agentleFS
link7373link7373/agentic-bi-team/.claude/skills/experiment/SKILL.md

experiment

Skill4 starsChanged 6 days agoMITView on GitHub

Design or read out a controlled experiment (A/B test) — hypothesis, power analysis, pre-registered primary metric, guardrails, randomisation unit, then sample-ratio and effect-size analysis with a clear ship/no-ship decision. Use to plan a test before launch or to analyse one after it ends.

29 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use cost, decide and 9 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • maxvaegaSkill

    cost

    Estimate what Claude Cowork / Claude Code conversations would cost if run on the pay-per-token Claude API instead of the subscription. Reports token counts and cache-aware API cost in USD and EUR, for the current conversation or for the whole history (cost per conversation). Trigger when the user asks things like "how much would this conversation cost on the API", "API cost of my Cowork conversations", "token usage / token count", "cost per conversation", "/cost", or wants to compare subscription vs API pricing.

    Also in 3 projects that use this

  • dirienSkill

    decide

    Three-agent adversarial debate protocol for strategic decisions. Two advocates steelman opposing positions while a moderator identifies risks and synthesizes a binding consensus. Use when choosing between alternatives, evaluating tradeoffs, or making high-stakes decisions.

    Also in 3 projects that use this

  • javatarzSkill

    wiki

    Access and manage the local GitHub Wiki clone. Use this skill when you need to read wiki content (architecture, design docs, guides), update wiki documentation, check wiki status, or push wiki changes. The wiki contains high-level system design and architectural documentation.

    Also in 3 projects that use this

  • HeshamFSSkill

    loop

    Run a command or check on a recurring interval

    Also in 3 projects that use this

  • VercelSkill

    workflow

    Creates durable, resumable workflows using Vercel's Workflow SDK. Use when building workflows that need to survive restarts, pause for external events, retry on failure, or coordinate multi-step operations over time. Triggers on mentions of "workflow", "durable functions", "resumable", "workflow sdk", "queue", "event", "push", "subscribe", or step-based orchestration.

    Also in 3 projects that use this

  • ibahgatSkill

    autopilot

    Run autonomous multi-stage execution loops until acceptance criteria pass or blockers are reached.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

6 rules. 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 27 edited versions we read keep these.

  • 5

    Distinctive

    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
  1. Distinctive:

    Default to design if no results exist yet.

    line 13
  2. Distinctive:

    Add 1–3 guardrail metrics (things that must not get worse: latency, refunds, churn).

    line 26
  3. Distinctive:

    Variance reduction (CUPED): If the metric has a pre-experiment history per unit, plan to adjust for it.

    line 38
  4. Distinctive:

    Check for interference: If one unit's treatment can change another's outcome — marketplaces, shared inventory, pricing, social features — …

    line 50
  5. Distinctive:

    Check sample ratio mismatch (SRM) first: If the actual split deviates from the intended allocation beyond chance, the assignment is broken …

    line 63
  6. Very specific:

    Close the loop: Update experiments/README.md with the decision, log the ruling in knowledge/decision-log.md, and if the test changed a …

    line 100

“Common” is measured: kept by at least 6 in 10 of the 27 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Experiment — A/B Test Design & Readout

Owner: data-scientist, with metrics-steward on metric definitions and analytics-engineer for the analysis dataset. Args name the test and mode, e.g. /experiment design new checkout flow or /experiment readout pricing-page test.

This skill has two modes. Design runs before launch; Readout runs after the test ends. Default to design if no results exist yet.

Mode A — Design (before launch)

