Project Chat

Why project chat is always agent mode

A project chat needs everything chat mode strips out:

• <project_context> — the project's file tree, open files, and runtime state. Without it the AI can't reason about your codebase.
• The full tool set — inject_template, sandbox read/write/exec, package install. Building requires mutation.
• All skills — react-best-practices, mobile-app-development, webapp-builder, etc. The right skill has to be discoverable or output quality collapses.
• The planning engine — multi-step builds ("add auth, then a dashboard, then wire them up") need decomposition + sub-chat workers + FlowVerifier.

Removing any of these to save tokens would make project chats useless — you'd be talking about a project without the context or tools to act on it. So the toggle is hidden entirely in project chats; the full agent system prompt is always attached.

What's in the project context

The <project_context> block injected into the system prompt includes:

The selectStarterTemplate flow (see Templates) bootstraps this context the moment you pick a template — it clones the repo, installs deps, and starts the dev server so the AI has a live, runnable project from turn one.

Per-project skill filtering

Not every skill is relevant to every project. A Vue project doesn't need the React skill; an API backend doesn't need react-native-component. Amplify filters skills per project via getProjectSkills(projectId) from the projectSkillsStore.

// Per-project skills: the project can toggle skills on/off in settings.
// Only enabled skills appear in <available_skills> for that project's chats.
const skills = getProjectSkills(projectId);
// → ["react-best-practices", "frontend-design", ...] (project's enabled subset)

You can toggle skills on or off per project in the project Settings panel. Disabling a skill for a project removes it from that project's <available_skills> block — it won't be discovered or loaded, saving tokens on every turn of that project's chats. Other projects are unaffected.

• How per-project filtering interacts with the skill directories

  Skills are still loaded from the three directories (app/lib/skills/, design/skills/, user_skills/) globally by the SkillLoader. The per-project filter is applied after loading: getProjectSkills(projectId) returns the intersection of loaded skills and the project's enabled list. So:

  • A skill must be installed (in one of the 3 dirs) and enabled for the project to appear.
  • Uninstalling a skill removes it from all projects.
  • Disabling a skill for one project keeps it available everywhere else.

  See Built-in Skills → Skill directories for the loading process.

Memory & vector search per project

Project chats also have access to Amplify's Vector Database for long-term memory scoped to that project. This lets the AI recall decisions, file relationships, and prior fixes across sessions — even after the WebContainer resets.

• Context persistence: BM25 + embeddings over the project's files and past chat turns, stored in IndexedDB.
• Project-scoped: memories are keyed by projectId, so switching projects swaps the memory store.
• Budget-aware: a 70% context-budget threshold controls how much memory is injected per turn (see AI Workflow).

This is why project chats are heavier on tokens than standalone chats — but also why they're dramatically more useful for real engineering work.

When to use project chat

For quick questions that don't need the codebase, start a standalone chat (which defaults to Chat Mode) instead — it'll be cheaper and just as accurate for pure Q&A.