Kpi Framework

KPI framework

You decide what to measure. You do not build the dashboard, you do not wire up the events, you do not write the monthly report. Your deliverable is a metric definition document: one north-star metric, a small set of input drivers that causally feed it, paired guardrails, and a calibrated target with a baseline and a date.

Most measurement work fails upstream, before any chart exists. Teams instrument 40 KPIs and none of them lead anywhere. They optimize a lagging output nobody can move. They celebrate a vanity number. They set "double it this month" and watch it get gamed. Your job is to kill those failures at the source by forcing four decisions:

1. What single output predicts long-term value? (the north star)
2. Which 3-5 controllable inputs cause it? (the driver set)
3. What breaks if we over-optimize it? (the guardrails)
4. What target — baseline, magnitude, date — is honest and ungameable?

Answer those four and hand the result to ../analytics/SKILL.md to instrument and ../dashboard/SKILL.md to display. If you find yourself choosing chart types or writing SQL, you have left this skill.

The one artifact

Everything you produce collapses into a single table. Nothing leaves this skill with the baseline, target, or date column blank — an unfilled target is a decision you skipped, not a decision you made.

The columns are not decoration. "Definition" must be unambiguous enough that two analysts querying independently get the same number — that means a concrete event, a time window, and a denominator. See references/definition-and-targets.md for how to write definitions that don't drift.

Step 1 — Pick ONE north star (the output)

The north star is an output / lagging metric. Why: it's the scoreboard for value delivered, deliberately too broad to act on directly. You don't push the north star — you push the inputs and watch the north star move. One per team; more than one means no team actually owns the outcome.

Express delivered value as a rate or ratio, not a raw count. Why: raw counts grow with time and headcount and hide health — "total users" goes up even as the product dies.

• Bad: total registered users
• Good: weekly active teams that completed a core action / all active teams

It must predict long-term retention or revenue. If the number can climb for a quarter while the business erodes, it is not a north star. The test: would you bet next year's retention on this number rising? If not, keep looking.

Vanity reject test. Followers, page views, likes, total signups — vanity unless tied to a downstream outcome (conversion, revenue, retention). 10k followers with zero sales lift is the canonical example. If a candidate metric can double with no change in value delivered, reject it and say why in the doc.

Source candidates from a lens — AARRR (acquisition/activation/retention/referral/revenue) or HEART (happiness/engagement/adoption/retention/task-success) — then narrow to one. references/metric-catalog.md lists candidate north stars and driver sets per business type (SaaS, marketplace, content, e-commerce, B2B sales-led).

Step 2 — Build the driver set (3-5 inputs)

The north star is the scoreboard; the inputs are the plays you actually run.

Each input is leading, directly controllable, and a concrete instrumentable event. Why: if the team can't influence it through their own work, it's not an input — it's another output, and chasing it is vanity. "Engagement" and "satisfaction" are not inputs; they're abstractions you cannot ship against.

• Bad: increase engagement
• Good: % of new teams with >=1 saved item in the first 7 days

Each input must plausibly cause the north star. Why: a metric tree connects every node to its parent (the outcome) and its children (the inputs). A standalone number has no defense against gaming; in a tree, gaming one node shows up as distortion in its neighbors. Draw the tree so the causal claim is explicit and falsifiable:

        Weekly Active Teams (north star, output)
        /              |                 \
Time-to-first      Week-1 saved        Invites accepted
core action        items (>=3)         within 7 days
(leading)          (leading)           (leading)

Cap the set at 5. Why: more than five inputs is sprawl — focus dilutes, nobody owns the list, and you're back to the 40-KPI swamp you came to escape. If you have eight candidates, the work of this step is cutting three.

Hand the final event list — exact events, windows, denominators — to ../analytics/SKILL.md to instrument. You define them; analytics implements them.

Step 3 — Add guardrails / countermetrics

"When a measure becomes a target, it ceases to be a good measure." — Goodhart's Law (Charles Goodhart, 1975)

Single-metric optimization gets gamed. Optimize sales volume alone and reps discount to the floor; optimize Average Handle Time alone and agents hang up on unsolved problems.

Every target gets a paired shadow metric representing the foreseeable harm. Why: the guardrail is what catches the gaming before it costs you. The pair must measure the thing that breaks when someone over-optimizes the target.

A guardrail does not need a stretch target — its target is usually "must not get worse than baseline." Write it into the table anyway, with ongoing as the date.