The skill behind this guide: Elena - Business Analyst AI Skill. Loads the questioning order and the acceptance-criteria discipline described below into Claude in one upload - $14.99, one payment, no subscription.
View the Elena skill →To use Claude as a business analyst, give it the request quoted exactly, the current process step by step, and the specific failure you are trying to fix - in that order - before you ask for a single requirement. Left to itself, Claude documents the solution the stakeholder proposed rather than the problem underneath it, and fills the gaps in your process map with plausible steps it invented. Below is the input order, five prompts that produce checkable output, and the failure modes worth re-reading every draft for.
What do you give Claude, and in what order?
Order matters because Claude anchors on whatever it reads first. Lead with a proposed solution and every later answer bends toward it. Feed it in this sequence:
- The request, quoted exactly. Paste what the stakeholder actually said or wrote, not your tidied-up version. "Sales needs a dashboard" and "Sales can’t tell which deals are stuck" produce completely different documents, and the second is usually what they meant.
- Who is affected and who signs off. Roles, and which of them hold a veto. Otherwise Claude writes for one generic "user" and you get requirements nobody owns.
- The current process, step by step. Every step gets an actor, a system, a trigger and an output. Include the manual workarounds - the spreadsheet one person maintains, the Slack message that functions as an approval. That is usually where the real requirement is hiding.
- The specific failure. Not "it’s inefficient". What breaks, how often, and what it costs: "three of about forty invoices a month get keyed twice, and finance spends roughly a day a month reversing them." If you don’t have the number, write that you don’t - that is a research task, not something to let Claude fill in.
- The constraints. Deadline, budget, systems that cannot be touched, compliance rules, the integration that took nine months last time.
- What already exists. Current docs, ticket exports, screenshots of the existing screens. Claude reads pasted text and attached files. It does not read your Jira.
Only then ask for output, and ask for one artefact at a time: current-state map, then gap list, then stories. Request a full BRD in one shot and you get a document where every section is roughly 60% right and you cannot tell which 40% is wrong.
What does Claude get wrong at business-analysis work?
It documents the solution, not the problem. Ask for requirements for "a new approvals dashboard" and you get dashboard requirements. It will not ask whether the approval step should exist at all. You have to force that split explicitly; it never does it unprompted.
It fills gaps with invented process steps. Tell it "then it goes to finance" and it returns three plausible finance sub-steps, written in exactly the same confident register as the steps you actually supplied. Nothing in the output marks the difference. This is the most expensive error in the list, because invented steps get reviewed, nodded at, and built.
Acceptance criteria that restate the story. Unedited, it writes "As a manager I want to export the report so I can share it. Acceptance criteria: the manager can export the report." That is not testable and QA will hand it straight back.
Non-functional requirements written as adjectives. "The system must be fast, scalable and user-friendly" shows up in nearly every first draft. No number, no measurement method, nothing anyone can sign off against.
One more on long threads: by message twenty it has quietly re-included scope you cut at message six. Re-state the agreed scope every few turns and check the document against it.
Which prompts actually work?
1. Split the problem from the proposed solution.
Here is a request, quoted exactly: <paste>. Do not write requirements yet. Separate what was asked for (a proposed solution) from the problem it implies. Then list the five questions whose answers would most change what we build. Mark every line you inferred rather than read with [INFERRED].
2. Map the current state without letting it invent.
Below are the process steps I know. Reflect them back as a numbered flow with actor, system, trigger and output per step. Where I have not given you a step, write [UNKNOWN - confirm with owner] instead of filling it in. Do not add steps.
3. Force acceptance criteria to be testable.
For each acceptance criterion, write the exact test a QA engineer would run: starting state, action, observable result. If you cannot write the test, the criterion is too vague - rewrite it until you can.
4. Put numbers into the non-functional requirements.
Rewrite every non-functional requirement so it contains a number and a way to measure it. Any you cannot quantify, remove from the list and put in a separate section headed "Needs a target from the business".
5. Catch scope drift.
Agreed scope: <list>. Read the document above and list every requirement in it that is not covered by that list.
Prompt 2 is the one that saves the most rework. Prompt 3 is the one people skip and regret.
Elena carries the input order and the [INFERRED] / [UNKNOWN] marking rules above as standing instructions, so you don’t have to re-type them into every thread.
View Elena - Business Analyst AI Skill →How do you install it?
The skill is a ZIP with a SKILL.md file at its root. In the Claude desktop app go to Customize → Skills → +, upload the ZIP, and toggle it on. Claude reads the SKILL.md and shows you a summary of what the skill does before you enable it.
One thing to check first: code execution must be enabled under Settings → Capabilities. Skills themselves work on every plan, including Free. Without code execution the skill will not run.
On a free Claude plan, or in ChatGPT or Gemini, the fallback is to open SKILL.md in any text editor and paste its contents into custom instructions. You lose the on/off toggle; the behaviour is the same. Plan requirements checked August 2026 - Anthropic changes plan gating fairly often, so confirm the current requirement in the Claude app before you buy.
When should you not use Claude for this?
It cannot run a stakeholder interview, and it has no access to your Jira, your data model, your CRM’s real field limits, or your team’s velocity. Anything it says about what a stakeholder wants is a guess dressed as a finding. If you do not already know how the current process runs, Claude cannot find out for you - it will hand you a confident map of a process it has never seen.
In regulated work where each requirement must trace back to a signed-off source, the drafting speed is a liability unless you keep the provenance yourself. And on a genuinely novel domain problem, with no close equivalent in its training material, it falls back on generic patterns and you get a template again.
Where it earns its money: the second and third drafts, expanding an agreed scope into stories, gap analysis between two process maps, and flushing out the requirement nobody wrote down.
Who is this for?
Business analysts who want a faster second draft, product owners writing stories, founders specifying a build for a contract developer, and operations leads redesigning a process they already understand. The work & business skills collection covers the neighbouring roles.
Same discipline, but the process maps and acceptance criteria are written in Salesforce terms - objects, record types, validation rules, sharing model.
View Owen - Salesforce Business Analyst AI Skill →In summary:
Give Claude the request verbatim, the current process, and the specific failure - in that order - then make it mark everything it inferred. For general requirements and process work, Elena - Business Analyst AI Skill ($14.99); for Salesforce-specific analysis, Owen - Salesforce Business Analyst AI Skill ($14.99). Both are ZIP uploads, no coding, 30-day money-back guarantee.
Elena - Business Analyst AI Skill
Upload one ZIP and Claude interrogates the request before it documents it - marks its own inferences, refuses to invent process steps, and writes acceptance criteria QA can run. One payment, no subscription.
KissMySkills is a marketplace of 1,000+ AI skills, 158 prompt packs, 55 agents & free tools for Claude, ChatGPT & any AI chat.