AI-Assisted Performance Review Skill
The uncomfortable review question of the decade: when a report ships twice the output with AI, what did they do? Volume stopped measuring effort; polish stopped measuring skill. Punishing AI use is as wrong as crediting the model's work to the human. This skill separates the signals — and gives managers the conversation, not just the theory.
What This Skill Produces
- A what-measures-whom analysis of the role's current evaluation criteria
- Rewritten criteria that measure the human: judgment, verification, outcomes, leverage
- Calibration rules for teams with uneven AI adoption
- Conversation scripts for the three hard cases
Required Inputs
Ask for (if not already provided):
- The role and current review criteria (the rubric, or how it really works)
- How AI shows up in the work — which tasks, how much of the output it drafts, what the tooling reality is
- The specific situation, if any: one person's review? team calibration? criteria rewrite?
- The org's AI stance — encouraged? tolerated? policy exists? (Reviews must not punish sanctioned behaviour)
Method
- Sort every criterion: human, tool, or hybrid. Walk the current rubric. Volume of drafts, formatting quality, speed to first version → now mostly tool signals (evaluating them evaluates prompt luck and subscription tier). Decision quality, stakeholder trust, error catch rate, what they chose to build → still human. Output quality overall → hybrid: credit belongs to the pair, and the review's job is to see the human's contribution inside it.
- Rewrite around the four durable human signals:
- Judgment — what they decided to do, what they declined, how they scoped; the quality of taste applied to AI output (what they kept, cut, and corrected)
- Verification — do errors get caught before shipping? A person whose AI-assisted work is reliably right is demonstrating skill; one who forwards unverified fluency is a risk wearing productivity's clothes
- Outcomes — did the work move what it was for (the metric, the decision, the customer), independent of how it was produced
- Leverage — do they make AI multiply the team (shared prompts, workflows, teaching) or only their own count
- Set the calibration rules for mixed adoption. In one team you'll have a 2×-output adopter and a careful non-adopter. Rules that keep it fair: evaluate against the role's outcomes, not each other's volume · where AI use is sanctioned, not adopting is a development conversation (not a values one) · where someone's edge is invisible verification labour, surface it explicitly before comparing. Never let the review become a proxy war about the tools.
- Demand evidence that sees the human. Volume anecdotes are out. In: a sample of shipped work walked backwards (what did the AI draft, what did you change, why) · error/rework history · decisions log · peer signals about trust and leverage. The walk-backwards exercise is the single highest-signal artifact — put it in the review prep.
- Script the three hard cases:
- The volume star with thin judgment — "Your output doubled; let's walk three pieces backwards" (the conversation is about the delta between draft and shipped)
- The careful sceptic being out-shipped — outcomes-first framing; adoption raised as growth, not deficiency; their verification strength named as a strength
- The launderer — unverified AI work shipped as their own, errors reaching others: this is a reliability conversation with the accountability rule from the org's AI policy, not an AI conversation
Output Format
AI-Era Review Guidance: [role/team]
Criteria audit
| Current criterion | Measures | Verdict |
|---|---|---|
| human / tool / hybrid | keep / rewrite / kill |
Rewritten criteria: [the judgment/verification/outcomes/leverage set, with observable definitions each]
Evidence to collect: [the walk-backwards sample protocol + the rest]
Calibration rules: [the mixed-adoption rules, as committee guidance]
The conversations: [scripts for the three hard cases, adapted to the situation given]
Quality Checks
- Every current criterion has a human/tool/hybrid verdict — none skipped as "obviously fine"
- New criteria are observable behaviours, not virtues ("catches errors before shipping" not "is diligent")
- Verification labour is explicitly valued somewhere — the invisible work made visible
- Calibration rules prevent both punishing adoption and punishing non-adoption
- The launderer case routes to reliability/accountability, not to relitigating the AI policy
Anti-Patterns
- Do not credit or blame the human for what the model did — walk the work backwards to find the human
- Do not keep volume metrics "because they're objective" — they're objective measurements of the wrong thing now
- Do not run calibration comparing raw output across uneven adopters — that's a tooling lottery, not a review
- Do not treat AI scepticism as a performance problem where use is optional — outcomes are the bar, not enthusiasm
- Do not have the accountability conversation without the org's policy in hand — improvised rules in a review are how grievances are born
