Draft Performance Reviews With Claude Without Exposing Staff Data

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Draft Performance Reviews With Claude Without Exposing Staff Data.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Draft Performance Reviews With Claude Without Exposing Staff Data.

A manager can draft performance reviews with Claude by keeping dated, factual notes per person, removing names and health or personal details, and working in a dedicated Project with memory kept separate. Claude turns the notes into balanced draft wording against your competency framework; the manager checks it for bias, sets the rating and owns every word.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

The line between drafting help and an AI decision matters more than any setting. Anthropic's usage policy lists evaluating a worker's performance among its high-risk areas: when Claude's recommendation shapes such a decision, a qualified person must review it with the authority to change it, and the person affected must be told AI was used. Internal drafting that is not the final recommendation is excluded. So decide the rating yourself, from your own evidence, and ask Claude only to help you say it clearly and fairly.

Which Claude plan and settings keep review notes private?

A Claude Team plan is the safer home for staff review notes, because Team does not train on your content by default, while Free, Pro and Max users must switch model improvement off themselves. Enterprise matches Team and adds custom retention controls.

What mattersFree, Pro, Max (consumer)TeamEnterprise
Price (claude.com/pricing, 22 Nov 2026)$0; Pro $20 a month ($17 billed annually); Max $100$25 a seat a month, $20 billed annually; 2 to 150 seats$20 a seat a month billed annually plus usage; 20-seat minimum
Model trainingUsed to improve Claude when the model-improvement setting is on (Settings, then Privacy)None by defaultNone by default
Deleted chatsGone from history at once; from back-end storage within 30 daysYou delete chats from your history at any timeCustom retention controls available
Thumbs up or down feedbackStores the whole conversation for up to 5 yearsSame, but owners can switch off the "Rate chats" option for everyoneSame as Team
Private projectProjects are private; Free allows up to fiveChoose "Keep it private" when creating the projectAs Team

The feedback row catches people out. Clicking thumbs up on a nicely worded review sends the entire conversation, notes included, to Anthropic's feedback store for up to five years, on consumer and business plans alike. Do not rate review chats. On a Team plan, an owner can remove the temptation by turning off Rate chats under Organization settings, then Data and Privacy. For the wider question of when a Team plan pays for itself, whether Claude Team is worth it for a small business runs the numbers.

Setting up a private reviews Project, step by step

A dedicated Project keeps the framework, rating definitions and instructions in one place, and its memory settings let you stop one person's review leaking into another's. The steps follow Anthropic's help pages on projects.

  1. Open Projects from the sidebar and click + New Project. Name it plainly, such as "Reviews 2026". Claude does not read the project name or description, so keep personal details out of both anyway.
  2. On a Team plan, choose Keep it private when asked about visibility. Do not share the project; each manager should hold their own.
  3. Click Set project instructions, paste the instructions from the prompt section below, and click Save instructions.
  4. With the + button, add only shared, non-personal material to project knowledge: your competency framework, the review form, rating definitions and one anonymised example of a good review. Individual notes go into individual chats, not project knowledge.
  5. On the project's page, under Memory, turn off Use account memory. Anthropic's help page explains that each project has its own memory, but new project memories can be referenced in chats outside projects when account memory is on.
  6. For each employee, start a new chat in the project and turn off Memory in the "+" menu before sending the first message. That chat then neither uses nor adds to the project's memory, so nothing about one person carries into the next person's chat.

The project is one of three ways to work, and the trade-offs are worth seeing side by side:

ApproachPrivacyEffort each timeBest for
Ordinary chat, notes pasted inChat saved to history; memory applies unless switched off for that chatPaste framework and instructions every timeA single review, once
Private project, memory off per chatFramework stored once; each person's chat kept apartPaste notes onlyA manager with several reviews a year
Incognito chat outside a projectNot saved to history or memory; on Team and Enterprise, deleted within 30 days unless flaggedPaste everything every time; the chat cannot be reopened once closedOne-off sensitive drafting, such as a difficult review

Incognito chats are not available inside projects, which is why the per-chat memory switch matters here. If you prefer to work outside a project, incognito chats are not saved to history or memory, but you lose the stored framework and instructions and have to paste them every time. Setting up Claude Projects as a shared team assistant covers projects that are meant to be shared; this one is deliberately the opposite.

