| name | designing-analytics-projects |
| description | Suggestions for scoping and writing an Analytics Project Brief — the one-page artifact that defines problem, metrics, counter-metrics, stakeholders, methodology, success criteria, and pre-mortem before any analysis begins. Use when the task is to draft, review, or critique a project brief, scope an analytics project, define KPIs, identify counter-metrics or blockers, or prepare a stakeholder map. Not for technical implementation — see ml-modeling, statistical-modeling, or data-warehousing for that. |
Designing Analytics Projects — the Brief
These are suggestions based on the user's CEU MSBA course Designing Analytics Projects (ECBS5228A), taught by Eduardo Arino de la Rubia (ex-Meta, ex-Domino). Source: https://github.com/earino/designing-analytics-projects.
The course centres on one artifact: the Analytics Project Brief. Everything in this skill exists to help write one well.
Cardinal Rule — Use What You're Given
A brief is only as good as its grounding in the scenario.
If the scenario, scoping notes, stakeholder list, data dictionary, prior analysis, or any other local file mentions a number, name, metric, system, or business rule — use it verbatim. Do not invent.
When information is missing, say so explicitly ("not specified in scenario; recommend confirming with X") rather than fabricating a baseline, headcount, KPI, or stakeholder name. Made-up specificity is the single most common failure mode for AI-drafted briefs — it looks professional and is wrong, which is worse than vague.
If the user gives a scenario file, read it first and extract:
- Company name, business model, current metrics (registered users, MRR, NPS, etc.)
- Named stakeholders, their roles, and their stated motivations / KPIs
- Available data tables and known data-quality caveats
- Stated constraints (time, sample size, privacy, can't survey, etc.)
- The exact ask (what was requested vs what's actually needed)
Quote those facts directly in the brief instead of paraphrasing into something more generic.
When to Use
- "Draft / review / critique an Analytics Project Brief" or "project brief" or "scoping doc."
- A scenario file is present and the deliverable is a written brief.
- The user is preparing to talk to stakeholders, kicking off a new analysis, or sanity-checking scope.
- Any mention of: counter-metrics, guardrails vs tradeoffs, Goodhart's Law, pre-mortem, Power-Interest Grid, stakeholder map, decision criteria, "what breaks if we succeed."
Do not auto-trigger for purely technical tasks (training a model, writing SQL, building a pipeline). The brief is the pre-code artifact. Once the brief is approved and you're ready to set up the actual repository, switch to the analytics-project-setup skill.
The 10 Sections (one-line each)
- Problem & Decision — what decision will this inform; who actually decides; why now; one-sentence hypothesis.
- Metrics — primary metric defined SQL-precisely (event/table, grain, eligibility, time window) + 2–3 counter-metrics labelled Guardrail (must not worsen) or Tradeoff (may worsen within bounds).
- Stakeholder Map — Power-Interest Grid (4 quadrants) + named Champions + named Blockers with their motivation (budget / ego / workload / KPI conflict).
- Methodology — 1–3 methods, each tied to a specific hypothesis and the data required, plus Stop/Go data-validity checks.
- Scope & Deliverables — In Scope, Out of Scope (the line that prevents creep), concrete deliverables.
- Success & Decision Criteria — analytical success vs business success, decision forum + action owner, pre-committed decision table ("if we find X, we will do Y; if inconclusive, …"), action thresholds.
- Timeline — milestones with dates, not vibes.
- Risks & Assumptions — assumptions, risks with L/M/H likelihood × impact, mitigations.
- Ethics & Privacy — PII? bias against protected groups? GDPR review? mitigations.
- Pre-Mortem — "It's 3 months from now and this failed. What happened?" Tell the causal story ("we did X, Y happened, because Z"). This surfaces the risks Section 8 misses.
A blank template lives in the course repo and as snippets/brief_template.md in this skill. Use it as the literal scaffold.
Quality Bar — what makes a brief strong
The course's rubric (see syllabus.md) rewards four things, in this order:
- Metric definition precision. Not "conversion rate" but "users with
signup_complete on day 0 → users with ≥1 app_open on calendar day 7, eligible cohort: web signups in last 6 months." If you can't write the SQL, the definition isn't done.
- Counter-metrics that show adversarial thinking. What breaks if we hit the target? Sugar-diet growth, zombie retention, casual-user alienation, brand-trust erosion. Two to three is the right number — five looks like padding.
- Stakeholder analysis that names blockers and their motivation. "Head of Growth" is a placeholder. "Head of Growth — bonus tied to signup volume, will resist any onboarding friction" is analysis.
- A pre-mortem that surfaces non-obvious risks. Not "the data could be bad" — a causal story: "By month 3 the recommendation was shipped, retention didn't move, and the post-mortem found that Learning Paths were correlated with retention because engaged users self-selected into them, not because the feature caused engagement."
High-Leverage Patterns
Counter-metric framing — for each candidate primary metric, ask: what's the laziest way to hit this number, and what would break? Examples from the course cheatsheet:
| Primary metric | Lazy way to hit it | Counter-metric |
|---|
| Conversion rate | Cut the funnel down to power users | Revenue per visitor |
| D7 retention | Spam push notifications | Notification opt-out rate |
| Subscription conversions | Gut free-tier limits | Free-user retention, brand trust |
| MAU | Send re-engagement to dormant users | Engagement depth ("zombie retention") |
| Power-user revenue | Optimise only for top 1% | Casual user satisfaction |
Pre-mortem prompt that consistently produces useful output:
Imagine it's 3 months from now. We shipped what this brief proposes. The project failed — not in a vague way, but specifically. Tell the causal story in 3 sentences: what we recommended, what happened, and the reason it didn't work that we missed today.
