| name | PuerTS Agent Development Guide |
| description | Guide for developing LLM agents based on PuerTsAgent framework — covers resource directory structure, system-prompt, skills, builtin modules, and best practices. |
PuerTS Agent Development Guide
This skill provides the complete guide for developing LLM agents based on the PuerTsAgent framework.
⚠️ Important — Required Packages: This framework depends on PuerTS core, v8, and agent packages. Before proceeding, you must check the Unity project's Packages/manifest.json to verify that the following packages are installed:
com.tencent.puerts.core
com.tencent.puerts.v8
com.tencent.puerts.agent
If any of these packages are missing, they must be installed first before using this skill. you MUST first invoke the "Install PuerTS for Unity" skill to AUTO complete the setup.
⚠️ Prerequisite: Before using this skill to create or modify an agent, you must know the agent's resource directory path under Unity Resources/ (e.g., maze-runner, editor-assistant). If the user has not provided this path, ask the user which Resources/ subdirectory to use before proceeding. All agent files (system-prompt, skills, builtins) are organized under this path.
⚠️ Builtin Language Choice: When creating builtin modules, you must ask the user whether they want to write builtins in TypeScript or JavaScript:
- TypeScript → Create a dedicated TS project (with
package.json, tsconfig.json, esbuild.mjs) whose build output (outdir) points to Assets/Resources/<agent-name>/builtins/. See TypeScript Project Setup for Builtins section below.
- JavaScript → Write
.mjs files directly in Assets/Resources/<agent-name>/builtins/.
Overview
In this framework, an agent's behavior and capabilities are entirely defined by files in a resource directory (Resource Root). Different resource directories represent different agents — Unity editor assistants, in-game AI characters, or any role you need.
Key Concept: Both system-prompt.md.txt and files under skills/ are essentially prompts (natural language instructions for the LLM). The only difference is their loading strategy:
system-prompt.md.txt — Always present in context (persistent). Put prompts here that are needed for most tasks.
skills/*.md.txt — Loaded on demand by the LLM via loadSkill. Put prompts here that are not needed for every task — they are loaded only when the LLM determines they are relevant.
Creating an Agent
Initialize an agent in C# via its resource path:
var agent = new AgentScriptManager();
agent.Initialize("maze-runner", () =>
{
Debug.Log("Maze runner agent is ready!");
});
The first parameter of Initialize is the resource directory path under Unity Resources/. The framework loads the full agent definition from this directory:
system-prompt.md.txt — Role definition (System Prompt)
skills/ — Domain skill documents (loaded on demand, optional)
builtins/ — Built-in helper modules (executable JS modules)
Example Directory Structure
Resources/maze-runner/
├── system-prompt.md.txt # Role definition (System Prompt)
├── skills/ # Domain skills (loaded on demand, optional)
└── builtins/ # Built-in modules (loaded on demand)
├── maze-control.mjs # Maze movement and status query
└── screenshot.mjs # Screenshot to observe the maze
All resource files are placed in Unity's Resources/ directory.
File extension convention: Unity Resources does not support .md and .mjs as TextAsset, so:
- Markdown files use
.md.txt suffix
- JS module files use
.mjs suffix
1. System Prompt — Role Definition
Purpose
system-prompt.md.txt defines the agent's identity, personality, capabilities, and behavioral rules. It is injected at the beginning of the LLM's system prompt and is the core of the agent's "persona."
This is a persistent prompt — it is always present in the LLM's context for every conversation turn. Therefore, put only the most frequently used instructions here — things the agent needs for the majority of its tasks. Avoid stuffing rarely-used domain knowledge here, as it permanently consumes context tokens.
File Location
<resource-root>/system-prompt.md.txt
Format
Plain text / Markdown format, write role definition directly, no front-matter needed.
Example — Editor Assistant
You are a helpful AI assistant running inside Unity via PuerTS (a TypeScript/JavaScript runtime for Unity). You can help with game development, scripting, and general questions. Be concise and practical.
Example — Maze AI (excerpt)
You are a Maze Explorer AI — an intelligent agent that navigates through 3D mazes by observing, reasoning, and acting.
## Your Capabilities
You can control a player character in a 3D maze using two builtin modules:
- **maze-control**: Move in compass directions (north/south/east/west) with `movePath()` and query obstacle distances with `getPlayerStatus()`
- **screenshot**: Capture the game view to visually observe the maze
## Goal Description
Your goal is to reach the **maze exit marker** — a tall RED pillar with a bright RED glowing ring.
The System Prompt describes the AI's identity, available modules, goals, exploration loops, navigation rules, etc. It can be lengthy — it defines the AI's behavior pattern throughout the task.
2. Skills — Domain Skills
Purpose
A Skill is an on-demand loaded Markdown document providing domain-specific operation guides and rules for the LLM. The LLM actively loads skills into context via the loadSkill tool when needed.
