Creates a calibration report from your review data, showing rating patterns by manager and which managers to talk to before calibration.
Analyze a performance review cycle
This skill turns a pile of review files into one report your leadership team can use before and during calibration. Every calibration finding is calculated on your data: managers who rate everyone at the top, managers who give every report the same score, top and bottom ratings with no evidence behind them, and where employees and managers disagree. It leads with the top five findings and the managers to talk to first.
When to use it
- Reviews are in and calibration is next week.
- You suspect some managers rate everyone high but can't show it.
- Leadership wants a summary of the cycle and you have hundreds of written answers.
- You need a promotion shortlist that rests on evidence.
- You want to compare this cycle with the last one.
What you get
One report, as a document to forward or an interactive page to filter during the calibration meeting:
- Top five findings: each with a headline, the numbers, plain words, and who to talk to.
- The manager list: each manager with a flag, the numbers behind it, and one or two neutral questions to ask them before calibration.
- Executive summary: 300 to 500 words.
- Calibration detail: one section per angle, with the threshold, the table, and the finding.
- Cohort overview: leadership, managers, individual contributors.
- What employees are saying: company-wide themes with counts and quotes attributed by department only.
- Promotion-ready shortlist: capped at 15% of headcount.
- Language and bias flags, person by person, and review process and form issues.
- HR-only follow-up: on its own page, not for distribution.
- Methodology: every threshold, what was calculated, and what was judgement.
How it works
- It reads what leadership asked for, if there's a note or email, and uses their words for section names.
- It takes in the review data. Any mix of self, manager, peer and upward reviews, in any format. It works out the structure and the rating scale itself and profiles the data before reading any text.
- It confirms who reports to whom. It sorts everyone into leadership, managers and individual contributors, and shows the breakdown for you to correct before running anything.
- It asks about values or competencies, format, and manager names. Employees are anonymised by default; managers are labelled unless you choose real names.
- It calculates eight calibration angles. Distribution, everyone at the top, no differentiation, lenient and severe managers, compression, extreme ratings without evidence, employee and manager alignment, and completeness. Each shows its threshold so you can argue with it.
- It reads every written answer. It doesn't sample. It clusters themes, counts distinct people, checks what employees raise that managers never mention, and flags review language such as vague praise or recency.
- It verifies before delivering. Every quote is re-checked against the source and every number recomputed.
Calculation and judgement stay apart. The calibration angles are calculations on ratings already in your file. Its own scoring of written comments is a judgement, labelled as such, and never stands in for a rating someone gave.
What it never does. It doesn't estimate: no "most managers" without the number. It doesn't invent a quote. It doesn't infer gender, ethnicity or age from a name. It doesn't give a verdict on a person: every flag is for human review, and a manager who rated everyone at the top may have an exceptional team. Managers with fewer than three rated reports get numbers, not flags. Pay findings are reported as employee sentiment, and decisions on them need HR or legal review. It never sends or shares the report: you decide who sees what.
How to run it
Install the Effy AI plugin for Claude or ChatGPT. One install adds every Effy AI skill, this one included. Then connect the tools you want the skill to read from, such as email, calendar, your drive or your meeting notetaker, and ask for what you need in plain words: the assistant picks the right skill.
Before you start: export the cycle's responses with ratings, comments, reviewer type, manager and team. Replace names with codes if your AI policy requires it. Add last cycle's export if you want the comparison.
- Attach the file and ask, for example: "Effy, please analyze all the collected review responses."
- Confirm the org breakdown and choose the format.
- Read the top five findings and the manager list, and check the HR-only page before you share anything.
Earlier in the cycle, "Audit your review form" improves the questions and "Draft a performance review" helps participants write with evidence.