How to use Elixcit

Elixcit turns a business feature, even one described in a single sentence, into a measurable, quality checked work item in your project tool. It does that by interviewing you. The better you start the conversation and answer its questions, the better the ticket. This page shows what good and bad prompts look like, with real examples.

What it does How a session goes Good prompts Bad prompts (and fixes) Answering the questions Do & don't Limits & troubleshooting

1 What Elixcit does (and doesn't)

Elixcit is an AI business analyst. It interviews you about the business feature you need, turns it into a well defined, measurable requirement, and then creates the right work item plus its discipline subtasks in your ticketing tool (via MCP). It behaves like a disciplined analyst: it asks targeted questions before it builds anything. The requirement comes first, the ticket second.

Before it asks you anything, it learns from everything it can reach: the existing backlog (prior requirements, conventions, related work), the product codebase (the real modules, architecture and naming), and the business rule documents (underwriting guides, eligibility and validation tables, which is where the actual thresholds live). It asks only the questions none of those sources can answer, and it cites real values from the business rules instead of guessing. That way its output is grounded in your actual project, not in generic advice.

When the requirement is solid, it shows you a preview of the ticket exactly as it will look in your project tool. Only after you say yes does it create the work item and its discipline subtasks ([Solution Design], [BE], [FE], [Infrastructure], [Testing], [Business Sign Off]), with Gherkin acceptance criteria attached.

Requirements first, ticket second

Never a ticket from a vague idea. The requirement is drafted and validated before anything is created.

Ask, don't invent

If something is unclear, it asks you a targeted question instead of guessing. Unknowns are recorded, never made up.

Quality gate

A ticket is only created after the requirement passes a quality checklist and you explicitly confirm.

Grounded in real sources

Backlog, codebase and business rules are read first; acceptance criteria cite the real thresholds and values.

Every suggestion says why

When it proposes something, it names the source in the same breath: a business rule document ("the eligibility workbook caps this at 21"), the codebase ("the admin dashboard already has a roles table") or an existing ticket. If nothing backs it, it is labelled as an assumption for you to confirm.

Design aware

Have a design link? Share it. It's used for context and attached to the ticket.

Two outputs, not one

Every run yields the ticket(s) and Gherkin acceptance criteria the team reuses directly for testing.

It will: interview you, find duplicates and related tickets, cite real thresholds from business rules, tell you why it suggests each option and where it found it, write measurable acceptance criteria, create the ticket + subtasks after your approval, and analyse or improve an existing ticket if you paste its link.
It won't: write or change application code, estimate story points, modify tickets you didn't ask about, or create anything in your project tool before you explicitly approve the preview.

2 How a session goes: the flow and its gates

Elixcit follows a fixed, auditable sequence. Two gates control progress: the quality check and your explicit approval. Failing a gate loops back to more questions, never forward to an unfinished ticket.

1

Set the direction & describe the feature

Pick a direction on the welcome screen: Business requirement (questions in business language; uses the backlog and business rules only at first — lower token use, recommended) or Technical implementation (scans the codebase from the start for tech leads — higher token use; the app asks you to confirm before selecting it). Then describe the requirement in a sentence, or paste an existing ticket link to analyse and improve it.

2

It gathers context first

Before asking you anything, it searches the backlog for duplicates and related tickets and reads the business rule documents. For Technical implementation it also scans relevant parts of the product code from the start; for Business requirement the codebase is used later, when drafting the Solution Design subtask. What it finds shapes the questions, and you'll notice it citing what it found ("I see an existing roles table…", "the eligibility workbook caps this at…").

3

The interview

Targeted questions, a few at a time, usually as numbered options. They cover only the genuine gaps the sources couldn't answer: success criteria, edge and error cases, who's allowed to do what, scope boundaries. "I don't know" is a valid answer: it's recorded as an open question for Solution Design, never guessed. If your answers contradict each other, it surfaces the conflict openly instead of silently picking a side.

4

Quality check & acceptance criteria

The draft is validated against a quality checklist: clear scope, testable, nothing critical missing. The acceptance criteria are written as Gherkin scenarios (Given / When / Then) covering every rule and boundary, with the real values cited from the business rules.

