#System prompt

Every lite-agent run starts from a system prompt that grounds the model in your workspace: where it may work, which tools to prefer, and which skills and subagents exist. The SDK builds a good one for you with buildSystemPrompt — workdir, model name, available skills, and available subagents are all filled in automatically — and lets you replace or extend it when your agent needs its own voice, rules, or domain context.

#Usage

Pass system to createLiteAgent (or systemPrompt to query) to replace the built-in prompt entirely:

import { createLiteAgent } from "@lite-agent/sdk";
import { anthropic } from "@lite-agent/provider";

const agent = createLiteAgent({
  model: anthropic(),
  modelName: "claude-sonnet-4-6",
  workdir: process.cwd(),
  system: "You are a release-notes assistant. Answer concisely, in Markdown.",
});

query() accepts the same string under the name systemPrompt:

import { query } from "@lite-agent/sdk";

for await (const ev of query({
  prompt: "Draft release notes for v0.4.0.",
  model: anthropic(),
  cwd: process.cwd(),
  systemPrompt: "You are a release-notes assistant. Answer concisely, in Markdown.",
})) {
  // ...
}

#Appending to the default prompt

A full override throws away the built-in grounding. To keep the default and add your own rules, call buildSystemPrompt yourself and concatenate — it is the same builder the SDK uses internally, exported from @lite-agent/sdk:

import { createLiteAgent, buildSystemPrompt } from "@lite-agent/sdk";

const agent = createLiteAgent({
  model: anthropic(),
  modelName: "claude-sonnet-4-6",
  workdir: process.cwd(),
  system:
    buildSystemPrompt({ workdir: process.cwd(), skills: "" }) +
    "\n\n## House rules\n- Never commit directly to main.\n- Run pnpm test before finishing.",
});

buildSystemPrompt(opts) takes a SystemPromptOptions object and returns the prompt string:

#What the default prompt contains

buildSystemPrompt produces a compact prompt with these sections:

• Identity & boundary — "You are lite-agent, a coding agent operating in <workdir>"; never access paths outside it.
• Core principles — prefer tools over prose.
• Files — use read_file / write_file / edit_file / delete_file instead of shell equivalents; use bash for running commands and searching.
• Task planning — for 3+ step work, plan with TaskCreate and track with TaskUpdate (see Tasks).
• Skills — pull specialized knowledge on demand with load_skill, followed by the auto-discovered skill list.
• Subagents — when subagents exist, how and when to delegate with the Agent tool (see Subagents).

Two capabilities extend the prompt at assembly time rather than through system:

• Setting outputSchema appends a ## Final answer section instructing the model to call final_answer exactly once.
• The Tasks reminder is injected per turn as a <system-reminder> message, not into the system prompt.

#See also

• Structured output — the ## Final answer section appended when outputSchema is set.
• Tasks — the per-turn task-list reminder.
• Subagents — the subagent list the default prompt surfaces.
• Getting started — install and run your first agent.