Skills are prompts too — just like system-prompt.md.txt, they are natural language instructions. The key difference is that skills are not always in context. If a piece of knowledge is not needed for most tasks the agent handles, it belongs in a skill file rather than in the system prompt. This keeps the base context lean while still making specialized knowledge available when the LLM needs it.
File Location
<resource-root>/skills/<skill-name>.md.txt
File Format
Skill files use YAML front-matter for metadata, body is Markdown content:
---
name: <skill-id>
description: "<one-line description shown to the LLM>"
---
(Markdown body — the actual skill instructions)
| Front-matter Field | Required | Description |
|---|
name | ✅ | Unique identifier, LLM calls loadSkill with this name |
description | ❌ | Short description shown in loadSkill available skills list |
Example
skills/puerts-interop.md.txt:
---
name: puerts-interop
description: "PuerTS JS ↔ C# interop rules: CS/puer globals, out/ref params, generics, operators, Array/List indexer access, async/Task."
---
...
Design Tips
- description should be accurate: LLM decides whether to load based on description
- content should be thorough: skill docs are the LLM's only reference for specific tasks
- one domain per skill: keep responsibilities single for precise loading
- system-prompt vs skill decision rule: If the agent needs this knowledge in >80% of tasks → put it in
system-prompt.md.txt. Otherwise → put it in skills/ for on-demand loading
3. Builtin — Built-in Helper Modules
Purpose
Builtin modules are JavaScript modules providing pre-built helper functions for the LLM's evalJsCode tool. Unlike Skills, Builtins are actually executable code, not documentation.
File Location
<resource-root>/builtins/<module-name>.mjs
Source code can use TypeScript, JavaScript, or any language compilable to .mjs. The framework only cares about the final .mjs files in the resource directory.
Module Contract
Every Builtin module must export the following two string constants:
| Export | Type | Description |
|---|
summary | string | Short summary (~one line), always shown in evalJsCode tool description |
description | string | Detailed function signatures and usage, LLM reads via import() |
Additionally, the module's exported functions are the actual capabilities the LLM can call in evalJsCode.
Example — builtins/unity-log.mjs (shown as TypeScript source)
export const summary = `**unity-log** — Unity console log access (retrieve and summarize recent logs). Read \`.description\` to see available functions and their signatures.`;
export const description = `
- **\`getUnityLogs(count?, logType?)\`** — Get recent Unity console logs.
- \`count\` (number, default 20): Number of log entries to retrieve (1-50).
- \`logType\` (string, default \`'all'\`): Filter by type — \`'all'\`, \`'error'\`, \`'warning'\`, or \`'log'\`.
- Returns an array of log entry objects: \`{ timestamp, type, message, stackTrace? }\`.
`.trim();
export function getUnityLogs(count: number = 20, logType: string = 'all'): LogEntry[] {
const logsJson = CS.LLMAgent.UnityLogBridge.GetRecentLogs(count, logType);
return JSON.parse(logsJson);
}
Example — builtins/maze-control.mjs (TypeScript source, excerpt)
export const summary = `**maze-control** — Control the player in the maze. \`movePath([{dir, steps}, ...])\` executes a multi-segment path. \`getPlayerStatus()\` returns position and obstacle distances. Read \`.description\` for details.`;
export const description = `
- **\`movePath(segments)\`** — Move the player along a multi-segment planned path.
- \`segments\` (array): Array of \`{ dir: string, steps: number }\`.
- Returns: \`{ success, stepsCompleted, blocked, reachedGoal, position, message }\`
- **\`getPlayerStatus()\`** — Get the player's position and obstacle distances.
- Returns: \`{ position, northDistance, southDistance, eastDistance, westDistance, reachedGoal }\`
`.trim();
export async function movePath(segments: PathSegment[]): Promise<MoveSequenceResult> {
const directionsJson = JSON.stringify(segments.map(s => s.dir));
const distancesJson = JSON.stringify(segments.map(s => s.steps));
const resultJson = await new Promise<string>((resolve, reject) => {
CS.LLMAgent.MazePlayerBridge.MoveSequenceV2(directionsJson, distancesJson, (json: string) => resolve(json));
});
return JSON.parse(resultJson);
}
export async function getPlayerStatus(): Promise<PlayerStatusResult> {
const resultJson = await new Promise<string>((resolve, reject) => {
CS.LLMAgent.MazePlayerBridge.GetPlayerStatus((json: string) => resolve(json));
});
return JSON.parse(resultJson);
}
Design Tips
- summary should be brief: it always consumes tokens, just convey the module's purpose
- description must list complete signatures: LLM must read description before calling functions
- functions should validate parameters: LLM may pass incorrect params, validate and throw meaningful errors
- top-level await is supported: modules can use
await at top level (e.g., async initialization)
- C# bridge classes must be pre-implemented on the Unity side
4. Complete Development Workflow
Step 1: Create Resource Directory
Resources/<agent-name>/
├── system-prompt.md.txt
├── skills/
└── builtins/
Step 2: Write Role Definition
Edit system-prompt.md.txt to define the AI's identity and behavioral rules. For example, a maze AI should describe:
- Role identity (maze explorer)
- Available modules (maze-control, screenshot)
- Goal (reach the red exit marker)
- Exploration loop (observe → plan → act)
- Navigation strategy (right-hand rule, dead-end detection, etc.)
