Needs Analysis Prompt Chain: From Interview Notes to a Learning Plan Section in One Claude Conversation

What This Builds

A five-prompt chain that runs inside a single Claude conversation and carries a business unit's raw interview notes through to a draft section of your annual learning plan. The stages are a needs summary, a cause classification, a ranked list of priorities, the plan section itself, and a final audit of that section against your notes. You stop after every stage, read what Claude produced, correct it, and only then send the next prompt.

The point is not speed alone. A single giant prompt hides the reasoning, so a wrong assumption in the middle ends up as a confident paragraph in the plan. In a chain, each stage is small enough for you to catch the mistake where it starts.

Prerequisites

• A Claude account. The chain works in a regular conversation, and a paid plan (Pro at $20/month) helps if your notes are long, because long conversations use up free usage limits faster. Total ongoing cost of this build: your Claude plan, nothing else.
• Interview notes for one business unit, already abstracted (see the data note below)
• Your ranking criteria and weights, written down before you start. Examples: business impact, number of people affected, urgency, cost to address. The criteria are yours to choose.
• Rough limits for the planning period: team capacity, a budget band or "not stated", and any fixed deadlines
• A place to keep each stage's output as you approve it, such as a Google Doc or a Word file

Before you paste anything. Interview notes about a business unit often contain names, performance complaints about individuals, and hints of reorganizations that have not been announced. Replace every name with a role or a code (I-1, I-2). Describe the unit rather than the leader ("customer support, tier 1"). Cut anything that identifies a person's performance issue. If a note mentions a planned restructure, product launch or headcount change that staff have not been told about, either remove it or replace it with a neutral phrase such as "a change to team structure". Your company's AI policy and approved tool list decide whether Claude may see this material at all, and a free consumer chat plan may use conversations for training unless you have switched that setting off. If you are unsure, ask IT or your security team before you start.

The Concept

Think of a good consultant reading interview notes with a colleague looking over their shoulder. The colleague asks after each step, "Show me what you have so far." A prompt chain puts that colleague in the conversation, and the colleague is you.

Each prompt does one job and hands its result to the next. Because everything sits in one conversation, Claude can see the earlier stages without you pasting them again. That is also the risk. If stage two is wrong and you let it pass, stages three to five build on the error. The checkpoint after each prompt is the safeguard, and skipping it defeats the technique.

Build It Step by Step

Part 1: Prepare the inputs

1. Open a new Claude conversation. Do not reuse an old one, because leftover context from another project can leak into your plan.
2. In your notes document, label each interview with a code: I-1, I-2 and so on. Keep one block per interview.
3. Write your criteria and weights on one line each. For example: "Business impact, weight 40. People affected, weight 25. Urgency, weight 20. Effort to address, weight 15." Weights should add up to 100 so the arithmetic is easy to check.
4. Write the constraints in one short paragraph: designer or trainer capacity for the period, a budget band as a range (or "not stated"), and any deadline that cannot move.

Part 2: Run the five prompts

Send one prompt at a time. Replace every square-bracket placeholder with your own text before sending. The text inside the code blocks is exactly what Claude receives.

Prompt 1: The needs summary

You are helping a learning and development manager turn stakeholder interview notes into a needs summary. Work only from the notes below. Do not add facts, numbers or names that are not in the notes. Where something is unclear, say so instead of guessing.

Business unit: [business unit, described by function, not by the leader's name]
Business goal the stakeholders named: [goal in one sentence]

Interview notes, one block per interview, each labelled with a code:
[paste abstracted notes here, with codes I-1, I-2 and so on]

Produce three things:
1. A needs summary of no more than 12 items. Number each item N1, N2 and so on, and tag each with the interview codes it came from.
2. Points where the interviews disagree with each other.
3. Open questions I should take back to the requesters before we decide anything.

Checkpoint 1. Read every numbered need against the notes. Delete any need you cannot trace to an interview code. Add any need Claude missed. Send the open questions to the requesters if they matter, and reply to Claude with the corrected list: "Use this corrected list as N1 to N[last number] from here on."

Prompt 2: Classify the likely cause

Using only the corrected needs N1 to N[last number] and the interview notes above, classify the most likely cause of each need as one of:
- Skill or knowledge gap
- Process or tooling problem
- Management or expectation problem
- Motivation or incentive problem
- Unclear

For each need give: the label, the evidence from the notes (cite the interview code), what evidence would change the label, and whether training alone would plausibly fix it. Use Unclear whenever the evidence is thin. Do not assume a cause the notes do not support.

Checkpoint 2. This is the most important review in the chain. You know the business, and Claude does not. Change any label you disagree with and say why in your reply. A need labelled as a process or management problem should not reach the ranking as a training request. If you disagree with Claude about many items, the notes are probably too thin, and the better next step is another stakeholder conversation.

Prompt 3: Rank the priorities

Rank the needs labelled Skill or knowledge gap, plus any I have re-labelled as mixed, against these criteria and weights:
[criterion 1, weight]
[criterion 2, weight]
[criterion 3, weight]
[criterion 4, weight]

Constraints for the planning period:
[team capacity, budget band as a range or "not stated", fixed deadlines]

Show a table with one row per need: the need number, a score from 1 to 5 for each criterion, the weighted total, and one sentence of reasoning. State every assumption you made when scoring. Then list the needs you set aside because training is not the fix, with one line each on who should own them.

