Prompt Chain: Turn Anonymized 360 Data into Session Talking Points in Three Checkable Steps

What This Builds

Asking an AI to read raw 360 data and hand you a finished debrief in one pass tends to blur three different jobs together: finding the themes, deciding which ones matter most, and turning that into something to say out loud. This chain forces those three jobs into separate steps in the same conversation, so you can check and correct each one before it feeds the next, instead of trying to spot an error buried inside one long answer.

Prerequisites

• A Pro subscription ($20/month), useful here for the longer context window a full set of anonymized rater responses needs.
• A completed anonymization pass on the 360 data: no rater names, no job titles specific enough to identify someone, no verbatim quotes traceable to a single person.
• Familiarity with reviewing an AI's output critically rather than accepting the first draft.

The Concept

A prompt chain is just a conversation with a plan. Instead of asking one big question and hoping the answer covers everything, you ask three smaller questions in sequence, each one building on the answer to the last. Step one only extracts themes. Step two only ranks them. Step three only turns the ranked list into talking points. Because each step produces something short enough to read carefully, you catch a wrong theme before it becomes a wrong talking point three steps later.

Build It Step by Step

Part 1: Prepare the anonymized input

Before opening Claude, take your raw 360 data (rater responses, interview notes, or assessment narrative) and strip it down to anonymized statements: no names, no titles specific enough to identify someone in a company of any size, no verbatim quote that could be traced back to a single rater. Paraphrase distinctive phrasing rather than quoting it directly. This is the slowest part of the process and it happens before Claude sees anything.

Part 2: Run the three-step chain

Open a single conversation in claude.ai and run these in order, reading the output of each before sending the next.

Step 1: Extract themes

Here is anonymized 360 feedback data for an executive I'm coaching (no
names, titles, or identifying detail, paraphrased rather than quoted
directly):

[paste anonymized data]

Extract the recurring themes across these responses. List each theme
with a one-sentence description and roughly how many of the responses
touched on it. Don't rank them yet, just surface what's actually there.

Read the theme list. If something looks like it came from a single outlier response rather than a real pattern, say so before moving on.

Step 2: Rank by development impact

From the themes you just listed, rank them by likely development impact
for this executive, meaning which ones, if addressed, would most change
how they're experienced by the people around them. Explain your
reasoning for the top three in one or two sentences each.

Check the ranking against your own read of the material. You know this client's context; the ranking is a starting point for your judgment, not a replacement for it.

Step 3: Draft talking points

Using the top two or three themes from your ranking, draft five specific
coaching talking points for my next session with this client. These
should be open questions or discussion prompts I can use, not
conclusions I hand the client. Keep the language client-facing and
non-clinical.

Part 3: Verify before the session

Before the session, check the talking points against two things: your own memory of the client's stated goals, and the client's actual assessment consent form, to confirm this kind of AI-assisted processing is covered by what they agreed to. If it isn't documented in the consent language, that's a conversation to have with the client or your own supervision resource before you rely on the output in session.

Real Example: A VP Working on Delegation

Setup: Ten anonymized rater responses for a VP of Operations, paraphrased and stripped of names, titles, and direct quotes.

Input (Step 1): The anonymized data pasted into the chain.

Output (Step 1): Four themes surfaced: delegating decisions rather than just tasks, listening under pressure in escalations, inconsistent follow-through on commitments made to peers, and strong technical credibility that sometimes overshadows collaborative input.

Output (Step 2): Delegation and listening under pressure ranked highest, since multiple raters connected both to the same recurring pattern: the VP steps in and takes over rather than coaching a direct report through a mistake.

Output (Step 3): Five talking points, including: "What would it look like to let someone on your team make the wrong call once, and coach them through the recovery instead of taking it back?"

Time saved: Breaking the task into three short exchanges instead of one long request makes each stage easy to check, which cuts the amount of second-guessing and re-reading a single dense output usually requires.

What to Do When It Breaks

• Step 1 surfaces a theme that only one rater actually raised → Push back in the same conversation: "How many of the ten responses actually mention this?" and ask it to revise the theme list if the count doesn't support calling it a pattern.
• Step 3 produces conclusions instead of open questions → Re-run the step with a sharper instruction: "Rewrite these as questions only, no statements telling the client what to do."
• Identifying detail creeps back in at step 2 or 3 → This can happen if your anonymized input still contained enough context (a distinctive project name, an unusual title) for Claude to infer specifics and echo them back. Reread each step's output for anything that could point to a real person before it reaches your session notes.
• The chain drifts if you leave and come back later → Keep the whole three-step sequence in one sitting when you can. If you must pause, re-paste the prior step's output when you resume so the next step has the right context.

Variations

• Simpler version: Run all three questions as one combined prompt when the 360 data is short and the themes are obvious. Save the full chain for denser or ambiguous data sets.
• Extended version: Add a fourth step that drafts a short written summary of the top themes (still non-attributable) for the client's own file, separate from the talking points you'll use live.

What to Do Next

• This week: Run the chain on your next scheduled 360 debrief and compare the talking points against what you'd have written unaided.
• This month: Keep a private note of which themes the chain surfaced accurately versus which ones needed correction, so you learn where it tends to overreach.
• Advanced: Pair this chain with a Claude Project loaded with your own de-identified debrief frameworks, so step three draws on your established talking-point style instead of a generic one.

A Note on Confidentiality

360 feedback is the most identifying material you handle: named raters, specific incidents, and often language distinctive enough to trace back to one person even after a light edit. The anonymization in Part 1 is not a formality, it's the entire control that makes this chain acceptable to run at all under your client's assessment consent and the ICF Code of Ethics, which treats this kind of data as confidential to the engagement. If any step of the chain produces output that references a detail specific enough to identify a rater or the client's employer, stop and rework the anonymized input rather than editing the AI's answer after the fact. When the data is too dense or too identifying to anonymize confidently, do the first pass by hand and save the chain for material that's already clean.