Skip to content

Chat Mode

Chat Mode is Amplify's lightweight conversation mode. It uses a very small system prompt containing only the parts necessary to talk to the AI and get great results, and it disables some skills and tools that aren't relevant to a quick back-and-forth. The goal is simple: save tokens so your API budget goes further on casual Q&A, brainstorming, and explanations.

What chat mode is for

Chat mode is the default for ad-hoc, non-project conversations — the kind where you just want to ask a question, get an explanation, or bounce an idea around. It is not a "dumber" mode; the same model and the same streaming pipeline are used. What changes is how much context Amplify injects around your message.

ConcernChat ModeAgent Mode
System promptMinimal — identity + output rules onlyFull — identity, skills, project context, memory, tools
SkillsA reduced setAll available skills
ToolsCore only (no inject_template, no sandbox mutation)All tools, including sandbox + template injection
Token budget per turnLowHigh
Best forQ&A, explanations, brainstormingBuilding, editing, running code

The minimal system prompt

In chat mode the system prompt is stripped down to the essentials. The full construction logic lives in the AI Workflow doc; in chat mode only these sections are included:

  1. Identity — who Amplify is and how it should format answers.
  2. Output rules — markdown, code fences, concise answers.
  3. Available tools (trimmed) — only the read-only / conversational tools.

Everything else is omitted:

  • <available_skills> is either empty or contains only a tiny subset.
  • <user_memory> and <user_context> are still attached (they're small and high-value), but <project_context> is not, because chat mode is not project-scoped.
  • Skill-discovery instructions ("you MUST first call list_skills…") are removed, since most skills are disabled anyway.

The result is a system prompt that's typically a few hundred tokens instead of several thousand — a direct saving on every single turn of the conversation.

Disabled skills & tools

To keep the token budget low, chat mode disables capabilities that only make sense when you're actively building something in the sandbox.

  • Disabled tools

    The following tools are not exposed to the model in chat mode:

    • inject_template — cloning a starter template into the workspace is pointless without a project.
    • Sandbox mutation tools (file writes, shell exec, package install) — there's no project to mutate.
    • list_design_systems / get_design_system — design systems are only relevant when generating UI code.

    Read-only tools (web search, reading a file you attached, etc.) remain available.

  • Disabled skills

    Heavy design and build skills are filtered out:

    • webapp-builder, mobile-app-development, react-native-component, react-component
    • html-page, docx (document generation)
    • Any user_skills/ skill tagged build-only

    Lightweight, conversational skills (e.g. api-integration for answering "how do I paginate a REST API?") stay enabled.

Why disable instead of hide?

Skills and tools are part of the system prompt. If they stayed enabled, their descriptions would be sent on every request — costing tokens even when never invoked. Disabling them removes their descriptions from the prompt entirely, which is where the real saving comes from.

Switching to Agent mode mid-chat

You are not locked into chat mode. If the conversation grows beyond a simple Q&A — you want the AI to actually write and run something — you can switch to Agent Mode mid-chat.

When you switch:

  1. The chat system prompt is replaced by the full agent system prompt (skills, project context, all tools).
  2. Your message history is preserved — the AI keeps the full conversation context and continues as if nothing happened, just with more capabilities available.
  3. Only the system prompt changes; the back-and-forth you already had is not re-sent or summarised, so there is no context loss.

This makes chat mode a great starting point: begin cheap, and escalate to the full agent only when the task demands it.

Not available in project chats

The chat ↔ agent toggle is not available in Project Chat. Project conversations always run in agent mode because project context, skills, and sandbox tools are mandatory for working on a codebase.

When to use chat mode

Continue to Agent Mode