How should review notes be structured before Claude sees them?

Review notes work best as dated, factual evidence grouped under your objectives and competencies, with a label instead of a name, the manager's provisional rating already decided, and no health, family or other personal details.

Here is a filled-in notes page for one account handler at an illustrative insurance broker (eight staff: six account handlers and two administrators). Everything Claude needs is here; nothing it does not need is.

EMPLOYEE: Handler C        ROLE: Account handler, commercial
PERIOD: Jan to Oct 2026    MANAGER'S RATING (decided): Meets, strong

OBJECTIVES
1. Renewal retention above 88%: achieved 91% (118 of 130 renewals)
2. New business: $45,000 commission target; achieved $38,500
3. File audits: 2 of 3 audits clean; Q2 audit found missing
   signed proposal forms on 3 files (fixed within a week)

EVIDENCE BY COMPETENCY
Client care
- 14 Mar: handled a fleet client's claim complaint, client renewed
- 2 Jul: two thank-you emails from clients forwarded by the director
Accuracy
- Q2 audit issue above; no repeat in Q3
Teamwork
- Covered Handler F's renewals for 3 weeks in August

FEEDBACK RECEIVED
- From senior handler: "reliable, could speak up more in meetings"

DEVELOPMENT AGREED LAST YEAR
- Complete advanced commercial lines course: done in June

MANAGER'S VIEW
Strong year on retention; new business short by $6,500, partly
because of the August cover. Next year: new business pipeline.

Most managers' notes start nowhere near that. Here is the same handler's material before and after an hour of tidying, section by section:

SectionBefore (raw notes)After (structured)
Retention"Retention good"91%: 118 of 130 renewals, target 88%
New business"Bit behind"$38,500 against $45,000; three weeks of August spent covering a colleague
Accuracy"Audit issue in the summer"Q2 audit: 3 files without signed proposal forms, fixed within a week; Q3 clean
Client care"Clients like her"14 Mar complaint resolved and client renewed; two thank-you emails on 2 Jul
Teamwork"Quiet in meetings"Senior handler's feedback quoted; August cover noted as a positive

The "before" column is what most first drafts are written from, and it explains why AI-written reviews so often sound vague: the vagueness was in the notes. Notice what the structured notes contain: numbers, dates, sources and a decision. That is what lets Claude draft something specific. Notes that say "good year, bit quiet" produce reviews that say the same thing at greater length.

If pulling each handler's figures from the team's KPI spreadsheet is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

What should you strip out, and what can stay?

Strip names, client details, health, family and sickness information from review notes, and keep the dated work evidence, because the evidence is all the draft needs.

InformationTreatmentWhy
Name, email, staff numberReplace with "Handler C"Claude needs a role, not an identity
Client namesReplace with "a fleet client"Client data has its own confidentiality duties
Health, sickness reasons, pregnancy, disability, family mattersRemove entirelySensitive, rarely relevant to the wording, and risky in any tool
Grievances, disciplinary cases, union mattersRemove; handle with HR or an adviserA review draft is the wrong place for them
Salary and pay-rise amountsRemovePay is decided separately; keep it out of review wording
Objectives, results, dated examples, feedback quotesKeep, with sourcesThis is the evidence the draft is built on
Your provisional ratingKeep, clearly marked as decidedStops Claude drifting towards its own view

A manager in a team of eight may find that "Handler C" plus a role and a set of results still identifies the person to anyone inside the firm. That is fine: the goal is to keep identity and sensitive detail out of the tool, not to hide it from yourself.