Checkpoint 3. Recompute two weighted totals by hand or in a spreadsheet. Chatbots make arithmetic slips, and a wrong total changes the order. Then treat the ranking as a starting point for a conversation with your leaders, not as a decision. Change any score you can justify with what you know and tell Claude which scores you changed.

Prompt 4: Draft the plan section

Draft the learning plan section for [business unit] covering the top [3] ranked needs after my corrections. For each need include: the on-the-job behavior we want to see, the audience, the suggested type of approach (for example workshop, job aid, coaching or on-the-job practice, with no vendor names), how we will know it worked, dependencies and risks, and what I still need to confirm with [requester's role].

Use only what appeared earlier in this conversation. Where you need a fact I have not supplied, write [TO CONFIRM] instead of inventing it. Length: about [400] words, plain business prose, no headings beyond one per need.

Checkpoint 4. Search the draft for [TO CONFIRM] and resolve each one yourself. Read the "how we will know it worked" lines critically. If a line promises a measurable result you have no way of collecting, rewrite it to something you can actually measure.

Prompt 5: Audit the section against the notes

Review the plan section above against the interview notes and the criteria I gave. List:
1. Any statement in the plan that the notes or my criteria do not support.
2. Any need from the classification stage that was dropped without an explanation.
3. Any recommendation that depends on training fixing a problem you labelled as a process or management problem.

Then give a revised version of the section with those problems fixed.

Checkpoint 5. Compare the revision with the version you approved at checkpoint 4. Accept changes that remove unsupported claims. Reject changes that add new material, because nothing new should appear at this stage.

Part 3: Save and test

1. Copy the approved section into your learning plan document and note the date and the interview codes it rests on.
2. Keep the conversation in your approved workspace so you can show how each conclusion was reached if a leader asks.
3. Test the chain once on an old, finished case where you already know the right answer. If the chain would have led you somewhere different, look at which checkpoint would have caught it.

Real Example: Customer Support Tier 1

Setup: A learning and development manager has three abstracted interview blocks about a customer support team. The stated goal is "fewer escalations from tier 1 to tier 2 by the end of next quarter". Criteria: business impact 40, people affected 25, urgency 20, effort to address 15. Constraint: one designer for six weeks, budget "not stated".

Input: Interview I-1 says new agents do not know how to read the product's error codes. I-2 says agents escalate because the refund approval limit is unclear. I-3 says the weekly team huddle was dropped two months ago and agents no longer hear about product changes.

What the chain does:

• Prompt 1 produces needs N1 (error-code reading), N2 (refund limit confusion), N3 (no channel for product updates), and an open question about who owns the refund policy.
• Prompt 2 labels N1 as a skill gap with evidence from I-1, N2 as a process problem (the limit is a policy question, not a training gap), and N3 as a management problem. It notes that training alone will not fix N2 or N3.
• At checkpoint 2, the manager agrees with all three labels and adds a note that N3 belongs to the support director.
• Prompt 3 ranks only N1, and lists N2 and N3 as set-aside items with owners. The manager finds N1 the only item that qualifies and recomputes its weighted total by hand.
• Prompt 4 drafts one plan section for the error-code need, with two [TO CONFIRM] markers: the list of error codes in scope, and who signs off on the job aid.
• Prompt 5 flags one claim, "reduces escalations by the end of the quarter", as unsupported by the notes. The revised text says the team will track escalations tagged with error-code causes and review them at quarter end.

Output: One approved plan section, two items handed back to their real owners with a note, and a record of where each conclusion came from.

Time saved: The manager still reads everything, so the saving is in drafting and cross-checking rather than in judgment. Expect the first run to take longer than doing it by hand, and later runs to take noticeably less once your criteria and constraint paragraph are reusable.

What to Do When It Breaks

• Claude adds needs that were not in the notes → At checkpoint 1, delete anything without an interview code. Remind Claude in your reply: "Use only the notes. Remove any need without a code."
• Later stages contradict earlier ones → Long conversations drift. Paste your approved lists back in a short message ("Approved needs: ...") before Prompt 3 and Prompt 4 so the latest version is close at hand.
• Every need is labelled a skill gap → Claude leans toward training because that is what you asked about. Ask it to argue the opposite for the three highest-ranked needs: what would have to be true for each to be a process or management problem.
• Weighted totals do not add up → Recompute yourself, tell Claude the correct numbers and ask it to redo the ranking with those.
• The chain fails without any error → A stage passes with output that looks fine but rests on a bad assumption. The only defence is the checkpoint. If you find you are approving stages in seconds, slow down and trace at least three claims per stage back to an interview code.
• Usage limit reached mid-chain → Start a new conversation and paste in your approved outputs from Part 3. Then continue from the next prompt.

Variations

• Simpler version: Run only Prompts 1 and 2 to prepare for a conversation with a requester, and stop there.
• Extended version: Save Prompts 1 to 5 in a Claude Project's instructions so every business unit gets the same treatment. Compare the ranked lists side by side when you build the company-wide plan.

What to Do Next

• This week: Run the chain on one business unit and note which checkpoint caught the most problems.
• This month: Tighten your criteria and constraint paragraph, and reuse them across units.
• Advanced: Feed the approved plan sections into a ChatGPT or Claude Project that holds your annual learning plan, so later requests are ranked against everything already planned.

Advanced guide for Learning & Development Manager professionals. These techniques use more sophisticated AI features that may require paid subscriptions.