Stakeholder Power × Interest in 30 seconds — for each named person:
| High Interest | Low Interest |
|---|
| High Power | Manage closely (weekly updates, pre-brief) | Keep satisfied (don't surprise them) |
| Low Power | Keep informed (channel for advocacy) | Monitor (FYI only) |
Then for each High-Power-High-Interest person, decide: Champion or Blocker? If Blocker, what's the motivation (KPI conflict, budget, ego, workload, prior burn)? Pre-brief privately before any group meeting — no surprises.
Anti-Patterns to Flag in Reviews
- Generic stakeholder list copied from a slide deck (no names, no motivations).
- Primary metric without a SQL-grade definition (no event, no grain, no eligibility, no time window).
- No counter-metrics, or counter-metrics that are just other primary metrics.
- "Explore the data" as the methodology — that's not a project, it's a fishing trip.
- No "Why now?" — without urgency, the brief will not get prioritised.
- Decision criteria written after results are in. Pre-commit, in the brief.
- Pre-mortem as a generic risk list ("data could be incomplete"). It must be a story, not a checklist.
- Inventing numbers ("we estimate ~20% lift") when the scenario gave none. Say "to be confirmed with stakeholder" or use the scenario's stated numbers.
- Treating "Out of Scope" as optional. It's the section that protects you when the ask quietly grows.
Worked Examples in the Repo (read at least two before drafting)
In templates/examples/ there is one fully-worked brief per foundational analysis:
Pick the example whose analysis type matches the scenario at hand and mirror its section depth, table formats, and tone.
The LearnLoop scenario is a strong reference for the depth and style expected on a real submission brief.
Foundational Analysis Cheatsheet (which one fits the scenario?)
Match the business question to an analysis type, then mirror that example brief.
| Question | Analysis | Watch out for |
|---|
| Where do prospects drop off? | Funnel | Cross-device tracking, missing events |
| Who gets credit for the conversion? | Channel Attribution | No "right" model — find where they disagree |
| Did this campaign actually cause the lift? | Campaign Effectiveness | Correlation ≠ causation; pull-forward; contamination |
| Are unit economics healthy? | CAC / LTV | Blended CAC hides bad channels; LTV on margin not revenue |
| Do users come back? | Retention | "Sugar-diet growth" hides churn; correlations ≠ drivers |
| Who are the heaviest users and why? | Power User | Don't alienate the casual majority |
| What's broken / what are we losing? | Failure Analysis | Manual sampling first; size by impact, not volume |
| Why do users upgrade / pay more? | Expansion & Monetisation | Free-user churn from over-monetisation |
| Do products help or cannibalise each other? | Ecosystem | Selection bias on multi-product users |
Source: cheatsheet.md in the course repo (the one-A4 exam cheat sheet — concentrated wisdom).
Suggested Workflow When Drafting from a Scenario
- Read the scenario twice. Highlight every number, name, system, and constraint.
- Pick the foundational analysis (table above). Open the matching example brief side-by-side.
- Draft Section 1 (Problem & Decision) first and resist the urge to skip to methodology. If you can't name the decision and the decision-maker, the brief is not ready.
- Write the metric definition next, SQL-precise. If the data dictionary doesn't support it, say so.
- Add counter-metrics by asking "what breaks if we hit this number?" Two or three.
- Stakeholder map from the named people in the scenario (do not invent roles). For each High-Power person, decide Champion vs Blocker and the motivation, quoting the scenario where possible.
- Methodology + Stop/Go checks. What must be true before you trust the result?
- Decision criteria pre-committed before you skip ahead to Pre-Mortem.
- Pre-mortem last — it depends on everything above.
- Re-read the scenario. Did you contradict any stated fact? Did you invent any number? Fix.
From Brief to Technical Setup
Once the brief is approved, the next step is technical project scaffolding. Use the analytics-project-setup skill to:
- Create the analytics repository.
- Create the folder structure (
dev/prod split, data/, numbered notebooks).
- Set up shared config/constants,
.gitignore, pre-commit hooks.
- Write
README.md using the structured template in the setup skill.
- Create
AGENTS.md so AI agents understand the project context.
The brief defines what you're building and why. The setup skill handles how the repo is structured.
Further Reference
- Course repo (the source of truth): https://github.com/earino/designing-analytics-projects
- Designing Experimentation Guardrails, Airbnb Engineering — the canonical counter-metrics reference.
- Getting to Yes (Fisher, Ury, Patton), ch. 1–3 — for the influence / blocker chapters of any brief.
- Goodhart's Law: "When a measure becomes a target, it ceases to be a good measure." — the reason counter-metrics exist.
- Géron, ML project checklist — his "Frame the problem" section is a complementary checklist to the Analytics Project Brief.
Companion skills:
analytics-project-setup — technical scaffolding, folder structure, AGENTS.md, environments. Use after the brief is approved.
ml-modeling — for building and evaluating predictive models.
statistical-modeling — for inferential/explanatory analysis.
data-warehousing — for data pipeline architecture.
Suggestions, not gospel. The single highest-leverage habit: quote the scenario, don't paraphrase it; flag missing info, don't invent it.