# Winning Ad Formula Finder > Winning Ad Formula Finder - Skill: `raneqbarber/winning-ad-formula-finder` (Agent Skill) - Install (CLI): `npx skillmds@latest add raneqbarber/winning-ad-formula-finder` - Raw SKILL.md: https://api.skillmd.com/api/skills/raneqbarber/winning-ad-formula-finder/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: raneqbarber (https://skillmd.com/u/raneqbarber) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/raneqbarber/winning-ad-formula-finder --- # Winning Ad Formula Finder Finds the **combinations** that outperform, and names them so a creative team can brief from them. Creatives do not think "I'll use a soft CTA." They think "bold copy, no product screenshot, soft CTA." That is a formula, and it is the level at which a brief actually gets written. Single-trait findings are the foundation underneath; formulas are the deliverable. The hard part is not finding combinations. It is being honest about what a combination proves. --- ## 1 · Operating contract - **A formula is a classification, not a proof of interaction.** Comparing `A+B` against everything else cannot establish that A and B work together. Never write "synergy" as a demonstrated effect — see §4 Step 3. - **Never publish a formula that failed the validator.** Formulas inherit the tier and robustness tag of the pattern behind them. - **Never brief a formula without its trait DNA.** "Make more like this" is not briefable. The specific trait values are the deliverable. - **Never publish a formula under its trait codes.** `cd_06=bold + vd_02=no` never enters team vocabulary. Coin a real name first — see Step 5. - **Report the ad count on every formula.** A recipe resting on five ads is a hypothesis wearing a name. - **Never claim a formula generalizes beyond the segments it was observed in.** --- ## 2 · When to run this Run after `creative-pattern-validator` on the surviving pair and trio patterns. Feeds `creative-brief-builder` and `performance-guidelines-builder`. Do not run it on unvalidated candidates. A named formula is far stickier than a trait code — once a team adopts "the Confessional Hook," it survives long after the evidence for it has been retracted. Naming an unvalidated pattern is how a false finding becomes doctrine. --- ## 3 · Required inputs and preflight ### Must have | Input | Notes | |---|---| | Validated patterns | Pairs and trios that cleared the validator, with tier and robustness tag | | Per-ad values | For each component trait individually, and for the combination | | Component single-trait results | The individual lift of each component, needed for classification | | `ads_in_pattern` | Minimum 5 | ### Improves the output Spend share, segment breakdown, matched-pair results for the same traits, and the concentration figure per formula. ### Preflight 1. **Validated?** *Halt on anything that did not clear the validator.* Naming precedes adoption. 2. **Component results present?** Classification compares the combination against its strongest single component. Without those, classification cannot run — report the formula unclassified rather than guessing. 3. **Minimum 5 ads.** Below that, report as a hypothesis, do not name it. 4. **Tier and robustness carried through?** They must appear on the formula, not be dropped in translation from the pattern. --- ## 4 · The procedure ### Step 1 — Assemble the recipe Each formula is 2–5 trait values appearing together in at least 5 ads. Record the exact trait DNA — this is what gets briefed. ### Step 2 — Compute group performance From summed totals across the ads carrying the full combination: group lift, spend, ad count, distinct campaigns and months, and concentration. ### Step 3 — Classify against the strongest single component Compare the combination's lift to the lift of its **strongest individual component**: | Class | Condition | What it means for briefing | |---|---|---| | **Amplifying** | Beats its strongest component by >15% | Brief the whole stack — the combination is the insight | | **Additive** | Within ±15% | Traits work independently. Only worth naming when spend justifies it | | **Suppressing** | Falls >15% below its strongest component | Do not combine these. Often the most valuable finding | **Two things must be said plainly in the output.** First, why the baseline is the strongest component and not the sum of components. The additive baseline systematically over-predicts, because traits are correlated — ads with trait A very often carry trait B, so summing double-counts the shared ads. Measured against an additive baseline on a real portfolio, 54% of pairs fell below the line and "antagonistic" became the default verdict, which is an artefact and not a finding. The strongest-component comparison also happens to be the question a creative actually asks: *does adding this second thing help?* Second, and more important: **this classification is descriptive.