System prompt

OpenClaw builds a custom system prompt for every agent run. The prompt is OpenClaw-owned and does not use the pi-coding-agent default prompt.

The prompt is assembled by OpenClaw and injected into each agent run.

Provider plugins can contribute cache-aware prompt guidance without replacing the full OpenClaw-owned prompt. The provider runtime can:

• replace a small set of named core sections (
  text

  Copy

  interaction_style
  ,
  text

  Copy

  tool_call_style
  ,
  text

  Copy

  execution_bias
  )
• inject a stable prefix above the prompt cache boundary
• inject a dynamic suffix below the prompt cache boundary

Use provider-owned contributions for model-family-specific tuning. Keep legacy

prompt mutation for compatibility or truly global prompt changes, not normal provider behavior.

The OpenAI GPT-5 family overlay keeps the core execution rule small and adds model-specific guidance for persona latching, concise output, tool discipline, parallel lookup, deliverable coverage, verification, missing context, and terminal-tool hygiene.

Structure

The prompt is intentionally compact and uses fixed sections:

• Tooling: structured-tool source-of-truth reminder plus runtime tool-use guidance.
• Execution Bias: compact follow-through guidance: act in-turn on actionable requests, continue until done or blocked, recover from weak tool results, check mutable state live, and verify before finalizing.
• Safety: short guardrail reminder to avoid power-seeking behavior or bypassing oversight.
• Skills (when available): tells the model how to load skill instructions on demand.
• OpenClaw Self-Update: how to inspect config safely with
  text

  Copy

  config.schema.lookup
  , patch config with
  text

  Copy

  config.patch
  , replace the full config with
  text

  Copy

  config.apply
  , and run
  text

  Copy

  update.run
  only on explicit user request. The owner-only
  text

  Copy

  gateway
  tool also refuses to rewrite
  text

  Copy

  tools.exec.ask
  /
  text

  Copy

  tools.exec.security
  , including legacy
  text

  Copy

  tools.bash.*
  aliases that normalize to those protected exec paths.
• Workspace: working directory (
  text

  Copy

  agents.defaults.workspace
  ).
• Documentation: local path to OpenClaw docs (repo or npm package) and when to read them.
• Workspace Files (injected): indicates bootstrap files are included below.
• Sandbox (when enabled): indicates sandboxed runtime, sandbox paths, and whether elevated exec is available.
• Current Date & Time: user-local time, timezone, and time format.
• Reply Tags: optional reply tag syntax for supported providers.
• Heartbeats: heartbeat prompt and ack behavior, when heartbeats are enabled for the default agent.
• Runtime: host, OS, node, model, repo root (when detected), thinking level (one line).
• Reasoning: current visibility level + /reasoning toggle hint.

OpenClaw keeps large stable content, including Project Context, above the internal prompt cache boundary. Volatile channel/session sections such as Control UI embed guidance, Messaging, Voice, Group Chat Context, Reactions, Heartbeats, and Runtime are appended below that boundary so local backends with prefix caches can reuse the stable workspace prefix across channel turns. Tool descriptions should likewise avoid embedding current channel names when the accepted schema already carries that runtime detail.

The Tooling section also includes runtime guidance for long-running work:

• use cron for future follow-up (
  text

  Copy

  check back later
  , reminders, recurring work) instead of
  text

  Copy

  exec
  sleep loops,
  text

  Copy

  yieldMs
  delay tricks, or repeated
  text

  Copy

  process
  polling
• use
  text

  Copy

  exec
  /
  text

  Copy

  process
  only for commands that start now and continue running in the background
• when automatic completion wake is enabled, start the command once and rely on the push-based wake path when it emits output or fails
• use
  text

  Copy

  process
  for logs, status, input, or intervention when you need to inspect a running command
• if the task is larger, prefer
  text

  Copy

  sessions_spawn
  ; sub-agent completion is push-based and auto-announces back to the requester
• do not poll
  text

  Copy

  subagents list
  /
  text

  Copy

  sessions_list
  in a loop just to wait for completion

When the experimental

tool is enabled, Tooling also tells the model to use it only for non-trivial multi-step work, keep exactly one step, and avoid repeating the whole plan after each update.

Safety guardrails in the system prompt are advisory. They guide model behavior but do not enforce policy. Use tool policy, exec approvals, sandboxing, and channel allowlists for hard enforcement; operators can disable these by design.

On channels with native approval cards/buttons, the runtime prompt now tells the agent to rely on that native approval UI first. It should only include a manual

command when the tool result says chat approvals are unavailable or manual approval is the only path.