Eight reviews in two evenings at an insurance broker

The illustrative broker's director had eight reviews due in November. In previous years each took about 75 minutes to write from scattered notes, ten hours in all, and the last few were noticeably thinner than the first.

This year the director spent the first evening on notes: about 40 minutes per person to gather dated evidence and decide each rating, five hours in total. The second evening was drafting. For each person, a new chat in the private project with memory off, the notes pasted in, and Claude's draft back in under a minute. Editing took about 15 minutes per review, two hours in all.

The bias check, run on all eight anonymised drafts together, found three problems worth fixing:

  • One administrator's draft was 40% shorter than the others and used "helpful" and "pleasant" where the handlers' drafts cited results. The notes themselves were thin; the director added four dated examples and redrafted.
  • Two handlers with similar retention figures were described differently: one as "confident", the other as "can come across as abrupt". The director found no evidence for the second phrase in the notes, and removed it.
  • One draft praised a handler for "a strong year" largely on the strength of October, the month before the review. The notes showed a weak first quarter that the draft skipped. The wording was rebalanced across the whole period.

Total time: seven hours against ten, with all eight reviews held to the same standard of evidence. The bigger gain was consistency, not speed.

The drafting prompt, and what a good draft looks like

Save these instructions in the project so every chat drafts the same way:

PROJECT INSTRUCTIONS: Performance review drafting

You draft review wording from a manager's notes. You never decide or
suggest a rating; the manager's rating is given and final.
- Use only evidence in the notes. Never invent examples, numbers,
  feedback or achievements. If a section has no evidence, write
  "[NO EVIDENCE IN NOTES]".
- Describe behaviour and results, not personality.
- Every criticism needs a dated example and a suggested next step.
- Cover the whole period, not just recent months.
- Use the competency headings from the framework file.
- Plain, warm, direct. 350 to 500 words.
- End with three development objectives for next year, each with a
  measure and a date.

And an illustrative extract of what came back for Handler C, with the notes above:

Client care
Handler C retained 118 of 130 renewals (91%), above the 88% target.
In March, C resolved a fleet client's claim complaint and the client
renewed; two further clients wrote to thank C in July.

Accuracy
The Q2 file audit found three files without signed proposal forms.
C corrected all three within a week, and the Q3 audit was clean.
Next step: a pre-renewal checklist for proposal forms.

New business
C reached $38,500 of the $45,000 commission target. Three weeks
covering a colleague's renewals in August reduced prospecting time...

What I would change: "reduced prospecting time" explains the shortfall but should not excuse it entirely, so the director added one sentence on what the pipeline looked like before August. I would also check that "warm" has not turned into flattery; the line "a valued member of the team" appeared in an earlier draft with no evidence behind it and was cut.

Which follow-up questions improve a Claude review draft?

Four follow-up questions improve almost any Claude review draft: where each claim came from, what the employee would challenge, how each criticism reads as behaviour plus next step, and what the measurable objectives are. Ask them in the same chat, after the first draft:

Ask ClaudeWhat it revealsWhat happened with Handler C's draft (illustrative)
"List every claim in the draft and the line of my notes it came from."Invented or stretched claimsOne sentence, "a go-to person for complex risks", had no source; cut
"Which sentences would the employee most likely challenge, and why?"Weak evidence and loaded wording"Could speak up more" flagged as vague; rewritten with the meeting context
"Rewrite each criticism as behaviour, example and next step."Softened or personal criticismThe audit issue gained a clear next step: a pre-renewal checklist
"Turn the development section into three objectives, each with a measure and a date."Development that is too loose to review next yearSee the objectives below

The objectives that came out of the last question, after the director's edits, were concrete enough to review in twelve months:

ObjectiveMeasureBy
Build a new-business pipeline20 qualified prospects in the CRM, 6 quotes issued30 June 2027
Zero audit findings on proposal formsAll three quarterly audits clean31 December 2027
Lead one team meeting slot a monthTen short case updates presented31 December 2027

A second example: an electrical contractor's apprentice review

The same routine works for very different roles once the evidence changes. Take an illustrative electrical contractor reviewing a second-year apprentice. The evidence sources are the site supervisor's weekly notes, the college progress report, the firm's safety observation log and the apprentice's own logbook.

