System prompt

WednesdayAI assembles the agent’s system prompt from a set of named sections at run time. The default prompt is comprehensive; you can slim it down with a preset, toggle individual sections, inject custom content, and control which project context files are included. All configuration is under agents.defaults.systemPrompt (shared across agents) or agents.<agentId>.systemPrompt (per-agent override). Changes take effect when the gateway restarts. Restart: systemctl --user restart openclaw-gateway.

Presets

The fastest way to change the prompt’s size and scope is a preset:

// ~/.openclaw/openclaw.json
{
  agents: {
    defaults: {
      systemPrompt: {
        preset: "lean",   // "default" | "lean" | "minimal"
      },
    },
  },
}

Preset	Effect
`"default"`	All sections enabled. Full-featured context, longer prompt.
`"lean"`	Disables documentation, runtime, model aliases. Compacts tooling and skills sections. Uses progressive skill rendering. Good balance of capability and token cost.
`"minimal"`	Disables documentation, runtime, model aliases, reply tags, messaging, voice, reactions, silent replies, heartbeats. Tooling renders names only. Skills compact + progressive. Lowest token cost.

Per-agent override

Named agents can override the shared default. The per-agent config merges over agents.defaults.systemPrompt:

{
  agents: {
    defaults: {
      systemPrompt: { preset: "lean" },
    },
    assistant: {
      systemPrompt: {
        preset: "default",   // this agent uses the full prompt
      },
    },
  },
}

Section toggles

Individual sections can be turned off or customised without changing the preset. Use the sections map with the section name as the key:

{
  agents: {
    defaults: {
      systemPrompt: {
        preset: "default",
        sections: {
          documentation: "off",    // remove the slash-command reference
          modelAliases: "off",     // remove the model alias list
          heartbeats: "off",       // disable heartbeat task list injection
        },
      },
    },
  },
}

Section catalogue

Section	Mode type	Content
`identity`	Standard	Agent name, personality, and identity context
`tooling`	Generated	Tool definitions and calling instructions — `"compact"` reduces detail, `"names"` lists only tool names
`toolCallStyle`	Standard	How the agent should format and sequence tool calls
`safety`	Standard	Safety and behaviour guidelines
`cliQuickReference`	Standard	Slash-command cheat sheet (`/new`, `/reset`, `/model`, etc.)
`skills`	Generated	Loaded skill descriptions — `"compact"` condenses, `"names"` lists names only; `progressive: true` renders on demand
`memory`	Standard	Memory read/write instructions
`selfUpdate`	Standard	Self-update capability instructions
`modelAliases`	Generated	Model name-to-provider alias mappings
`workspace`	Standard	Workspace directory and file layout
`documentation`	Standard	Documentation access and usage instructions — disabled in `lean`/`minimal` presets
`sandbox`	Standard	Sandbox mode restrictions and warnings
`authorizedSenders`	Standard	Who is authorised to give the agent instructions
`time`	Standard	Current date and time
`workspaceFilesNotice`	Standard	Notice about workspace file availability
`replyTags`	Standard	Reply formatting tag instructions
`messaging`	Standard	Messaging-specific formatting rules (disabled in `minimal`)
`voice`	Standard	Voice mode formatting rules (disabled in `minimal`)
`groupContext`	Standard	Group chat context and mention awareness
`reactions`	Standard	Reaction handling instructions (disabled in `minimal`)
`reasoningFormat`	Standard	Reasoning and thinking format instructions
`projectContext`	Config object	Project context files — see Project context files
`silentReplies`	Standard	When to suppress outbound replies (disabled in `minimal`)
`heartbeats`	Standard	Heartbeat task list (injected during heartbeat runs)
`runtime`	Runtime	Runtime and model info — `"minimal"` trims detail; disabled in `lean`/`minimal` presets

Mode type key:

Type	Accepted values
Standard	`"default"` \| `"off"` \| `"prepend"` \| `"append"` \| `"replace"`
Generated	`"default"` \| `"compact"` \| `"names"` \| `"off"`
Runtime	`"default"` \| `"minimal"` \| `"off"`
Config object	`ProjectContextSectionConfig` — see project context files section

Project context files

These files are read from the agent workspace directory and injected into the system prompt:

