measurement-experimentation-ops
Skill3 starsChanged 49 days agoMITView on GitHub
Gatekeeper for 'is this result real before scaling': SRM, peeking, contamination, lag and multiple-testing checks; picks causal, screening or infrastructure test mode plus Meta/Google lift tools. When no volume, dying accounts or no rep allow a lift study, prescribes the honest weaker method instead of a fake lift.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- slilbudget/knowledge-delta-skillsword for word, first seen Sep 28, 2026
Other versions1 project uses a version of this.
Word for word · 1
- slilbudget/knowledge-delta-skills★ 3100% same
Edited · 0
None found.
Wrote their own version · 0
None found.
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository17 other files it gives its agents.
- senior-buyer-opsSkill
- tracker-opsSkill
- deep-researchSkill
- knowledge-delta-skill-architectSkill
- web-scrapingSkill
- design-stack-pickerSkill
- normcore-webSkill
- responsive-adapterSkill
- google-adsSkill
- google-feed-opsSkill
- google-grey-opsSkill
- meta-adsSkill
- meta-grey-opsSkill
- js-obfuscationSkill
- secure-codingSkill
- tiktok-adsSkill
- tiktok-opsSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
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
- Distinctive: line 25
Use for high-throughput creative hunting where being fast beats being certain.
- Distinctive: line 26
Never present a screen result as validated.
- Very specific: line 43
SRM: check the RANDOMIZED-UNIT split (a 50/50 arriving 55/45 = broken randomization/logging → invalid) — on assignment counts, NOT on …
≈ “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
Measurement & Experimentation Ops
The other skills act on measured differences; this decides whether a difference is real or noise before they do. Status: unmeasured — no baseline task has been run for this skill; treat sections as unvalidated until one is.
Pick the testing mode by the decision at stake
Three modes, different evidence bars — match to the cost of being wrong:
- Causal: estimates incrementality (no design proves causality without assumptions). Two sub-modes — a randomized experiment (one treatment vs a non-overlapping control/holdout, pre-sized) is the strongest; quasi-experimental causal estimation (GeoLift synthetic control, pre/post) is the fallback when you can't randomize. Use for expensive, hard-to-reverse bets: offer, funnel, landing page, "does this channel even lift sales." Cost: volume + discipline + often a Meta rep.
- Screening (directional): many concepts in one ad set / parallel ABO cells; delivery is UNEQUAL by design, so a "winner" is a hypothesis, not a proof. Use for high-throughput creative hunting where being fast beats being certain. Never present a screen result as validated.
- Infrastructure (isolate infra variance): hold the CREATIVE fixed, vary one infra axis (domain / proxy cluster / account batch) across a balanced set to attribute delivery/ban/CPM differences to infra, not creative. The grey inversion of a normal test — see meta-grey-ops/06.
Feasibility gate (grey reality — check BEFORE promising a clean test)
Causal measurement often isn't available on grey/small-account buys: too little volume to power a holdout, accounts die mid-test, no clean pixel signal, no rep for a sandboxed Conversion Lift. When you can't run causal, SAY SO and drop to the best affordable proxy (geo holdout, pre/post with tracker truth, screening) — label it directional, don't dress a screen up as a lift study. Choosing the honest weaker method beats a "causal" test that's silently contaminated.
Validity traps (each one silently flips a conclusion)
- SRM: check the RANDOMIZED-UNIT split (a 50/50 arriving 55/45 = broken randomization/logging → invalid) — on assignment counts, NOT on spend/impressions/conversions (those diverging is a delivery effect, not SRM).
- Peeking: Meta's A/B "end test early if a winner is found" — leave off and run the pre-set window unless Meta's sequential rule is verified (unpublished) (02).
- Contamination: overlapping audiences between cells — Advantage+ broad bleeding into manual cells; duplicated winners cannibalizing in the auction → not clean groups. Use the A/B tool's non-overlapping split, or geo separation.
- Conversion lag: judging before the payout event matures counts spend against unripe conversions → every fresh cohort looks like a loser. Window ≥ lag; nowcast if you must decide early (tracker-ops/03).
- Multiple testing: screening tolerates chance winners (you re-test anyway); a causal decision needs the bar corrected for the number of comparisons.
- Underpowered: "no significant difference" ≠ "no effect" — size first (01).
Route references
| Need | Reference |
| Sizing (MDE/power as decision rules), SRM, peeking, contamination, lag, inconclusive handling | references/01-experiment-design.md |
Meta tools: A/B Test, ad_study API, Conversion Lift, Brand Lift, GeoLift, Robyn/MMM, Andromeda implication | references/02-meta-measurement-tools.md |
| Google tools: Experiments/drafts, PMax experiments, Conversion Lift, Meridian MMM, brand-search incrementality, plus two 2026 confounders — read before attributing any Google result to your own change | references/03-google-measurement-tools.md |
Buy mechanics → meta-ads (its /09 owns single-account diagnosis & test-design intake) or google-ads (its /08 owns the diagnostic tree and unit economics); this skill owns the validity/incrementality layer above both. Counting & cohort truth → tracker-ops. Portfolio decisions on the result → senior-buyer-ops.
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.