The owner's structured notes covered 46 weeks: 41 supervisor notes, of which 33 were positive and 8 raised an issue; a college report showing 4 of 5 modules passed and one resit due in January; 2 safety observations (one praise for isolating a circuit correctly before work started, one reminder about wearing eye protection when drilling); and the owner's decided rating of "on track".

Two things differed from the office reviews. First, the draft needed to separate what the apprentice controls (attendance, safe working, effort on the resit) from what the firm controls (the range of jobs offered), and Claude did this well once the instruction said so. Second, the safety reminder had to stay in, stated plainly with the date, because a review that leaves out a safety point is a review someone may regret later. Claude's first draft folded it into "generally safe working practices"; the owner restored the specific example.

How do you check a Claude draft for bias and fairness?

Check a Claude-drafted review for bias by comparing it with the evidence and with your other reviews, looking for unsupported claims, personality words, recency, uneven depth and coded language. Coded language includes words that hint at age, gender or background rather than describing the work.

Claude can help with the comparison if you give it all the drafts at once, still anonymised, in a fresh chat with memory off. Ask: "For each draft, count the dated examples, list every adjective about the person, flag any claim with no evidence in its notes, and compare tone and length across drafts." Then judge the results yourself. The checks that matter most:

CheckWhat it looks likeFix
Unsupported claims"A natural leader" with no exampleAdd evidence or delete the claim
Personality, not behaviour"Abrasive", "bubbly", "not a people person"Describe what happened, when and its effect
RecencyThe last two months dominate a ten-month periodRequire one example per quarter in the notes
Uneven depthSome drafts cite results, others only adjectivesBring thin notes up to the same evidence standard
Coded language"Energetic young", "emotional", "lacks polish"Remove and replace with evidence about the work
Softened criticismA real problem reduced to "some room to grow"State the issue, the example and the next step plainly

One caution: the comparison is only as fair as the notes. If you keep detailed notes for some people and sparse ones for others, the drafts will reflect that, however neutral the prompt. Writing fairer performance reviews with AI goes deeper into calibration across a team.

Who decides the rating, and what do you tell staff?

The manager decides the rating, before Claude drafts anything, and the employee should be told that an AI tool helped with the wording. Both protect the employee and keep you on the right side of Anthropic's usage policy.

Three rules point the same way. Anthropic's usage policy, as above, requires a qualified person to review any AI recommendation about a worker's performance and requires disclosure when a decision rests on one. Data-protection law such as the GDPR gives people the right not to be subject to a decision based solely on automated processing where it significantly affects them (Article 22). And the GDPR's right of access (Article 15) means an employee can ask for a copy of the personal data you hold about them, which can include your notes. If you employ people where the GDPR or similar law applies, write every note as if the employee may one day read it, and ask an HR adviser about anything uncertain.

A disclosure line on the review form is enough for most teams: "I used an AI writing tool to organise my notes into this review. The examples, judgements and rating are mine." It costs nothing, it is true, and it answers the question before anyone asks it.

If you want help fitting AI drafting into how your business runs, I offer one-to-one AI consultations for small businesses.

Where AI-drafted reviews go wrong

These are the mistakes that turn a time-saver into a grievance.

An achievement nobody achieved

A draft credited a handler with "winning two new fleet accounts". The notes mentioned two fleet quotes, one of which was lost. Claude had smoothed the story. The "[NO EVIDENCE IN NOTES]" rule and a line-by-line read against the notes catch this; never sign a sentence you cannot point to evidence for.

One person's details in another's review

A manager drafted all eight reviews in one long chat. The sixth draft mentioned a client from the third person's notes. Use a new chat per person with memory off, as in the setup steps.

