ad-optimization
Skill1 starsChanged 3 months agoMITView on GitHub
Turn Meta Ads findings into specific, safe, approved changes — kill/scale/iterate decisions, budget shifts, creative swaps, fatigue fixes, learning-phase-aware scheduling of edits. Use when the user wants to improve results, decide what to pause or scale, reallocate budget, fix creative fatigue or rising CPA/CPL, or asks "what should I change?
In this repository4 other files it gives its agents.
- ad-analysisSkill
- campaign-managementSkill
- opportunity-spottingSkill
- pulseSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 11
Read-only by default: Every live change needs the user's explicit approval of that specific change.
- Distinctive: line 12
Never fight the optimizer: No pausing segments off breakdown-average CPA (see the ad-analysis skill's breakdown-effect reference).
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Meta Ads Optimization
Optimization is an operating system, not a bag of tricks: a target number, a decision tree run on a cadence, and change discipline that never fights the delivery optimizer. Every change is proposed → approved → executed → logged → checked.
Prime directives
- Read-only by default. Every live change needs the user's explicit approval of that specific change. Execution details and hard limits:
references/safety-governance.md— load it before executing anything. - Never fight the optimizer. No pausing segments off breakdown-average CPA (see the ad-analysis skill's breakdown-effect reference); no daily knob-fiddling; batch edits; respect learning phase.
- Every swap needs a replacement. Pausing without replacing = shrinking. What to build in place of each failure type is in the decision tree.
- Log everything. In a workspace with
optimization/CHANGELOG.md, append every executed change immediately: date, entity, change, reason, expected outcome, check date.
The target number: TCPL
Optimize to target cost per qualified lead/result, not raw CPL. Derive (pick the strongest available, label the source):
- From economics:
TCPL = target cost per sales-meeting × (qualified-lead → meeting rate)— or for e-comm, breakeven CPA = AOV × margin. - From history:
TCPL = trailing-30-day cost per qualified lead × 0.80(a stretch, not a fantasy). - Both available → take the MIN.
Raw platform leads lie; qualification lives in the CRM. If the workspace
_config/account.mdhas targets, use those.
Derived thresholds (from TCPL)
| Rule | Formula | Meaning |
| Minimum data before judging | spend ≥ 3 × TCPL | Below this, WAIT (at 3× with zero results, ~95% confidence it's not luck) |
| Swap line | cost/QL > 1.5 × TCPL | Persistent breach → swap the ad |
| Kill line (hard) | spend ≥ 3 × TCPL and 0 conversions | Pause now (still requires ≥7 days or ≥20 clicks unless spend is runaway) |
| Ad-count ceiling | (daily budget × 14) ÷ (2 × TCPL) | More live ads than this = starvation; sweet spot 6–10 |
| Tests per week | (monthly budget × 0.20) ÷ (3 × TCPL) ÷ 4 | Sustainable test velocity |
| Learning budget | daily ≥ (TCPL × 50) ÷ 7 per ad set | Else knowingly learning-limited |
Weekly cadence
- Monday — decision day. Pull rolling 14-day data (
--daily), run the decision tree (references/decision-tree.md) on every active ad, execute approved swaps/graduations, log. - Wednesday — test day. Launch the week's new tests (paused → approved → live); run day-7 delivery checks on last week's launches.
- Friday — budget day. Scaling moves, rollbacks, reallocations per
references/scaling-budget.md. - Monthly — full review against TCPL trend; re-derive TCPL if the funnel shifted.
Decision quick card (full tree in references/decision-tree.md)
- Spend < 3×TCPL → WAIT
- Delivery ~zero after 7 days (CBO chose not to spend it) → KILL (the auction rejected the creative; rebuild hook/visual, not copy)
- Zero platform leads at 3×TCPL → SWAP, abandon the concept
- Leads but ~zero qualified → SWAP, keep format, add ICP-filtering specificity
- Cost/QL ≤ TCPL → winner track; 1–1.5× → MONITOR; >1.5× → SWAP (same angle, new execution)
- Graduate to scaling only when ALL: ≥5 QLs · QL rate ≥60% · cost ≤ TCPL · ≥14 days live · ≥1 QL in last 7 days
- Fatigue (scaling ads): frequency 3.0–3.5 or cost +20% → build 2 iterations now; >3.5 or +40% → swap immediately
Scaling & budget rules (details in references/scaling-budget.md)
- Vertical: +20% per 3–5 days (30% absolute max) — bigger jumps reset learning. Rollback if cost >1.5× target after a step: cut 20–30%, wait 7 days, then scale horizontally instead.
- Portfolio: ~70% proven / 20% promising / 10% experiments; structural 80/20 scaling-vs-testing campaign split.
- Learning-limited remediation ladder: broaden audience → raise budget → optimize a higher-funnel event → consolidate ad sets.
Change hygiene
- Significant edits (budget >~20%, targeting, creative, optimization goal) reset learning — batch them into one session per ad set, never a drip.
- One variable at a time on anything you call a test.
- Set the check date (3–7 days) at execution time; on the check date, verify against the expected outcome and close the changelog loop.
References
references/decision-tree.md— full two-stage tree, swap-replacement rules, graduation, fatigue playbook, quality scoring.references/scaling-budget.md— scaling protocols, kill rules with data gates, 70/20/10, seasonality, measurement hierarchy.references/safety-governance.md— mandatory before any live execution: approval gates, paused-by-default, rate limits, account-ban failure modes.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

