Guided mode

A proactive, PM-style layer on top of the agent that flags gaps, reviews output, and points at the next step. · Recommended

Guided mode turns the agent from a tool you summon into a co-pilot that speaks up at the right moments. It's on by default for every project and fully skippable — flip one toggle to "I'll drive" and the traditional route is completely unaffected. Everything on this page is deterministic and free on every plan, including Free and Starter; the paid interview layer is covered separately in Interviews and decisions.

The toggle

At the top of the agent panel is a per-project switch:

• Guide me (default) — the agent proactively shows briefs, reviews, and coverage prompts.
• I'll drive — the agent only responds when you message it. No cards, no prompts.

Figure: The Guide me / I'll drive toggle.

The choice is saved per project, so you can guide one project and drive another. When the toggle is off, the panel renders nothing but the toggle row itself — there is no behaviour change to the section buttons or the chat.

The first time you open a guided project you'll see a one-time intro card explaining what guided mode does, with two buttons — Guide me and I'll drive. Pick either; the card never shows again.

What guided mode does

When guided is on, the agent contributes three kinds of card. Each one is computed locally from your project's current state, so producing it costs no tokens. But a one-line summary of every brief and review you're shown is recorded into the conversation transcript — that's deliberate, so the agent's next reply already knows what you were just told instead of asking you to repeat it. The practical consequence: those summaries then travel with the conversation as ordinary context on later turns.

1. Arrival brief — "Where you are"

When you land on a section (Features, Personas, Pages, Schema, and so on), the agent posts a short brief pointing at what's thin and the single best next step. Examples:

• On Features with none yet: "No features yet — they're the spine everything else hangs off."
• On Pages with orphans: "3 pages aren't tied to any user story — orphan pages tend to become scope creep."

A brief carries up to two one-click chips (the same pipeline chips as elsewhere), so you can act without typing. It's dismissible, and you see at most one per section per visit — if nothing on the current section is noteworthy, the agent stays silent rather than talk for the sake of it.

Figure: An arrival brief with a fix chip.

2. Post-generation review — "Generation review"

After a pipeline finishes, the agent posts a review of what just happened: what was created ("Created 8 user stories, 3 pages") and what's still thin, each with a one-click fix chip. Dismiss it once you've acted or read it.

3. Coverage check — "Check my coverage"

A Check my coverage chip is available on every project route, guided mode or not. It runs a deterministic gap report — features without stories, stories without criteria, pages without stories, personas without stories — and renders it as a card. Because it's pull-based and free, it works even when you're in "I'll drive" mode.

💡 Tip: Run Check my coverage right before Prepare for Dev. Closing gaps first makes the developer handoff far richer.

What changes when guided is on

The cards are the visible half. Guided mode also changes how the conversational agent answers when you ask for its judgement rather than issue an instruction:

• It proposes, instead of only advising. With Guide me on, if acting on its own recommendation means changing the spec and the agent is confident in that change, it stages the change as a confirmation for you to approve or reject — nothing is written until you approve. If it isn't confident, it asks the single question whose answer would most change the outcome. With I'll drive on it stays strictly advisory: it offers the change in one line and waits for you to say the word.
• It won't quietly reverse a decision. Once your decision log has entries, a request that contradicts one is surfaced as a clarification naming the decision it contradicts, rather than being silently planned.

What it costs

Nothing to produce. Arrival briefs, post-generation reviews, and the coverage check are all computed from data the panel already holds — zero LLM tokens, no extra generation charges. The one-line note each leaves in the transcript is written locally too, so it costs nothing to create; it does become part of the conversation, so it counts as context on the messages you send afterwards. Only the optional pre-flight interview in Interviews and decisions calls the model, and that's a Pro-plan feature you can skip on every run.

⚡ Power-user hints

• The chips inside a brief or review submit the exact same task payload as the section buttons — billing, live progress, and version control are identical.
• Guided cards are session-scoped: dismissals reset on a fresh session and never follow you across devices, so a brief you closed can re-arm next time if the gap is still open.
• A brief only reappears on the same section if the underlying signal has changed materially — you won't be nagged with an identical card twice in one session.

↔ The traditional way

On a real team this is the product manager's job: noticing a feature has no stories, reminding everyone to write acceptance criteria before the handoff, and reviewing each artefact as it's produced. Guided mode encodes that discipline into the editor, so a solo founder gets the same "someone's watching the plan" safety net without a second person in the room.

What's next

• Interviews and decisions — the Pro-plan pre-flight interview and the decision log it builds.
• Prepare for Dev — consolidate the spec into a build-ready blueprint.