Prompt modes

OpenClaw can render smaller system prompts for sub-agents. The runtime sets a

for each run (not a user-facing config):

• text

  Copy

  full
  (default): includes all sections above.
• text

  Copy

  minimal
  : used for sub-agents; omits Skills, Memory Recall, OpenClaw Self-Update, Model Aliases, User Identity, Reply Tags, Messaging, Silent Replies, and Heartbeats. Tooling, Safety, Workspace, Sandbox, Current Date & Time (when known), Runtime, and injected context stay available.
• text

  Copy

  none
  : returns only the base identity line.

When

, extra injected prompts are labeled Subagent Context instead of Group Chat Context.

For channel auto-reply runs, OpenClaw can omit the generic Silent Replies section when the direct/group chat context already includes the resolved conversation-specific

behavior. This avoids repeating token mechanics in both the global system prompt and channel context.

Workspace bootstrap injection

Bootstrap files are trimmed and appended under Project Context so the model sees identity and profile context without needing explicit reads:

• text

  Copy

  AGENTS.md
• text

  Copy

  SOUL.md
• text

  Copy

  TOOLS.md
• text

  Copy

  IDENTITY.md
• text

  Copy

  USER.md
• text

  Copy

  HEARTBEAT.md
• text

  Copy

  BOOTSTRAP.md
  (only on brand-new workspaces)
• text

  Copy

  MEMORY.md
  when present

All of these files are injected into the context window on every turn unless a file-specific gate applies.

is omitted on normal runs when heartbeats are disabled for the default agent or is false. Keep injected files concise — especially , which can grow over time and lead to unexpectedly high context usage and more frequent compaction.

Large files are truncated with a marker. The max per-file size is controlled by

(default: 12000). Total injected bootstrap content across files is capped by (default: 60000). Missing files inject a short missing-file marker. When truncation occurs, OpenClaw can inject a warning block in Project Context; control this with (, , ; default: ).

Sub-agent sessions only inject

and (other bootstrap files are filtered out to keep the sub-agent context small).

Internal hooks can intercept this step via

to mutate or replace the injected bootstrap files (for example swapping for an alternate persona).

If you want to make the agent sound less generic, start with SOUL.md Personality Guide.

To inspect how much each injected file contributes (raw vs injected, truncation, plus tool schema overhead), use

or . See Context.

Time handling

The system prompt includes a dedicated Current Date & Time section when the user timezone is known. To keep the prompt cache-stable, it now only includes the time zone (no dynamic clock or time format).

Use

when the agent needs the current time; the status card includes a timestamp line. The same tool can optionally set a per-session model override ( clears it).

Configure with:

• text

  Copy

  agents.defaults.userTimezone
• text

  Copy

  agents.defaults.timeFormat
  (
  text

  Copy

  auto
  |
  text

  Copy

  12
  |
  text

  Copy

  24
  )

See Date & Time for full behavior details.

Skills

When eligible skills exist, OpenClaw injects a compact available skills list (

) that includes the file path for each skill. The prompt instructs the model to use to load the SKILL.md at the listed location (workspace, managed, or bundled). If no skills are eligible, the Skills section is omitted.

Eligibility includes skill metadata gates, runtime environment/config checks, and the effective agent skill allowlist when

or is configured.

Plugin-bundled skills are eligible only when their owning plugin is enabled. This lets tool plugins expose deeper operating guides without embedding all of that guidance directly in every tool description.

text
Copy
<available_skills>
  <skill>
    <name>...</name>
    <description>...</description>
    <location>...</location>
  </skill>
</available_skills>

This keeps the base prompt small while still enabling targeted skill usage.

The skills list budget is owned by the skills subsystem:

• Global default:
  text

  Copy

  skills.limits.maxSkillsPromptChars
• Per-agent override:
  text

  Copy

  agents.list[].skillsLimits.maxSkillsPromptChars

Generic bounded runtime excerpts use a different surface:

• text

  Copy

  agents.defaults.contextLimits.*
• text

  Copy

  agents.list[].contextLimits.*

That split keeps skills sizing separate from runtime read/injection sizing such as

, live tool results, and post-compaction AGENTS.md refreshes.

Documentation

The system prompt includes a Documentation section. When local docs are available, it points to the local OpenClaw docs directory (

in a Git checkout or the bundled npm package docs). If local docs are unavailable, it falls back to https://docs.openclaw.ai.

The same section also includes the OpenClaw source location. Git checkouts expose the local source root so the agent can inspect code directly. Package installs include the GitHub source URL and tell the agent to review source there whenever the docs are incomplete or stale. The prompt also notes the public docs mirror, community Discord, and ClawHub (https://clawhub.ai) for skills discovery. It tells the model to consult docs first for OpenClaw behavior, commands, configuration, or architecture, and to run

itself when possible (asking the user only when it lacks access). For configuration specifically, it points agents to the tool action for exact field-level docs and constraints, then to and for broader guidance.

Related

• Agent runtime
• Agent workspace
• Context engine