Step 3: Write Domain Skills (Optional)
Create .md.txt files in the skills/ directory with YAML front-matter and skill content.
The Maze Demo doesn't use skill files — all navigation rules are directly in system-prompt. When domain knowledge is extensive and doesn't need to be loaded every time, splitting into skills is more appropriate.
Step 4: Write Helper Modules
First, ask the user whether to write builtins in TypeScript or JavaScript:
- TypeScript: Create a dedicated TS project following the TypeScript Project Setup for Builtins template (Section 6). The project's build output will automatically go to
Resources/<agent-name>/builtins/.
- JavaScript: Write
.mjs files directly in Assets/Resources/<agent-name>/builtins/.
Step 5: Initialize the Agent
In C# code, initialize the agent with the corresponding resource path (see "Creating an Agent" section above).
5. Best Practices
System Prompt
- Keep it concise and clear, avoid excessive length (it always consumes context tokens)
- Focus on "who you are" and "what you should do", don't include framework runtime details
- Only include prompts needed for the majority of tasks — this is always-on context
Skills
- Split complex domain knowledge into separate skill files
- Include keywords in description to help LLM decide when to load
- The more detailed the better — this is the LLM's only reference for the task
- Remember: skills and system-prompt are both prompts; the difference is persistent vs on-demand loading
Builtins
- Each module focuses on one functional area (e.g., scene operations, log queries)
- Summary should be one line, description must list all function signatures completely
- All exported functions should validate parameters and throw meaningful errors
General
- One resource directory = one independent agent role
- Different agents are completely isolated, can have entirely different skills and toolsets
6. TypeScript Project Setup for Builtins
When the user chooses TypeScript for builtins, create a dedicated TS project with the following structure and files. Replace <agent-name> with the actual agent name (e.g., maze-runner).
Project Structure
Ts<AgentNamePascalCase>/ # e.g., TsMazeRunner/
├── package.json
├── tsconfig.json
├── esbuild.mjs # Build script — outputs to Resources/<agent-name>/builtins/
└── src/
└── builtins/
└── <module-name>.mts # TypeScript source files
package.json Template
{
"name": "<agent-name>",
"version": "1.0.0",
"description": "<Agent description> - TypeScript builtins",
"private": true,
"scripts": {
"build": "node esbuild.mjs",
"typecheck": "tsc --noEmit"
},
"devDependencies": {
"esbuild": "^0.25.0",
"typescript": "^5.4.0"
}
}
tsconfig.json Template
{
"compilerOptions": {
"target": "esnext",
"module": "nodenext",
"moduleResolution": "nodenext",
"rootDir": "src",
"strict": true,
"esModuleInterop": true,
"skipLibCheck": true,
"forceConsistentCasingInFileNames": true,
"declaration": false,
"sourceMap": false,
"noEmit": true,
"typeRoots": ["../Assets/Gen/Typing", "../puerts/unity/upms/core/Typing"]
},
"include": ["src/**/*.mts", "src/**/*.d.ts", "../Assets/Gen/Typing/csharp/index.d.ts"],
"exclude": ["node_modules"]
}
esbuild.mjs Template
import { build } from 'esbuild';
import { readdirSync, mkdirSync, existsSync } from 'fs';
import { join } from 'path';
const builtinSrcDir = 'src/builtins';
const builtinOutDir = '../Assets/Resources/<agent-name>/builtins';
const builtinFiles = existsSync(builtinSrcDir)
? readdirSync(builtinSrcDir).filter(f =>
f.endsWith('.mts') && !f.endsWith('.d.mts'))
: [];
if (builtinFiles.length > 0) {
if (!existsSync(builtinOutDir)) {
mkdirSync(builtinOutDir, { recursive: true });
}
const builtinEntries = builtinFiles.map(f => join(builtinSrcDir, f));
await build({
entryPoints: builtinEntries,
bundle: true,
format: 'esm',
outdir: builtinOutDir,
outExtension: { '.js': '.mjs' },
platform: 'neutral',
target: 'esnext',
sourcemap: false,
minify: false,
keepNames: true,
external: [],
define: {
'process.env.NODE_ENV': '"production"',
},
});
console.log(`[esbuild:<agent-name>] Built ${builtinFiles.length} builtins module(s) → ${builtinOutDir}/`);
} else {
console.log('[esbuild:<agent-name>] No builtins modules found in src/builtins/');
}
Setup Steps
After creating the project files:
- Run
npm install in the project directory to install dependencies
- Write
.mts source files in src/builtins/
- Run
npm run build to compile and output .mjs files to Assets/Resources/<agent-name>/builtins/
- Run
npm run typecheck for type checking (optional but recommended)