GATE: the requirement must pass the quality check ↺ not ready yet → back to the interview with targeted questions
5

Your approval

You get a preview of the ticket(s) exactly as they will appear in your project tool: type and summary, description, Gherkin acceptance criteria and the subtasks table. The preview is just a draft in chat, no ticket exists yet. Ask for changes as many times as you like.

GATE: nothing is created without your explicit yes ↺ you request changes → it loops back and revises
6

Created in your project tool

The work item and its discipline subtasks are created, the acceptance criteria travel with the ticket (and as a .feature file for the test suite), and a record of the whole interview is attached as a comment so every decision is auditable. You get the ticket links back in chat.

7

Business ↔ technical handoff (optional)

After a business requirement: Elixcit asks whether you want the technical breakdown for the team prepared now. If yes, it scans the codebase and updates the existing subtasks — [Solution Design], [BE], [FE], [Infrastructure], [Testing] — with implementation-ready detail (preview + your approval before anything is written to the ticketing tool).

Starting with technical implementation: it first searches for an existing business requirement for the same feature. If one exists, it offers to update that ticket's subtasks instead of creating a duplicate.

What you get at the end

  • A correctly typed work item under the right Epic, following the team's conventions and labels.
  • The six discipline subtasks: [Solution Design], [BE], [FE], [Infrastructure], [Testing], [Business Sign Off].
  • Gherkin acceptance criteria on the ticket and as a .feature file QA can bind automated tests to.
  • The interview record (question → answer) as a ticket comment, so nothing rests on memory.

You can write in Croatian or any other language and Elixcit will understand you, but it always answers in English. Tickets and acceptance criteria are always written in English.

3 Good prompts

A good opening prompt names who needs what capability and why. You don't need to know the technical details, that's what the interview is for.

✓ Good: clear actor, capability and outcome
Admins should be able to waive cancellation fees for a customer, with a reason recorded for audit.
Why it works: names the actor (admins), the capability (waive fees) and a constraint (audited reason). Elixcit can immediately check for related tickets and ask sharp follow up questions instead of basics.
✓ Good: improving an existing ticket
Analyse PROJ-2112. What's missing before the team can pick it up?
Why it works: pasting a ticket key or link puts Elixcit in review mode: it audits the ticket against its quality checklist and proposes concrete improvements.
✓ Good: brings the business rule with it
Instalments should only be offered to eligible customers. The rules are in the Payment Eligibility workbook (age, Delphi score, credit utilisation).
Why it works: points at the source of truth. Elixcit reads the workbook and writes acceptance criteria with the real thresholds (e.g. "Delphi score < 600 → declined") instead of placeholders.
✓ Good: scoped change to existing behaviour
When a policy renewal fails payment, the customer currently gets no notification. They should get an email within an hour, and support should see the failure in the admin dashboard.
Why it works: describes current behaviour, desired behaviour and both affected surfaces. The interview can go straight to edge cases (retries? bounced emails?) instead of reconstructing the basics.
✓ Good: honest about what you don't know
We need role based access control for the admin dashboard. I know which roles we need, but not how permissions should be enforced technically.
Why it works: declaring what you can and can't answer up front lets Elixcit ask you the role questions and park the enforcement questions on the [Solution Design] subtask for the tech lead.

4 Bad prompts, and how to fix them

None of these are "wrong" and the interview will eventually recover, but they waste your message budget on questions you could have avoided.

✗ Too vague: no actor, no capability
Make the app better for customers.
Problem: there is nothing to elicit against, so every question would be a guess.
Fix: "Customers abandon the quote flow at the payment step; we want to let them save a quote and finish it later."
✗ Several features in one prompt
Add RBAC to the admin dashboard, email notifications for renewals, and a monthly claims report export.
Problem: those are three separate requirements with three separate interviews; mixing them produces one mushy ticket.
Fix: run three conversations, one per requirement. Use New conversation between them.
✗ A solution disguised as a requirement
Add an "is_eligible" boolean column to the customers table.
Problem: that's an implementation decision, not a requirement. The team can't test "a column exists" against a business outcome.
Fix: state the outcome, "Instalments must only be offered to eligible customers", and let the design land in [Solution Design]. (If you genuinely want an engineering task, pick the Technical implementation direction.)
✗ Asking for things it doesn't do
Write the code for the fee waiver feature and deploy it.
Problem: Elixcit produces requirements and tickets, not application code or deployments.
Fix: "Admins should be able to waive cancellation fees". The resulting ticket is what the engineering team implements.
✗ Answering questions it didn't ask, all at once
(three pages of pasted notes, old emails and half decisions)
Problem: contradictory or stale content forces the agent to untangle instead of elicit, and the conversation burns through its message limit.
Fix: start with the requirement in one sentence; paste specific excerpts (a rules table, a decision) when a question asks for them.
✗ Sensitive data in the chat
Here's a real customer record so you get the idea: John Smith, DOB 12/04/1985, card 4532…
Problem: never paste real personal data, credentials or card numbers. The requirement doesn't need them.
Fix: use invented examples: "a customer born in 1985 with an average credit score".

