Skip to main content
Bootstrapping is the preparation ritual before each Agent run: assembling the System Prompt, loading tools, injecting user profile and memories, ensuring the Agent starts with the correct identity, capabilities, and context.

What Bootstrapping Does

On each Agent invocation, Zeus performs the following bootstrapping flow:
  1. Assemble System Prompt — Concatenate core prompts from multiple Bootstrap files in order
  2. Select Mode Prompt — Inject corresponding behavioral guidelines based on current mode (Agent / Ask / Plan)
  3. Dynamic Context Injection — Inject current time, user/project profile, relevant memories, Skills metadata, etc.
  4. Load Tool Set — Filter by mode and assemble available tools

Bootstrap Files

Zeus’s System Prompt uses a layered architecture composed of multiple Markdown files. Each file handles an independent responsibility and is concatenated in order at startup.

Core Files

FileResponsibilityContent Summary
CORE.mdCore identity & capability declarationProgramming languages, information processing, content creation, file handling, limitations
SOUL.mdPersonality & valuesProfessional text style (no emoji), accuracy, privacy, security principles
TOOLS.mdTool usage guideFive-layer tool hierarchy description and selection rules, dynamically replaces current available tool list
WORKFLOW.mdTask processing flowQuick response vs complex task, task planning rules, completion criteria
MEMORY.mdMemory system instructionsTrigger words, tool usage, type selection, scope, priority rules

System Prompt Full Description

View detailed content and design principles for each Bootstrap file

Mode Prompts

Different behavioral prompts are injected based on the current interaction mode:
FileModeBehavior
agent_mode.mdAgentFull tool access, proactive execution and modification
ask_mode.mdAskRead-only tools, analysis and Q&A, guides mode switching
plan_mode.mdPlanRead-only (no execution), research and planning, outputs structured plans
Modes take effect immediately at the invoke() entry point. In Ask mode, write tools are replaced with placeholders (progressive disclosure) — the Agent knows the tools exist but cannot call them, and guides the user to switch modes.

Dynamic Injection

On each run, the following context is dynamically injected into the System Prompt:
Injection ItemDescriptionSource
Current timeLets the Agent perceive timeSystem clock
User profileSummary of user preferences, skills, habitsMemory Profile
Project profileCurrent project’s tech stack, constraintsMemory Profile
Relevant memoriesHistorical memories related to current conversationMemory Search
Skills metadataAvailable skill names and descriptions (without full content)SkillsManager
Connector SkillsBrowser / Desktop control instructionsInjected based on connection status
MCP promptsUser-selected MCP Prompt templatesMCP servers
Resource filesContext files referenced by user via @Frontend request
Sandbox filesList of existing files in sandboxSandbox state
Chat attachmentsUser-uploaded images, documentsFrontend request

Skills Progressive Disclosure

Skills use a two-phase loading strategy to reduce initial token consumption:
PhaseContentToken Cost
Discovery (startup)Skill name + one-line descriptionVery low
Activation (on-demand)Full SKILL.md contentOn demand
Execution (on-demand)Script files + reference materialsOn demand
At startup, only the metadata list is injected. The Agent loads full content on demand via the load_skill tool.
  • Agent Loop — After Bootstrapping completes, enters the execution loop: Agent Loop
  • Workspace — Workspace initialization and file storage: Agent Runtime