** Comparing `A+B` against the rest of the segment answers "is this recipe associated with better performance than the segment remainder?" It does **not** establish interaction. Establishing that requires comparing: - `A+B` against `A`-only - `A+B` against `B`-only Until those comparisons exist, "amplifying" is a label, not a demonstrated effect. Where the data supports the component comparisons, run them and say so — that upgrades the claim materially. Where it does not, the output must say the classification is descriptive. ### Step 4 — Type by briefing team | Type | Contains | Briefs | |---|---|---| | **Copy direction** | All copy traits | Copywriters | | **Visual direction** | All visual traits | Designers | | **Creative direction** | Both | The whole team — highest leverage, because it describes a complete recipe | ### Step 5 — Name it Every formula ships as `[rename]` in an editable field. A human coins a real name before publishing. This is not decoration. *"The Confessional Hook"* enters team vocabulary the first time it is used and gets asked for by name six months later. `cd_06=bold + vd_02=no` never will. Once coined, the name is locked for the life of the formula — a renamed formula loses its accumulated meaning. ### Step 6 — Write the brief Per formula: the trait DNA, the classification with its baseline stated, the evidence tier, the segments observed in, the ad count, and what a creative should actually make. Plus, explicitly, what would need testing to move it from tier 3 or 4 to tier 1. --- ## 5 · Computation rules - **Minimum 5 ads per formula.** No exceptions. - **Baseline is the strongest single component**, never the additive sum. Report the additive figure in its own column so the choice is auditable. - **±15% bands** for classification. - **Group performance from summed totals.** - **Tier and robustness tag carry through** from the validated pattern. - **Never publish under trait codes.** --- ## 6 · Output contract ``` FORMULAS · 2026-Q2 · from 47 validated pair and trio patterns classification baseline: strongest single component (additive shown for audit) ⚠ classification is DESCRIPTIVE — see limits below [rename] · creative direction · tier 2 · campaign-robust DNA cs_01 = question + cd_03 = soft + vd_05 = illustrated_diagram evidence 21 ads · 6 campaigns · 5 months · 27% concentration lift +21% strongest component cs_01 alone +14% additive baseline +41% (over-predicts — components co-occur in 11 of 21 ads) class AMPLIFYING — beats strongest component by 100% briefs whole team make an illustrated explainer opening on a reader question, closing soft [rename] · copy direction · tier 4 · campaign-robust DNA cs_02 = outcome_led + cd_04 = 76_150 evidence 34 ads · 5 campaigns · 6 months lift +16% · strongest component +15% class ADDITIVE — traits work independently note named only because it carries 22% of spend [rename] · visual direction · tier 4 DNA vd_06 = heavy + vd_09 = none evidence 14 ads · 4 campaigns lift −31% · strongest component −12% class SUPPRESSING — worse together than either alone make nothing. This is a guardrail: never ship heavy on-image text with no image CTA. LIMITS — read before briefing These classifications compare each combination against the rest of its segment. That does not establish that the traits interact. Establishing interaction requires comparing A+B against A-only and against B-only. Where those comparisons were possible they are shown; where they were not, treat the class as a descriptive label and the formula as a test priority. component comparisons available: 2 of 3 formulas question + soft: vs question-only +14pp · vs soft-only +21pp → interaction supported heavy + none: insufficient ads carrying only one component → descriptive only ``` --- ## 7 · Failure and degraded states | Situation | What to do | |---|---| | Unvalidated pattern | Refuse. Explain that naming precedes adoption and adoption outlives retraction. | | Under 5 ads | Report as a hypothesis with its DNA. Do not name, do not classify. | | Component results missing | Report the formula unclassified. Never guess a classification. | | Component comparisons impossible | Say so per formula. Mark the class descriptive. | | Every formula is suppressing | Check the baseline first — this is the signature of an additive baseline having crept back in. | | High concentration | Report it beside the class. A formula from two campaigns is those campaigns. | | Formula spans one segment only | Say so. It is guidance for that segment, not the portfolio. | --- ## 8 · Worked example `cs_01 = question + cd_03 = soft + vd_05 = illustrated_diagram` — 21 ads, +21% lift. The strongest single component, `question` alone, returns +14%. The combination doubles it, so it classifies **amplifying**. The additive baseline would have predicted +41%, making the same formula look like a *failure* against its own components. It is not. The components co-occur in 11 of the 21 ads, so summing their individual lifts double-counts the same ads. This is why the baseline is the strongest component and why the additive figure is still printed — so anyone can check the choice rather than trust it. Then the honest part. Because the data contains enough ads carrying `question` without `soft` and `soft` without `question`, the component comparisons could be run: +14pp against question-only and +21pp against soft-only. Both positive, so interaction is genuinely supported here, and the output says so. For the `heavy + none` formula there were not enough single-component ads, so its classification stays **descriptive** — a label indicating where to look, not evidence that the two traits fight each other. The recommendation is still useful as a guardrail; it just is not a claim about mechanism. --- ## 9 · Self-check 1. Is every formula backed by a validated pattern with its tier carried through? 2. Is the baseline the strongest single component, with the additive figure shown for audit? 3. Does the output state that classification is descriptive unless component comparisons were run? 4. Where component comparisons were possible, were they run and reported? 5. Does every formula carry its trait DNA, ad count, campaigns, months and concentration? 6. Is every formula shipped as `[rename]` rather than under trait codes? 7. Is each formula typed by the team that briefs it? --- *MIT licensed.*