The whole HR file pasted in

A manager uploaded the full personnel file, including a sickness record and a grievance outcome, "for context". None of it belonged in the draft, and all of it was now in the chat. Delete that chat, and paste only the structured notes in future.

The same phrases in every review

Staff compared reviews and found "consistently goes above and beyond" in four of them. Ask Claude to vary nothing, and fix the cause instead: generic phrases appear when notes lack specifics. Add evidence, and the language becomes individual on its own.

Letting Claude "suggest" a rating

A manager asked which rating the evidence supported, liked the answer and used it. That is precisely the AI recommendation the usage policy's high-risk rules are about. Decide the rating first; if you want a sense check, ask Claude to list the evidence for and against your rating, not to choose one.

What should you keep and delete once a review is final?

Once a review is final, keep the signed review in your HR system and delete the Claude chat that held that person's notes; the project keeps only templates and instructions. On consumer plans, Anthropic says deleted chats leave your history immediately and its back-end storage within 30 days. On Team plans, chats stay in the product until someone deletes them, so the deleting is a step in your process, not something that happens on its own.

ItemWhere it livesHow long
Signed reviewHR system or personnel fileYour normal HR retention period
Structured notesWith the signed reviewSame as the review
Each person's drafting chatClaudeDelete at sign-off
The bias-check chat with all draftsClaudeDelete once every review is final
Framework, form, rating definitions, instructionsThe private projectKeep until you replace them

If a manager leaves, the reviews they wrote should already be in the HR system; do not plan on recovering anything from a former employee's Claude account. And when you next update the competency framework, replace the file in the project rather than adding a second version, because two frameworks in one project can produce drafts that mix their headings. For the broader policy on what staff may paste into AI tools, stopping staff pasting client data into free AI tools and using AI for HR without exposing employee data set the house rules this routine sits inside.

Questions managers ask about reviews and Claude

Is it safe to put employee review notes into Claude?

It can be, with the right plan and settings. Team and Enterprise plans do not train on your content by default. On Free, Pro and Max, turn off the model-improvement setting under Settings, Privacy before you start. In every case, use labels instead of names, leave out health and personal details, avoid the thumbs up and down buttons on these chats, and delete the chats once the review is final.

Can Claude decide an employee's performance rating?

It should not. Anthropic's usage policy lists evaluating a worker's performance as a high-risk area: if Claude's recommendation shapes the decision, a qualified person must review it with authority to change it, and the employee must be told AI was used. Data-protection law such as the GDPR also restricts decisions based solely on automated processing. Set the rating yourself, then ask Claude for wording.

Should I tell staff that AI helped write their review?

Yes, a short line is the honest default, and Anthropic's usage policy requires disclosure when a decision about someone rests on an AI recommendation. A sentence such as 'I used an AI writing tool to structure my notes; the judgements and the rating are mine' covers it. Staff may also be able to request copies of their personal data, including notes, so write notes you would be comfortable showing them.

Does Claude remember one employee's review when I draft the next?

It can if memory is on. Each Claude project has its own memory, and new project memories can be referenced outside the project when account memory is on. Turn off 'Use account memory' on the project's page, and turn off Memory in the + menu before the first message of each person's chat, so one review never colours another.

Can I use Claude's incognito mode for performance reviews?

Only outside a project. Anthropic's help page says incognito chats are available for chats outside projects, are not saved to history or memory, and on consumer plans are not used to improve Claude even if that setting is on. Inside a project, the equivalent is switching Memory off in the + menu before your first message.

Further reads

Sources: Claude pricing page; Claude help centre articles on projects, creating and managing projects, incognito chats and uploading files; Claude privacy centre articles on model training, data retention and model-improvement settings for consumer and commercial plans; Anthropic's Usage Policy (high-risk use case requirements); GDPR Articles 15 and 22 (all read 22 November 2026).

Want a second pair of eyes on your setup?

Bring the process you're trying to fix. On a 1:1 call we'll work out which AI tools should handle it, which parts stay manual, and how to roll it out without disrupting the team.

Book a 1:1 call with me