  1. Frame the decision. What ships if it wins? What's the cost of a false positive? If the answer doesn't change a decision, an experiment may be overkill — say so.
  2. State the hypothesis as a directional, testable claim: "Changing X will increase [primary metric] by at least [MDE] for [population]." The minimum detectable effect (MDE) is the smallest change worth shipping, set with the business — not whatever turns out significant.
  3. Pick the primary metric — one — from knowledge/metrics-catalog.md. Pre-register it. Add 1–3 guardrail metrics (things that must not get worse: latency, refunds, churn). Picking the winner after the fact from many metrics is how teams ship noise.
  4. Choose the unit of randomisation (user, account, session, cluster) and confirm the analysis unit matches it. Mismatch (e.g. randomise by user, analyse by session) breaks the statistics. Watch for cross-unit contamination/network effects.
  5. Power analysis. Compute required sample size and run duration from: baseline rate (from the catalog/marts), MDE, significance level (default α=0.05), and power (default 80%). State the assumptions. If the required duration is implausible, the MDE is too small — renegotiate, don't peek early. Two ways to need less traffic, both decided now, not after the result (analytics.md Part 5):
    • Variance reduction (CUPED). If the metric has a pre-experiment history per unit, plan to adjust for it; the sample needed falls by the squared correlation between the pre-period and in-experiment values. Record the covariate and its window in the design.
    • A capped metric for heavy-tailed outcomes (revenue per user): pre-register the winsorisation percentile.
  6. Decide how the test will be monitored. People will look at the dashboard before the end. Pick one and write it down: fixed horizon (interim looks are for guardrails and broken assignment only — no decisions), group-sequential (a stated number of looks with an alpha-spending rule), or always-valid intervals (safe to check any time, at the cost of width). An unplanned early stop on a significant-looking interim result is not a result.
  7. Check for interference. If one unit's treatment can change another's outcome — marketplaces, shared inventory, pricing, social features — randomise clusters or use a switchback, and size the test for the number of clusters, not the number of users.
  8. Write experiments/YYYY-MM-DD-<slug>/DESIGN.md: hypothesis, primary + guardrail metrics (with catalog links), randomisation unit, MDE, sample size, planned duration, the monitoring method and stop rules, any variance-reduction covariate, how secondary metrics and segments will be corrected for multiple comparisons, and what decision each outcome triggers. If the test collects or uses personal data in a new way, run /privacy-review first. For high-stakes or customer-affecting tests, confirm the design with the user before launch. Add a row to experiments/README.md.

Mode B — Readout (after the test ends)

  1. Check sample ratio mismatch (SRM) first. If the actual split deviates from the intended allocation beyond chance, the assignment is broken — stop and diagnose; do not interpret the result.
  2. Profile the data per standards/sql-and-data-standards.md: exposure counts per arm, date coverage, dedup at the randomisation unit, exclusions (and why).
  3. Estimate the effect on the pre-registered primary metric: point estimate with a confidence interval, at the randomisation unit. Report effect size in business terms, not just a p-value. Check every guardrail metric. Use exactly the analysis the design registered: the CUPED adjustment if one was planned (and say how much it narrowed the interval), the sequential boundary if looks were planned. If the metric is a ratio whose unit differs from the randomisation unit (clicks per session, revenue per order), get the variance from the delta method or a bootstrap over randomisation units — a row-level test understates it.
  4. Pressure-test (run the analytics.md Part 1 traps and Part 5): Twyman's law first — a surprisingly large effect is a bug until shown otherwise, so check logging, assignment, and the join before believing it; no peeking-based early calls; secondary metrics and segment cuts corrected for multiple comparisons (Holm for a confirmatory family, Benjamini–Hochberg for exploration) and labelled as hypotheses, not results; segment heterogeneity (does one segment drive it — Simpson's paradox?); novelty/primacy effects; outlier sensitivity on skewed metrics (report robust/median where the metric is heavy-tailed). Distinguish "no effect" from "underpowered."
    • Incremental effect, not just averages. The treated-vs-control gap is the causal lift — the raw treated rate is not. When the effect varies across segments, that heterogeneity is the seed of uplift modelling: the clean control group lets data-scientist model who is persuadable (largest incremental effect) so future targeting spends only where the treatment actually moves the needle.
  5. Decide and write experiments/YYYY-MM-DD-<slug>/RESULTS.md:
    • Verdict: ship / don't ship / inconclusive — with the effect estimate + CI and a confidence level.
    • Guardrail check, segment notes, and any caveats that change the conclusion.
    • Impact if shipped (quantified) and what to monitor post-launch (hand to performance-monitor).
    • For customer-affecting decisions, note the fairness trade-off explicitly (analytics.md Part 1) — equal error rates and equal predictive value can't both hold when base rates differ.
    • The provenance footer (standards/reporting-standards.md). A ship decision goes to analysis-reviewer before it is announced.
  6. Close the loop. Update experiments/README.md with the decision, log the ruling in knowledge/decision-log.md, and if the test changed a metric's interpretation, route it to metrics-steward.

Created by Colin Beck — https://www.linkedin.com/in/beckcolin/

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.