Talent Review Pack Assembly Chain for Calibration Sessions

What This Builds

A three-step prompt chain that takes the scattered inputs behind a calibration session (manager notes, a succession candidate list, and aggregated comp-position context) and turns them into one clean packet, organized by performance tier and ready-now versus needs-development status, before you walk into the room. Instead of reconciling three tools by hand every quarter, you run three prompts in sequence and get a packet a business leader can actually read.

The packet organizes the discussion. It never finalizes a rating, a succession call, or a pay decision. That stays with the people in the room.

Prerequisites

• Manager notes or performance summaries for the group being calibrated, ready to anonymize
• A current succession candidate list for any critical roles in scope
• Aggregated, banded compensation-position data (not individual pay figures) if comp context is part of your calibration process
• A Claude account and a ChatGPT account, both on a plan with enough context length for a full team's worth of notes
• Total ongoing cost if you do not already have these: $20/month for Pro, plus $20/month for Plus

The Concept

A prompt chain is just a sequence of prompts where each one builds on the output of the last, instead of you starting fresh every time. Here, one tool summarizes the raw notes, then you carry that summary into a second step that cross-references it against your succession list, then a third step formats the combined result into a packet. Each step does one job well, rather than asking a single prompt to do everything at once and getting a messy result.

Build It Step by Step

Part 1: Anonymize your source material

Before anything goes into a chatbot, replace every employee name with a role label or letter (Employee A, Employee B, or "Senior Analyst, Team 2"). Do this in your own document first, not inside the chat. Keep a separate, secure key mapping labels to names, stored outside any AI tool, for you to reference when the packet comes back.

Part 2: Step one, summarize the notes in Claude

Open a new Claude conversation and paste in the anonymized manager notes for the group:

Save this output. It becomes the input for the next step.

Part 3: Step two, cross-reference against succession, back in the same Claude conversation

Paste your anonymized succession candidate list into the same conversation and prompt:

Because this happens in the same conversation, Claude already has the performance summary in context and does not need it re-explained.

Part 4: Step three, format the packet in ChatGPT

Copy both outputs into a new ChatGPT conversation and prompt:

What you should see: A single formatted document, organized and scannable, explicitly labeled as a discussion aid.

Part 5: Re-attach names and review before the meeting

Using your secure key from Part 1, re-attach real names to the packet only in your own working copy, outside any AI tool. Read the full packet once for accuracy against what you actually know about each person, since the AI only saw what was in the notes, not your full context. This review is the step that makes the packet safe to bring into the room.

Real Example: Quarterly Calibration for a 12-Person Team

Setup: A quarterly calibration is coming up for Team B, 12 employees, with two people on the succession list for a senior analyst role opening next year.

Input: Anonymized manager notes for all 12, plus the two succession candidates' recent performance notes.

Output: The packet groups the 12 into three performance tiers with recurring themes (three people flagged for "strong technical work, needs more cross-team visibility"), and notes that one succession candidate looks ready now while the other has a recent note about missed deadlines worth raising before confirming their readiness status.

Time saved: Cuts calibration prep from roughly an evening of reconciling three sources by hand to about 30-40 minutes of running the chain and reviewing the result.

What to Do When It Breaks

• The summary in step one misses context you know is relevant → The notes you pasted were incomplete. Add the missing detail to your source notes and rerun step one rather than trying to patch the final packet by hand.
• Step two's succession cross-reference feels generic → Give it more to work with. A one-line note on why each person made the succession list in the first place produces a sharper "ready now versus needs development" call.
• The formatted packet in step three loses nuance from steps one and two → Ask ChatGPT directly: "What did you leave out from the input?" before finalizing, then add anything material back in by hand.
• You catch yourself starting to draft a final rating instead of discussion points → Stop and reread the one-line note at the top of the packet. That note exists to keep the packet from being mistaken for a decision it is not.

Variations

• Simpler version: Run all three steps in a single Claude conversation instead of switching tools, if you do not need ChatGPT's formatting and are comfortable with a plainer output
• Extended version: Add a fourth step that drafts three open-ended calibration questions per tier, so the packet also seeds better discussion instead of just a status report

What to Do Next

• This week: Run the chain once using your next upcoming calibration as the real test case
• This month: Save your three prompts in a personal notes file so each quarter starts from the same tested wording instead of rebuilding it
• Advanced: Extend the chain with an aggregated comp-position step (banded data only) for cycles where compensation context belongs in the calibration conversation

Advanced guide for HR Business Partner professionals. Performance notes, succession status, and compensation position are among the more sensitive data an HRBP handles. Anonymize every name before anything goes into a chatbot, keep the name-to-label key outside any AI tool, use aggregated comp bands rather than individual pay figures, and restrict access to the finished packet to the people who would already see this information in a calibration meeting.