All projects
03In ProgressCAPACITI training project

AI Chatbot & Conversational Workflow Project

Designing and refining a structured assistant conversation

AI-Assisted Productivity / Conversation Design

An AI-assisted conversational workflow built during CAPACITI training: structuring how an assistant should ask, confirm and respond, then reviewing and refining the instructions based on how the responses actually behaved.

  • CAPACITI AI skill training project — CAPACITI is a training programme, not an employer.
Open the live app
Prompt and instruction designAI assistant toolingWritten response review

01 — The Problem

An assistant conversation that is not deliberately structured drifts: it answers the wrong question, skips the confirmation step, or produces a tone that does not fit the setting.

02 — Objective

Structure a conversational flow with a defined role, defined scope, a confirmation step before acting, and a consistent response format — then refine it against observed responses.

03 — Approach — what I did

  • Defined the assistant's role, scope and the questions it should ask before responding.
  • Wrote instruction text specifying tone, response format and out-of-scope handling.
  • Ran sample conversations and reviewed the responses against the intended behaviour.
  • Revised the instructions where the responses drifted from scope or format.

04 — Tools used

Prompt and instruction designAI assistant toolingWritten response review

05 — Process

User message receivedIntent and scope checkClarifying question where information is missingStructured response in the defined formatOut-of-scope handoff where the request falls outside the defined role

06 — Output

  • A defined conversational flow with role, scope and confirmation steps.
  • Instruction text refined across review rounds.

07 — What I learned

  • How instruction wording changes assistant behaviour more reliably than adding more examples.
  • How to review conversational output against defined criteria rather than impressions.

08 — Limitations

  • This is a training project, not a deployed or production assistant. There are no users, no traffic and no performance figures.
  • No underlying model was built, trained or fine-tuned; existing AI tooling was used.

09 — Skills demonstrated

Prompt engineeringConversation structure designIterative evaluationDocumentation

10 — Project evidence

The assistant built in this project is available as a live app. The workflow diagram on this page is a conceptual illustration of the designed flow, not a capture of the interface.

Live app — smart-desk-sidekick.lovable.app