Draft OKRs That Pair Every Metric with the Counter-Metric That Keeps It Honest

Workflow area: Delivery & Reporting · Tier: Core · Hands-on time: ~30 minutes

Goodhart's law — when a measure becomes a target, it ceases to be a good measure — is not a risk in engineering OKRs; it is the default outcome. "Reduce incident count" teaches teams not to declare incidents. "Increase deploy frequency" produces confetti deploys. AI is genuinely good at drafting crisp OKRs, and even better at something leaders rarely ask it for: playing the smart, cynical team that will game them. That adversarial pass is what this playbook adds to a chore you already do quarterly.

When to use — and when not to

• Use it every planning cyclequarterly or half-yearly OKR drafting is where an hour of adversarial review saves a quarter of measuring the wrong thing.
• Use it on OKRs handed down to yourunning the alignment and gaming critique on your VP's draft is how you negotiate targets with evidence instead of vibes.
• Use it when a metric already smells gamedthe red-team prompt works retroactively: feed it the KR and the behavior you're seeing, and it will usually name the incentive you built.
• Skip it for team-health and learning goalsnot everything belongs in a measured KR; forcing mentorship or morale into numbers is how you Goodhart the culture itself.
• Skip the counter-metric machinery for genuinely low-stakes KRsa counter-metric per KR on a two-person team's internal-tools objective is process for its own sake; reserve the rigor for KRs that drive behavior.

The recipe

1. 1.

   Assemble the inputs. You need three things: the org- or business-level goals for the period (verbatim, not your paraphrase), your team's mission and current commitments, and last cycle's OKRs with their actual outcomes. Ten minutes of collection determines the quality of everything downstream.

2. 2.

   Draft candidate OKRs with alignment stated explicitly.

   Copy

   Draft OKRs for my team for [PERIOD].
   
   Org goals (verbatim): [PASTE]
   Team mission and context: [PASTE]
   Last cycle's OKRs and outcomes: [PASTE]
   
   Produce 2–3 Objectives, each with 2–4 Key Results. For every KR:
   - the metric, its current baseline (write [BASELINE NEEDED] if
     I gave you none), and the target
   - WHERE it is measured (which system produces the number) and
     WHO owns reporting it
   - one line tracing it to a specific org goal above. If you
     cannot trace it, mark [ALIGNMENT GAP] — do not invent a link.
   
   Outcomes over output: no KR may count activity (tickets closed,
   PRs merged) unless the activity is the business outcome.
   

3. 3.

   Run the measurability critique. Follow up in the same conversation: "Critique each KR as a measurement: is the number observable in a named system today, is the baseline real, can we tell success from failure at quarter end without argument? Flag any KR that needs a new dashboard to exist." Kill or fix anything that fails — a KR nobody can measure by week 3 becomes a KR everybody ignores by week 6.

4. 4.

   Red-team for Goodhart, then pair counter-metrics.

   Copy

   Now act as a smart, cynical team that will be graded on these
   KRs. For each one:
   1. GAMING VECTOR — the easiest way to hit the number while
      betraying the objective's intent.
   2. COUNTER-METRIC — one paired metric that gets worse when the
      KR is gamed, measured in a named system.
   3. Verdict: SAFE (hard to game), PAIR (needs its counter-metric
      reported alongside), or REDESIGN (invites the wrong behavior).
   

   Good output looks like:

   Copy

   KR: Reduce p1 incident count 30%
   GAMING VECTOR: reclassify p1s as p2s; discourage declaring.
   COUNTER-METRIC: customer-reported issues + mean severity of
   declared incidents (PagerDuty + support tickets).
   Verdict: PAIR — report both numbers together, always.
   
   KR: Cut PR review turnaround to <4h
   GAMING VECTOR: rubber-stamp approvals, review depth collapses.
   COUNTER-METRIC: post-merge defect escapes + revert rate.
   Verdict: PAIR
   

5. 5.

   Human pass: cut to what you'll actually drive. The model happily produces five objectives; a team drives two or three well. Cut, sanity-check every target against baselines yourself (models cheerfully propose "improve 30%" against baselines they never saw), and confirm each counter-metric will genuinely be reported alongside its KR — a counter-metric nobody publishes is a design document, not a guardrail.

6. 6.

   Publish the pairs, not just the KRs. Wherever the OKRs live, the counter-metric appears in the same row as its KR, and every review of the KR reviews both numbers. Announcing the gaming vectors to the team is optional but powerful: it says you know how the metric bends, and you're watching the pair.

Make it yours

• Small team or startup: one objective, three KRs, counter-metrics only where gaming is plausible. Your risk isn't gaming — it's KRs disconnected from survival metrics; weight the alignment check heaviest.
• Large org or GCC: run the alignment prompt against both the global org goals and your site or account-level goals — GCC teams serving a parent org fail OKRs by laddering to only one of the two. Standardize the KR + counter-metric pair format across leads so roll-ups stay comparable.
• Platform and infra teams: your customers are internal, so draft KRs from consuming-team outcomes (adoption, migration completion, toil removed) and let the red-team prompt specifically attack vanity platform metrics like "number of services onboarded."

Quality check

• Every KR traces to a real org goalany surviving [ALIGNMENT GAP] means you're about to spend a quarter on something the org didn't ask for; that's a conversation, not a KR.
• Baselines are real numbers you verifieda target without a baseline you checked is a wish with a percentage sign.
• Every PAIR verdict shipped with its counter-metriccount them; the gap between designed pairs and published pairs is your actual Goodhart exposure.
• No activity metrics wearing outcome costumesreread each KR asking "could we hit this while customers notice nothing?"; if yes, redesign.
• You can name who reports each number and where it comes fromownership ambiguity in week 1 is measurement theater by week 6.