Conversational Agent Overview
On this page
Conversational Agent Overview
A chat-based, project-aware co-pilot that edits your spec through natural conversation. · Recommended
Use the agent whenever you want a targeted change, a quick answer about your project, or a one-click generation — without leaving the editor. You can skip it entirely and drive every step from the section buttons, but it is usually the fastest way to refine what the pipelines produce.
What the agent is
The agent lives in the chat panel on the right side of the project editor. It is more than a chatbot: before every response it reads your project's current state — every persona, feature, user story, acceptance criterion, page, and schema entity — so new content fits what already exists and edits respect how entities relate to one another.
What it can do
| Capability | Examples |
|---|---|
| Add entities | Create features, personas, user stories, acceptance criteria, pages, tables |
| Edit entities | Rename a feature, update a persona's goals, refine a story |
| Remove entities | Delete outdated entities — with an impact analysis first |
| Answer questions | "How many features reference the admin persona?" |
| Analyse relationships | "What user stories depend on the authentication feature?" |
| Bulk operations | "Add acceptance criteria to every feature that has none" |
| Suggest improvements | "What's missing from my project?" |
| Run generations | One-click suggested actions that run the same pipelines as the page buttons |
Suggested actions
The panel shows contextual suggested action chips — in the empty state and above the message input. They come in three kinds:
- Pipeline chips (wand/sparkles icon) run a full generation — personas, features, user stories, criteria, pages, schema, or the business case — through exactly the same pipeline as the button on that section's page. Billing, live progress, and version control behave identically. The highlighted chip is the recommended next step for your project's current state.
- Prefill chips (pen icon) drop editable text into the input, such as "Find gaps in my spec" — nothing runs until you press send.
- Check my coverage runs a free, deterministic gap report (features, stories, criteria, pages, personas) as a card in the panel — no tokens spent, no message sent.
Chips are derived from what already exists: a project with features but no user stories suggests generating stories; once pages and criteria both exist, a "Sync criteria from pages" chip cross-checks acceptance criteria against your page layouts. Chip-triggered generations use whichever model you've selected in the input's model picker.
How it understands your project
Every message goes through a structured process:
- Intent classification — is this a question, a generation request, or an edit?
- Context retrieval — the relevant slice of your project is pulled into working memory.
- Impact analysis — for edits, it checks how the change ripples through related entities and warns you.
- Execution — it performs the operation and updates the editor in real time.
When to use the agent vs. regeneration
Both the agent and the section regeneration buttons can create and update content, but they serve different purposes.
Use the agent when you want to:
- Make a specific, targeted change ("Rename the Notifications feature to Alerts")
- Add a single entity with particular requirements ("Add a persona for a compliance officer")
- Ask questions about the project's current state
- Perform a multi-step operation in one request
- Understand how a change would affect the rest of the project
Use regeneration when you want to:
- Generate a full set of entities from scratch (e.g. all personas at once)
- Replace an entire section with fresh AI-generated content
- Start over on a particular step
- Produce content in bulk without specific instructions
💡 Tip: A simple rule of thumb — if you know exactly what you want changed, use the agent. If you want the AI to make broad decisions for you, use regeneration.
Key benefits
- Natural-language control — no forms or dropdowns; describe what you need in plain English.
- Contextual accuracy — the agent references existing data, so new content fits.
- Safety guardrails — destructive operations require confirmation and show impact analysis first.
- Real-time updates — changes appear immediately, with no page refresh.
- Conversation memory — it remembers earlier messages in the session, so you can build incrementally.
⚡ Power-user hints
- The model picker in the input applies to both chat replies and chip-triggered generations — switch to a stronger model for a tricky bulk edit, then switch back.
- Turn on Guided mode to have the agent behave like a product manager: flagging gaps when you land on a section, interviewing you before a generation, and reviewing the output. See 9.1 — Guided mode.
↔ The traditional way
Refining a spec by hand means a product manager editing a PRD document, manually reconciling every downstream story, ticket, and schema note each time a feature changes. The agent collapses that into a sentence — and, unlike a document, it already knows what every other entity is and warns you before a change breaks something.
What's next
- 8.2 — Using the agent — phrasing, example prompts, and how changes are applied.
- 9.1 — Guided mode — the proactive, PM-style layer that builds on the agent.