File	Default mode	Purpose
`AGENTS.md`	`inline`	Workspace context and session startup instructions
`SOUL.md`	`inline`	Agent personality and core values
`IDENTITY.md`	`inline` (max 4 000 chars)	Agent name, creature, vibe, emoji, avatar
`USER.md`	`manifest`	User profile — name, timezone, context
`TOOLS.md`	`manifest`	Local environment notes — devices, SSH hosts
`HEARTBEAT.md`	`runKind`	Periodic task list (injected during heartbeat runs only)

File mode values:

Mode	Effect
`"inline"`	Full file content injected into the prompt (bounded by `maxChars` if set)
`"manifest"`	File path and size listed; content is not injected (avoids bloating the prompt for large files)
`"runKind"`	Injected only during heartbeat-triggered runs, not during interactive chat
`"off"`	File skipped entirely

Default per-file limit: 20 000 characters. Hard ceiling on total prompt source reads: 8 MiB. Missing optional files are silently skipped. Leave HEARTBEAT.md empty (or absent) to skip heartbeat API calls. To change a file’s mode, use the projectContext.files config:

{
  agents: {
    defaults: {
      systemPrompt: {
        sections: {
          projectContext: {
            files: {
              "SOUL.md": { mode: "off" },
              "TOOLS.md": { mode: "inline", maxChars: 2000 },
            },
          },
        },
      },
    },
  },
}

Prompt cache

The prompt is split into a stable (cached) region and a volatile region at build time. The cache boundary controls where the split happens:

{
  agents: {
    defaults: {
      systemPrompt: {
        cache: {
          boundary: "auto",   // "auto" (default) | "off"
        },
      },
    },
  },
}

"auto" lets the runtime choose the optimal cache boundary for the current provider. Set "off" to disable prompt caching entirely (increases token cost, may help debugging).

Injecting content from a plugin

Plugin authors can inject content into the system prompt at run time via the before_prompt_build lifecycle hook — without touching admin config. See Hooks for the hook API; the return value accepts systemPrompt (appended after the built sections) and prependContext (inserted before the user turn).

Identity attribution

WednesdayAI attributes each agent run to a sender when it can, and labels how it resolved that identity. Two config surfaces drive this:

session.identityLinks maps platform-prefixed ids (for example telegram:123456789) to canonical identities. When an inbound sender matches a key, the run is attributed to that canonical identity with identitySource: "sender".
heartbeat.identity (and the equivalent identity field on cron jobs) attributes a scheduled system run to a registered user. The value must case-insensitively match a session.identityLinks key at run time; a match resolves with identitySource: "job-config", an absent or non-matching value leaves the run as a system run with no identity. Subagent runs may inherit the requester’s identity, validated against the same identityLinks keys (identitySource: "parent-session").

{
  session: {
    identityLinks: {
      alice: ["telegram:123456789", "discord:987654321012345678"],
    },
  },
  agents: {
    defaults: {
      heartbeat: {
        every: "30m",
        identity: "alice",   // attribute heartbeat runs to alice (must match an identityLinks key)
      },
    },
  },
}

Resolution precedence: inbound sender match → inherited requester identity (validated) → configured job/heartbeat identity → identity-less. Each successful injection beyond the sender path emits one runtime identity injected info log — the live observable for attribution. For model-provider calls, WednesdayAI derives opaque correlation ids rather than forwarding raw identities: providerConversationId (hashed agent + session) and hashedUserId (one-way hash of the sender identity). The provider prompt-cache key is deliberately keyed on agent + model only, never on conversation or user, so the shared tools+system prefix keeps one cache entry across every conversation under that agent and model. Related: Sessions — Identity and privacy.

Hooks — before_prompt_build and context.collect hook reference
Gateway configuration — gateway-wide settings
Agent tools — exposing capabilities to the agent
Sessions — session keys, reset lifecycle, and privacy boundary

Getting started

Gateway configuration

Channels

Automation

Observability

Core concepts

AI providers

Security

Maintenance

Admin

Channels (synced pages)

Gateway (synced pages)

Gateway security

Installation

Platforms

macOS app

More providers

Security

Web UI

Shohin

System prompt

System prompt

Presets

Per-agent override

Section toggles

Section catalogue

Project context files

Prompt cache

Injecting content from a plugin

Identity attribution

​System prompt

​Presets

​Per-agent override

​Section toggles

​Section catalogue

​Project context files

​Prompt cache

​Injecting content from a plugin

​Identity attribution

​Related

System prompt

Presets

Per-agent override

Section toggles

Section catalogue

Project context files

Prompt cache

Injecting content from a plugin

Identity attribution

Related