5 Answering the questions well

  • Answer what you know, park what you don't. "I don't know" is recorded as an open question on [Solution Design], it is never guessed.
  • Reply with numbers. When options are numbered, "1 and 3" is a perfectly good answer.
  • Cite sources instead of memory. If a rule lives in a spreadsheet, name the document or paste the relevant rows so the real values end up in the acceptance criteria.
  • Flag contradictions honestly. If your answer changes an earlier one, say so. The agent reconciles it explicitly rather than silently picking a side.
  • Push back on the preview. The preview step is exactly for that: "make scenario 2 stricter", "drop the FE subtask", "rename the summary". Nothing is created until you approve.

A realistic exchange

Example
YouAdmins should be able to waive cancellation fees
ElixcitI found PROJ-2197 (fee waiving, In Progress). Is your feature 1) an extension of it, 2) a change to it, or 3) separate? Also: which fees, cancellation only or admin fees too?
You3. Both fee types. Waiver needs a reason from a dropdown and an audit trail.
ElixcitWho can waive, any admin or a specific role? And is there a design for the admin screen?
YouAuthorised admins only. No design yet, and I don't know how enforcement should work technically.
ElixcitNoted, enforcement details go to [Solution Design] as open questions. …full ticket preview with Gherkin scenarios and 6 subtasks… Nothing is created yet. Shall I create these tickets?
YouYes, create.

6 Do & don't, at a glance

Guideline
DoOne requirement per conversation; start a new conversation for the next one.
DoName the actor, the capability and the business outcome in your first message.
DoPaste ticket links to analyse or improve existing tickets, and name business rule documents when relevant.
DoSay "I don't know" when you don't. It gets recorded, not invented.
DoReview the preview critically before approving; iterate as often as needed.
Don'tBundle several features into one prompt.
Don'tPrescribe the implementation when you mean a business requirement.
Don'tPaste real customer data, passwords or API keys. Ever.
Don'tAsk it to write code, estimate story points or deploy anything.
Don'tApprove a preview you haven't read. Approval is the moment tickets get created.

7 Limits & troubleshooting

SituationWhat's going on
"Message limit reached (30/hour)"Each user can send 30 messages per hour, as a fairness and cost guard. Take a break and continue; your conversation is kept for an hour.
Technical direction confirmationChoosing Technical implementation triggers a warning: scanning the codebase uses more AI tokens and costs more than Business requirement. Confirm only if you need the ticket grounded in code now.
"This conversation is very long"Conversations cap at 40 messages. Approve or wrap up, then start a New conversation. Long threads produce worse tickets anyway.
"Daily message budget" reachedThe whole service has a daily cap. Try again tomorrow, or ask the team to raise it.
Business then tech (or the reverse)After a business requirement is created, Elixcit asks if you want the BE/FE/Infra subtasks filled in from the codebase. With technical implementation, it searches for an existing business ticket first and offers to update it instead of duplicating.
Logged out unexpectedlyLogin tokens last 12 hours, and a service restart logs everyone out. Just log in again. An unfinished conversation is lost, but any created tickets are already in your project tool.
"Ticketing tool not connected" pillElixcit still works fully; instead of creating tickets it gives you markdown you can paste straight into your project tool.
It created something wrongTell it in chat and it will transition the item to Cancelled and note it (deletion is often not permitted in project tools). Every item it creates carries the agent's provenance label, so it's always identifiable.
Start a conversation →