Draft a North-Star Metrics Tree with Gaming-Risk Critique

Workflow area: Strategy & Roadmap · Tier: Foundation · Hands-on time: ~30 minutes

Choosing a north star is a judgment call no model should make. But two parts of the work are mechanical, and AI does them well: enumerating candidate metric trees exhaustively, and red-teaming each candidate the way a bonus-hungry team eventually will. Most metric disasters — support teams closing tickets without resolving them, engagement metrics rewarding rage — were foreseeable with one adversarial pass nobody ran.

When to use — and when not to

• Use it before the metrics workshoparrive with three critiqued candidate trees and the meeting debates trade-offs instead of brainstorming from zero.
• Use it on your existing metricsrunning your current north star through the gaming critique is twenty minutes and routinely finds an incentive problem you've been living with.
• Use it when metrics cascade to teamsthe input-metric layer is where gaming concentrates, because that's the layer tied to individual teams' goals.
• Skip the AI for the final selectionwhich trade-off to accept is a strategy decision owned by humans who understand the business model.
• Skip it if leadership hasn't agreed what the product is fora metrics tree can't resolve a strategy dispute; it will only formalize whichever side wrote the prompt.

The recipe

1. 1.

   Brief the model like a new hire. Product, business model, stage, and how the product creates value:

   Copy

   Product: [ONE PARAGRAPH — what it does, who pays, how]
   Business model: [e.g. "B2B SaaS, seat-based, land-and-expand"]
   Stage: [e.g. "post-PMF, 40K WAU, scaling"]
   Value moment: [WHEN A USER ACTUALLY GETS VALUE — e.g. "an expense
   report approved without manual chasing"]
   
   Propose 3 candidate north-star metrics. For each, build a tree:
   - North star (leading indicator of the value moment, not revenue)
   - 3-5 input metrics that mechanically drive it
   - For each input: which team could own it, and how it's measured
   
   For each candidate, state in one line what behavior it rewards and
   what it ignores. Do not recommend one yet.
   

2. 2.

   Pick the two trees worth attacking. Discard any candidate that's revenue in a costume (revenue lags value; a north star should lead it) or that your instrumentation can't actually measure — check each proposed metric against events that exist in Amplitude or your warehouse, not events the model assumed.

3. 3.

   Run the gaming-risk critique — the heart of this playbook.

   Copy

   Now attack this metrics tree as three adversaries:
   
   1. THE GAMING TEAM: a team bonused on each input metric. For each
      metric, what's the cheapest way to move it WITHOUT creating real
      user value? Be concrete and devious.
   2. THE PROXY SKEPTIC: where does each metric diverge from the value
      it proxies? Describe a realistic scenario where the metric goes
      up while users get less value.
   3. THE LONG GAME: what does 2 years of optimizing this north star
      do to the product? What quietly degrades because nothing in the
      tree measures it?
   
   For each attack, rate severity (high/med/low) and say whether a
   guardrail metric could catch it — and which one.
   
   Tree: [PASTE THE CANDIDATE TREE]
   

4. 4.

   Read for the attacks that sting. Good output looks like:

   Copy

   GAMING — input: "onboarding completions/week" (HIGH)
   Cheapest move: shrink onboarding to 2 trivial steps. Completions
   spike; activated-user quality drops. Guardrail: week-4 retention
   of onboarded cohort.
   
   PROXY SKEPTIC — north star: "weekly approved reports" (MED)
   Auto-approval rules inflate the count with zero manual-chasing
   pain relieved. Metric up, value flat. Guardrail: % approved
   without any reminder sent.
   

   Attacks that make you wince are the deliverable. For every high-severity attack, either add the guardrail metric to the tree or consciously accept the risk in writing.

5. 5.

   Choose and hand-write the rationale. Pick the surviving tree yourself and write the one-paragraph "why this metric, what we're accepting" note in your own words. Bring tree, attack log, and guardrails to the workshop — the team debates and finalizes; the AI never picks.

Make it yours

• Two-model adversarial split: generate trees in ChatGPT, critique in Claude (or the reverse) — a model is measurably more willing to find flaws in a tree it didn't write.
• Amplitude reality check: paste your actual taxonomy's relevant event names into step 1 so every proposed metric maps to instrumentation that exists; a north star you can't measure by Monday is a poster, not a metric. Wire the winner into a dashboard the same week.
• Team scale: before the workshop, send each lead one candidate tree plus the prompt from step 3 and have them run their own attack. Merged attack logs beat any single reviewer, and the exercise teaches the team to think in guardrails.

Quality check

• Every metric maps to a real eventverified against your instrumentation, not the model's assumption of what you track.
• Each input mechanically drives the north startell the causal story for every edge; correlation-shaped trees fall apart at cascade time.
• Every high-severity attack is resolvedguardrail added or risk accepted in writing; an unaddressed high-severity gaming path is a scheduled incident.
• The north star leads revenue, not repeats itif it moves only when revenue moves, it's a lagging indicator wearing a north-star badge.
• The rationale is in your voiceif you can't defend the chosen trade-offs without the AI's text in front of you, you haven't chosen yet.