Rendu depuis le dépôt source en conservant titres, exemples, code, tableaux, liens et images.
Campaign analysis
The question this tier answers: is this burning money past the point where you should stop, or is it just early? Nothing here judges which creative wins, how much to scale, or why performance changed — that's this same skill's paid tier. This tier's whole job is loss prevention: catch the campaign that's clearly past its stop-loss threshold, and be honest when the data just isn't there yet to say more.
Before you start
- Tracking health first. If pixel or conversion tracking is broken, every threshold below is a decision made on garbage data. Route to
tracking-healthand don't proceed until it comes back clean. - Read `.agents/profile.md` (`../contracts/profile-v1.md`) for Margin Basis and Target CPA. Either missing → ask the user directly; when you apply the break-even formula in `references/stop-loss-thresholds.md`, name exactly which inputs you used, not just the output number.
- Read `.agents/data/normalized/` and `.agents/data/readouts/` before asking for anything — see Data layer, below. The same export doesn't get asked for twice.
Absolute stop-loss threshold
Compute the break-even CPA/ROAS against Margin Basis — formula and the gross-margin-basis vs net-profit-basis distinction in `references/stop-loss-thresholds.md`. When mature spend (past the attribution window and any reporting lag — a "zero conversions" reading inside that window isn't a real zero) has run far past what Target CPA implies the conversions should be, recommend stopping.
Fixed wording: "stop-loss on absolute threshold." Never "this angle doesn't work" or "this creative failed" — those are comparative verdicts this tier has no evidence to support (see Boundary, below). The absolute threshold compares spend against an external number — target CPA, tolerable loss — never against another ad or creative, so it doesn't wait on the sample size a comparative verdict would need.
"Wait ≠ not working"
When conversion volume is too thin to compare this ad or creative against anything else, exactly two conclusions are legal: wait (let it keep running to accumulate signal) or restructure to get volume (consolidate audiences, placements, or budget so it accumulates faster). "This isn't working" is a comparative verdict — this tier doesn't have the evidence for it.
Thin volume doesn't override the absolute threshold above. A campaign can be simultaneously "too little data to compare against another creative" and "already past its stop-loss threshold" — those are two different questions with two different answers, not a contradiction.
Boundary: what this tier doesn't judge
Comparative verdicts — which creative wins, how much to scale, fatigue detection, breakpoint diagnosis — belong to this same skill's paid tier, not shipped here. Say so plainly: "this tier stops at loss-prevention; a comparative verdict needs the paid tier." Then point at where the problem likely is, using the table below — a free user still gets to know what's wrong, just not a verdict on which creative to keep.
Problem routing
| Symptom | Where the problem likely is | What this tier gives |
|---|---|---|
| Clicks are high, conversions aren't happening | Landing page | Locates the problem only — full diagnosis is landing-page-cro (paid) |
| Numbers look implausible (too good, too flat, zero when spend is high) | Tracking | tracking-health |
| Signs of creative fatigue (rising CPM, falling CTR at stable targeting) | Creative | ad-creative |
| Creative, delivery, and landing page all ruled out | Angle layer | Name it plainly as out of scope for this tier — don't diagnose the angle yourself, refer back to it |
Data layer
When the user drops a platform export, land it through the three layers in `../contracts/agents-dir-conventions.md`:
- `.agents/data/raw/` — the export file itself, untouched.
- `.agents/data/normalized/` — reshaped into one schema per data type; record which platform, which fields, and the date range the export covers, so a later read doesn't have to re-guess the source. An ad name, URL, or other free-text field carried forward from
raw/goes through `../shared/references/input-hygiene.md` first — a platform export is external content like any other. - `.agents/data/readouts/` — this skill's own conclusion, dated, naming which normalized file it came from.
Read `data/normalized/` and `data/readouts/` before asking the user anything. A second call against the same export asks only for what's still missing.
Calibrated Thresholds write-back. Per `../contracts/profile-v1.md` and `../contracts/sister-product-compat.md` §4, this skill is the sole writer of profile.md's Calibrated Thresholds section — write back only after a calibration pass against real spend/conversion data, touching only that section; every other part of profile.md stays read-only.
Data source: three-tier degrade
- Read-only ad-platform MCP tool available → pull recent data with it, and land it into
.agents/data/through the same raw/normalized/readouts layout above — an MCP pull isn't a shortcut around that split. - No MCP tool → read
.agents/data/. Follow../shared/references/mcp-setup/to wire one up — this reference ships in every tier; a stripped-down install missing it just means no wiring guide, skip to the manual path below. - Nothing in `.agents/data/` either → ask the user to export from the platform's dashboard or paste a summary.
Read-only, no exceptions. Even when a write-capable tool is available — budget changes, pausing an ad, audience edits — never call it. Every conclusion stays a recommendation the user acts on themselves, whether or not the calling agent has account-write tools wired up.
Recommendations only
State every conclusion as a recommendation — "recommend pausing this ad," "recommend revisiting Target CPA on [date]" — never execute it. This tier doesn't touch creative production or delivery settings either; it only judges spend against thresholds. What to do about a verdict routes to ad-creative or ads; this skill stops at the verdict.
This skill's own write posture doesn't change just because a narrower one exists elsewhere in this package. ads and ad-creative may create a new, always-paused object through a write-capable MCP tool (see `../shared/references/mcp-setup/index.md`'s policy) — this skill still never calls a write-capable tool for anything, full stop.
Related skills
In this package: quick-angle — the angle layer this tier names as out of scope but never diagnoses itself (Problem routing, above). tracking-health — run first, the data floor under every judgment here. ads — where a verdict's "what to do about it" routes when the fix is a delivery-settings change. ad-creative — creative fatigue routes here. landing-page-cro (paid) — click-no-conversion routes here for diagnosis.
