How to use the Host Observer for Grok Bot agent skill

What the host-observer skill makes the agent do

An agent skill is a folder containing a SKILL.md file with instructions that an AI model reads when a task matches the skill's description. The host-observer skill applies only inside the claude-mem observer pipeline. It tells the model how to answer observer jobs for Grok Bot.

The model must reply in one of two fixed formats. For idle, init, or no-useful-tool-result events, it returns an empty skip marker. For a finished unit of work, it returns one structured observation with a real title, 4 to 10 facts that cite file paths, and a narrative. Free-form prose is discarded.

• Skip marker for noise: <skip_summary reason="noise" />
• Observation types listed in the source: bugfix, feature, refactor, change, discovery, decision
• Titles must never name a tool
• The model should not call the memory save endpoint except as a last resort

Install the host-observer skill

Download the ZIP for host-observer from this page. Then choose one of the two install locations below.

For a personal setup, unzip the download into ~/.claude/skills/ so the host-observer folder with its SKILL.md sits inside that directory. For a single project, unzip it into the project's .claude/skills/ folder instead.

If you use the Claude apps rather than a folder on disk, upload the skill under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Per project: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills

Confirm you have the required setup

The skill does nothing useful on its own. The source lists three requirements: the claude-mem observer pipeline, Grok Bot or an OpenRouter-compatible local shim, and access to the outbox/inbox file layout that the bot uses.

If you do not run claude-mem's observer loop, you can skip this skill. It is not a general improvement for everyday agent tasks.

• claude-mem observer pipeline
• Grok Bot or an OpenRouter-compatible local shim
• Access to the outbox/inbox file layout used by the bot

How to trigger the skill with natural requests

The skill's description says to use it when fulfilling claude-mem observer jobs on Grok Bot. In practice, the model loads it when it receives an observer job in that context. You do not need a special command; the job itself is the trigger.

Keep requests within the observer context. Phrases that match the job, such as a claude-mem observer job on Grok Bot with an inbox item to answer, are the ones the description targets. Generic requests like help me write code will not match this skill's purpose.

• Good trigger: a claude-mem observer job on Grok Bot asking for an observation for inbox item {id}
• Poor trigger: a general coding question with no observer job

Examples of the output format

For a job where nothing useful happened, such as an init event with no tool results, the model should write only the skip marker to outbox/{id}.txt. For a finished change, it writes a single observation XML element. The source does not give a full observation example, so check your pipeline's expected schema before relying on exact tags.

Do not mix a skip marker and an observation in the same reply. Prose replies are dropped by the pipeline.

• Noise: write only the skip marker
• Finished work: write exactly one observation
• Never send both in one reply

Safety notes: scripts and network

The assessment for this skill states that it runs no scripts and uses no network. It is a text file of instructions for a model. It does not run code or fetch data itself.

The skill does tell the model not to POST to /api/memory/save except as a last resort. Review your pipeline before enabling it, since the outbox and inbox paths only make sense in that specific setup.

• Runs scripts: no
• Uses network: no
• Only SKILL.md is included

Tips for using the skill well

Read the SKILL.md file before installing so you know exactly what instructions the model will follow. Keep the outbox and inbox layout consistent with what the skill assumes. If you are a maintainer, the main value is consistent, machine-parseable output across many jobs.

• Read SKILL.md before installing
• Match your outbox/inbox paths to the skill's assumptions
• Use it only for claude-mem observer jobs