Kpi Framework Agent Skill

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:

• What single output predicts long-term value? (the north star)
• Which 3-5 controllable inputs cause it? (the driver set)
• What breaks if we over-optimize it? (the guardrails)
• 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.

| metric | type | definition (event + window + denominator) | leading/lagging | owner | baseline | target | target_date |

|---|---|---|---|---|---|---|---|

| Weekly Active Teams | north-star | teams with >=1 member completing a core action in a rolling 7-day window / all active teams | lagging | PM, Activation | 38% | 52% | 2026-Q4 |

| Time-to-first-core-action | input | median minutes from signup to first core action, new teams | leading | PM, Onboarding | 41 min | <15 min | 2026-Q3 |

| Week-1 saved items | input | new teams with >=3 saved items in first 7 days / new teams | leading | PM, Onboarding | 22% | 40% | 2026-Q3 |

| Invites accepted | input | invited members who activate within 7 days / invites sent | leading | Growth | 31% | 45% | 2026-Q4 |

| Support tickets / active team | guardrail | open tickets / weekly active teams (must not rise) | lagging | Support lead | 0.12 | <=0.12 | ongoing |

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.

| north-star / target you push | likely gaming move | guardrail to pair |

|---|---|---|

| Average Handle Time ↓ | agents close tickets prematurely | First Contact Resolution + Customer Effort Score |

| Activation rate ↑ | loosen "activated" definition, count trivial actions | week-4 retention of newly-activated cohort |

| Signups ↑ | buy low-intent traffic | activation rate of new signups |

| Revenue per order ↑ | aggressive upsell, hidden fees | refund rate + repeat-purchase rate |

| Sessions per user ↑ | dark patterns, notification spam | uninstall / unsubscribe rate |

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.

Step 4 — Set the target

This is where frameworks most often break: arbitrary numbers that discourage, or

sandbagged ones that drive nothing.

Baseline before target. Why: you cannot calibrate a target without knowing current

state. "Get to 50%" is meaningless until you know whether you're at 12% or 48%. If there is

no baseline, the first deliverable is "measure the baseline" — do not invent a target on

top of an unknown.

Magnitude must be calibrated — not arbitrary, not sandbagged. Why: targets that are too

ambitious hurt performance through burnout and shortcuts; targets that are trivially safe

drive no improvement. Ground the magnitude in the baseline (a defensible improvement band)

and the levers you actually have, not in a round number that sounds good in a deck.

• Bad: double activation this month
• Good: activation 38% → 52% by 2026-Q4, owner: PM Activation, based on onboarding rework + invite flow

Attach a date and an owner to every target. Why: a target with no date is a wish; a

target with no owner is nobody's job. A row missing either is incomplete.

See references/definition-and-targets.md for baseline measurement, improvement-band

calibration, and why round-number targets invite theatre.

Decision table — is this row a north star, an input, a guardrail, or noise?

| the metric is... | controllable by the team? | tied to delivered value? | → classify as |

|---|---|---|---|

| an output (outcome) | no (you steer it via inputs) | yes, predicts retention/revenue | north star (pick one) |

| an output | partially | yes, but could regress when pushing the NSM | guardrail |

| an input (a play) | yes, directly | causally feeds the north star | input driver |

| a count or output | no | no downstream outcome | noise / vanity — reject |

If a candidate is controllable but doesn't feed the north star, it's a distraction. If it's

tied to value but uncontrollable, it's either the north star itself or a guardrail. If it's

neither controllable nor value-tied, cut it.

Anti-patterns

| anti-pattern | why it bites | the fix |

|---|---|---|

| Vanity metric | grows without value moving; celebrates nothing real | tie to a downstream outcome or reject |

| 40-KPI sprawl | nothing leads, no focus, no owner | one north star + 3-5 inputs, cut the rest |

| Lagging-only | you can watch it but can't act on it | add controllable leading inputs |

| Un-actionable input | team can't influence it through their work | replace with a concrete shippable event |

| Arbitrary target | "double it" discourages or invites gaming | baseline first, then a calibrated band |

| Single number, no guardrail | gets gamed, breaks a neighbor silently | pair every target with a countermetric |

| Raw count as north star | rises with time/size, hides decline | use a rate or ratio tied to value |

| Never re-validated | metric stops predicting value, nobody notices | re-check predictiveness semi-annually |

Re-validation cadence

Re-validate the north star's predictiveness (does it still track retention/revenue?) and

the inputs' controllability (can the team still move them?) at least semi-annually.

Products and portfolios change; a metric that predicted value last year can quietly stop.

Evolve definitions transparently — version the doc, note what changed and why, so a metric

shift never looks like cooking the numbers.

Handoff

When the metric definition doc is complete, route the downstream work:

• Events to instrument (the exact inputs + windows) → ../analytics/SKILL.md
• What to display and how → ../dashboard/SKILL.md
• Recurring narrative around the numbers → ../reporting/SKILL.md
• Designing a test to move a specific input → ../ab-testing/SKILL.md
• Projecting a metric forward in time → ../forecasting/SKILL.md
• Cash/revenue economics, CAC/LTV behind the metric → ../unit-economics/SKILL.md
• Wiring KRs into an operating cadence → ../project-ops/SKILL.md
• Standing analytics models behind it all → ../business-intelligence/SKILL.md