openclaw-ops

Guide for creating and customizing AGENTS.md — the operating rules, procedures, and behavioral guardrails for an OpenClaw agent. Use this skill whenever the user is working on agent behavior, operating procedures, session startup sequences, memory management policies, group chat rules, safety red lines, tool usage priorities, or standing order references. Even if they just say "the agent should do X differently" or "add a rule for Y", this skill applies because operational rules belong in AGENTS.md. Also use when deciding what belongs in AGENTS.md versus SOUL.md — this skill clarifies the boundary.

Guide for BOOTSTRAP.md (first-run setup ritual) and BOOT.md (gateway restart checklist) in OpenClaw. Use this skill whenever the user is setting up a new agent for the first time, configuring onboarding flows, creating first-run discovery processes, or defining what should happen on gateway restart. Even questions like "what should happen when the agent starts for the first time", "how do I initialize a fresh agent", or "what runs on reboot" need this skill.

Use when an OpenClaw instance config is about to be read, changed, restored, split or explained — a model chain, channel, tool, plugin, skill, session or memory setting, a secret reference, an include, a gateway that refuses to start after an edit, a config the process appears to ignore, a change that looks applied and has no effect, an edit that removes lines, a stray backup or rejected sidecar file next to the config, or the question of what needs a restart.

Use when anything about OpenClaw is about to be stated or recommended that could have changed — a config key, a CLI flag or subcommand spelling, an auth method, a release channel, a current version, a model name, whether a feature exists or is deprecated — and whenever a claim needs a citation, two sources disagree, an instance's own documentation looks older than the project's, or the situation is offline and it must be decided what may still be said without a live source.

Use when a whole OpenClaw fleet job is in view rather than a single answer — first contact with a host whose instances are unknown, a fleet where every instance lost its provider login, an upgrade whose rollback story is unclear, a new instance that has to be stood up and proven isolated — and whenever the question is what the entire sequence looks like, which command owns which step, what a run of this plugin produces end to end, or where one step's output becomes the next step's input.

Use when an OpenClaw instance is misbehaving or suspected of it — health green but nothing is happening, a container restarting or unhealthy, an empty or ignored config, models returning unauthorized or reporting logged out, tokens that are present but empty, schedules failing or firing twice or stalling, memory search dead or paused or a state database growing without limit, keys missing inside the container, shared skills that look installed and do nothing, a version that does not match its siblings, or a port published where it should not be. Also use when a report, audit or repair needs a stable finding id.

Use when work touches an OpenClaw gateway instance or several of them on one host — inventory, status, health, logs, provider auth, config, secrets, memory, shared skills, upgrades, cloning — or when an instance is named or selected, or when the alternative would be a hand-written docker exec, a guessed path, or a remembered version or model name.

Guide for OpenClaw heartbeats — the periodic background turn that surfaces anything needing attention. Current releases keep heartbeat instructions in the heartbeat monitor's scratch (a database row, managed with `openclaw automations scratch`) and the cadence in `agents.defaults.heartbeat`; the HEARTBEAT.md workspace file is retired and never read at runtime, but older workspaces still carry one. Use this skill whenever the user needs background monitoring, periodic checks, standing orders integration, quiet hours configuration, a leftover HEARTBEAT.md, or wants to understand heartbeat mechanics and token costs. Even questions like "how do I make the agent check something periodically", "set up monitoring", or "reduce heartbeat token costs" need this skill.

Guide for creating IDENTITY.md — the agent's name, creature type, vibe, emoji, and avatar in OpenClaw. Use this skill whenever the user is naming their agent, choosing an emoji or avatar, setting up the agent's visual identity, or asking how agent identity works. Even simple questions like "what should I name my agent", "how do I set the bot emoji", or "change the agent's avatar" need this skill.

Use when a new OpenClaw instance is to be created from an existing one — cloning the reference, standing up a canary or a throwaway test instance, adding an instance for a new workload or tenant, picking a free gateway port for one, deciding what a new instance may share with its source and what it must not — and also when an instance created earlier behaves like its source, answers on the wrong port, has no credentials, or is suspected of not being isolated from the instance it was copied from.

Use when an OpenClaw instance or a fleet of them is being upgraded, or when the question is which version to move to — version drift between instances, what the current stable is, a release channel or a registry dist-tag, an image tag or digest pin, a soak or hold-back window, a gateway that will not start after an update, schedules that started firing several times per tick, a session that expired some time after an update, or what a given release added.

