle tested or untested on evidence, not on intent. Tested means it ran for at least `test.min_runtime` and reached `test.volume_floor`, or `test.click_floor` where the read was on CTR. Everything else is `needs_data`. 6. Read supply against `coverage.stale_spend_share`: work out how much of each ad set's spend in the period rode creatives older than the age that key names. The reading is on money, not on the launch date of the newest file, so an account shipping constantly can still fail it. Above the share, the account is spending its inventory and the gap list is the pipeline. 7. Exclude every creative below `delivery.never_delivered_floor` from the tested and untested judgement. That creative did not lose its angle, it never ran. Mark it `never_delivered`, keep it out of any ranking or cut list, and count its angle as untested rather than as tried and failed. A creative that cleared exposure but stayed under `delivery.ranking_floor` is a separate case: it ran and it may carry a verdict of its own, but its share of delivery is too thin for its angle to count as tested here, so that angle also stays on the untested list. 8. Name the untested angles worth a slot. Derive each from something the user supplied: an objection they hear, a reason customers gave, a segment nobody addressed. Do not derive one from category convention, and label every proposed angle `hypothesis` until it has run. 9. Order the gaps by how cheaply each can be produced from assets that already exist, and cap the recommended slate at `test.concurrent_variant_cap` per ad set. ## Decision rules - Distinct angles in an ad set below `coverage.angles_per_ad_set`: the gap is the finding, and the next round is angles rather than executions. - Two angles collapse into one: `cut` the redundant slot and say what it was duplicating. Redundant slots are the mechanism by which an account tests a lot and learns nothing. - Angle ran under `test.volume_floor`: `test` again with a real share, confidence `low`. Never `cut` an angle that never got a readable sample. - Every creative carrying an angle sat under `delivery.never_delivered_floor`: the angle is `needs_data`, not a loser. Record it as untested, keep it out of the cut list, and say the ad set never gave it a share of delivery. - Every creative carrying an angle ran but stayed under `delivery.ranking_floor`: the angle is `needs_data` on the tested question. Those creatives may still have been cut elsewhere in this pack; a cut on starved delivery is a verdict about that execution and is not evidence that the argument failed. - Spend riding creatives older than `coverage.stale_spend_share` allows: the finding is supply, and the first item in the gap list is production rather than another argument. - Angle covered in only one format: `rewrite` into the missing format before any conclusion about the angle is written down. - Proposed angle rests on a claim with no source: `approval_needed` under `claims.evidence_rule`. - Proposed angle falls in `claims.restricted_topics`: route to the platform's own policy page and issue no verdict here. ## Output format No verdict line, this is a map. Two tables. Covered angles: | Angle sentence | Ads | Ad sets | Formats | Ratios | Tested | Evidence label | Confidence | Decision | |---|---|---|---|---|---|---|---|---| Gaps worth a slot: | Proposed angle | Where it came from | Cheapest asset path | Evidence label | Decision | |---|---|---|---|---| What `high` means in this skill, since it reads a library rather than a test or a time series: the ad list covered the whole period, live and paused, every angle sentence was written from the creative itself rather than from the ad name, and the collapse was applied under `coverage.angle_distinctness`. The comparison family's `test.min_runtime` and `test.volume_floor` conditions belong to a test and are not borrowed here; where a row states whether an angle was tested, that row inherits those conditions and is labelled on the scale in full. Then `What this could not see`, `Missing data`, `Approval gates`. ## Practical example Illustrative made-up account data, not a real advertiser. 23 ads across 3 ad sets. Raw angle sentences: 11. After collapsing, 4 distinct angles remain, since 7 sentences reduce to one argument about being cheaper than an agency, phrased seven ways across seven images. Reading: ad set B carries 2 distinct angles, under `coverage.angles_per_ad_set`, and its 12 ads created the impression of heavy testing. Three of the four angles ran in a single format, below `coverage.formats_per_angle`, so no angle verdict is written for them. The claim that the price angle is the account's proven winner is refused: it ran in 7 ads but only 1 cleared `test.volume_floor`, and the rest are `needs_data`. Supply: 61 per cent of the three ad sets' spend in the period rode creatives launched before the age named in `coverage.stale_spend_share`, past that key's share, so production leads the gap list. Two of the 23 ads took under `delivery.never_delivered_floor`; their angles are recorded as untested rather than as tried, because they never ran. Gaps proposed: one objection angle taken from a support transcript the user supplied, one segment angle for the audience the library never addresses. Both labelled `hypothesis`. A third idea about a compliance benefit was withheld, because it sits in `claims.restricted_topics` and belongs on the platform's policy page first. ## Guardrails - Never count an ad as an angle. Count arguments after collapsing duplicates. - Never mark an angle failed on a sample below `test.volume_floor`. - Never invent a customer objection, a segment or a proof point to fill a gap. - Do not pause or launch anything. The slate is a recommendation.