Use when OpenClaw memory or its embeddings are involved — embedding calls failing authorization or reporting an invalid token, vector search paused or returning nothing useful, an index-identity warning, a last-index timestamp far in the past, search still poor after the provider was fixed, choosing or changing an embedding provider, model or key, a reindex, or a state database that keeps growing and needs retention or compaction. This skill owns the subscription-session-versus-embedding-key distinction; chat-side credential and OAuth problems are provider-auth.

Guide for OpenClaw's memory system — MEMORY.md curation, daily logs, memory flush, and vector search. Use this skill whenever the user asks about how the agent remembers things between sessions, wants to configure memory management, needs to understand daily logs versus long-term memory, or is setting up memory curation processes. Also applies to questions about vector search, memory security in groups, or "how does the agent remember".

Use when model-provider credentials are in question — unauthorized responses, an instance reporting itself logged out, tokens present but empty, an expiry approaching or passed, several instances losing the same account at once, a login that has to be performed, a choice between an API key, provider OAuth and a local CLI backend, billing that moved to metered tokens without a config change, a provider CLI reporting not-logged-in inside its container, or credential directories being shared, split or copied between instances. Embedding and vector-search failures are memory-ops, not this skill.

Use when secrets reach an OpenClaw instance through an injection wrapper and something about that is in question — a feature silently off while the config looks right, an instance receiving far fewer keys than its siblings, a plaintext env file inside the state tree, a token or client id sitting in a backup or identity copy, a key that has to be added, rotated, or proven delivered, a machine identity or its project binding, a secret reference in the config, or a restart whose safety depends on every referenced key still resolving.

Use when the security posture of an OpenClaw instance or of the whole fleet is in question — a gateway port that may be reachable from outside the host, firewall rules that read correctly but may not apply to published ports, an operator bearer token that may be shared between instances or sitting in a plaintext file, permissions on state, credential or identity trees, a metrics or admin endpoint answering without authentication, a secret found in a backup or an identity copy, a question about who can reach an instance and what that access grants, before exposing an instance to a new network or new people, and after any suspected compromise.

Techniques for analyzing OpenClaw session transcripts (listed through the CLI, exported as a trajectory bundle, then queried with jq) to understand user behavior, tool effectiveness, error patterns, token costs, and active hours. Use this skill whenever the user wants to analyze how their agent is performing, what users are asking, which tools work best, what errors occur, or how to improve the workspace based on real usage data. Also applies to questions like "what do my users ask about", "is my agent working well", or "show me session stats".

Use when skills or plugins are shared across OpenClaw instances on one host — shared trees mounted but empty, the same skill copied into every instance, a shared copy edited with no change in behaviour, an installed asset that never appears in the registered list, a plugin change that did nothing, a load candidate refused over ownership, a suspect install lock, or any request to deduplicate, promote, register, verify or roll out shared skills and plugins.

Guide for creating and refining SOUL.md — the agent's personality, values, tone, and boundaries. Use this skill whenever the user is working on who the agent should be, its communication style, persona traits, behavioral boundaries, or domain-specific character. Even requests like "make the agent more friendly", "add a boundary", "change the tone", or "the agent sounds too corporate" need this skill because personality and tone live in SOUL.md, not AGENTS.md.

Design patterns for standing orders — autonomous programs that give agents permanent operating authority for defined tasks. Use this skill whenever the user needs recurring tasks, scheduled agent operations, autonomous execution rules, or the execute-verify-report pattern. Also applies when discussing cron jobs with OpenClaw, periodic reports, automated monitoring, content scheduling, or any task the agent should do on its own without prompting each time.

Guide for the local tool notes of an OpenClaw agent — environment-specific tool notes, search priorities, device names, and local infrastructure details. On current releases these live in the `## Tools` section of AGENTS.md (the standalone TOOLS.md file is retired); older workspaces still carry TOOLS.md. Use this skill whenever the user is configuring which tools their agent should prefer, adding environment notes, documenting search tool priority orders, or setting up device/API references. Even questions like "which search tool should my agent use first", "how do I tell the agent about my devices", or "add tool notes" need this skill.

Guide for building USER.md — user profiles, preferences, roles, and communication styles. Use this skill whenever the user is setting up who will interact with the agent, mapping Telegram/Discord user IDs to names, configuring multi-user access, or defining communication preferences. Even questions like "how does the agent know who I am", "add a team member", or "set up user permissions" need this skill.

OpenClaw workspace architecture and file system overview — directory structure, file loading mechanics, character limits, scan categories (A-J), and subfolder organization patterns. Use this skill whenever the user asks about how the workspace is structured, what files get loaded automatically, character limits, what goes where, or how to organize their workspace. This is the foundational skill for workspace content — use it first when the user is new to OpenClaw or asks broad questions about workspace organization, even before more specific skills like agents-md or soul-md.