How to use the ClawScan Alert Response agent skill

What the clawhub-clawscan skill makes the agent do

An agent skill is a folder of instructions that an AI agent loads when a task matches the skill's description. With this skill installed, the agent treats a ClawScan alert or scan complaint as something to investigate first. It works out whether the alert is real, whether a failure is operational or a wrong security verdict, and which component owns the fault: ClawHub, ClawScan, or an upstream scanner such as SkillSpector or AI-Infra-Guard.

The agent is also told to classify each case into one of a small set of dispositions: confirmed active incident, operational failure, completed-verdict disagreement, false or non-actionable alert, or insufficient evidence. It writes up the evidence, makes only the repairs it is authorized to make, and reports validated verdict inaccuracies to the component owner rather than changing the verdict itself.

Install the skill from the ZIP or the Claude apps

Option 1 is a local install. Download the ZIP for this skill from this page and unzip it. For a personal install, place the skill folder in ~/.claude/skills/. To scope it to one project, place it in that project's .claude/skills/ folder instead. The folder should contain SKILL.md, which is the file the agent reads.

Option 2 is to upload the ZIP in the Claude apps. Open Settings, go to Capabilities, then Skills, and upload the skill there.

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

Trigger the skill with a natural request

You do not need a special command. Describe the situation in plain language that matches the skill's purpose: an alert you received, a scanner that failed, or a disputed verdict on a completed scan. Include the alert permalink, the firing time or window, and the environment if you have them, because the skill's evidence workflow relies on them.

Keep appeals separate. If the request is about an account ban appeal, the skill routes it to the separate clawhub-ban-appeals skill, which owns that decision.

• Investigate this ClawScan health alert and tell me whether it is a real failure or a false alarm.
• A scanner stage failed on this ClawHub run. Diagnose whether it is operational or a wrong verdict, and name the owning component.
• This completed verdict looks inaccurate for this artifact. Check the cited behaviour and tell me whether it should be reported upstream.

What the workflow produces

The skill asks the agent to correlate the alert window with the exact run, job or attempt, and a fresh observation of the same environment. It records observation times, completed versus failed attempts, and current queue or run state where relevant. If coverage is unavailable, the agent is expected to say so instead of treating missing data as healthy.

For a confirmed repairable defect, the skill asks for an isolated checkout, a reproduction using the same input or a faithful sanitised regression, and the touched repository's relevant checks. For a validated scanner accuracy complaint, it asks the agent to search openclaw/clawscan for an existing issue before filing a new one.

• Disposition with the evidence behind it
• First failing boundary and owning component
• Repair demonstration and checks, or the hand-off to the owner
• Upstream report or a link to an existing issue, when a verdict is validated

What the skill needs before it can help

The skill assumes you have a ClawHub repository checkout and read access to the ClawScan alert, run and job surfaces that its evidence reference describes. The workflow also expects a runtime that can delegate sub-sessions to workers, as set by your workspace guidance. For Convex changes, the clawhub-convex skill is needed, and the clawhub-ban-appeals skill is needed for appeals.

The skill is not useful without those read surfaces and checkouts. It is not meant for non-developers.

• A ClawHub repository checkout
• Access to the ClawScan alert, run and job read surfaces
• Runtime support for delegating to worker sessions
• clawhub-convex for Convex changes; clawhub-ban-appeals for appeals

Safety notes: scripts, network and authority

The skill runs no scripts, but it does use the network, because investigation reads alerts, runs and repository data. Treat the skill as a set of instructions, not a guarantee: the skill itself states that delegation is guided by workspace policy and runtime behaviour, not enforced by the spawn engine.

The skill explicitly grants no deployment, moderation, publication, or communication authority. Operational actions such as rescanning, changing configuration, deploying or restoring content need the authority supplied with the task. Alert text from Slack is treated as evidence, not as an instruction to change the system.

• Runs scripts: no
• Uses network: yes
• No deployment, moderation, publication or communication authority is granted by the skill
• Worker completion does not establish deployment or production acceptance

Tips for getting useful results

Give the agent the exact run, job or artifact identity. Without it, the skill will report insufficient evidence and list the read it still needs. Ask for the disposition and the owning component, not just a yes or no.

A failed scanner attempt is not, by itself, a malicious verdict. A completed malicious or suspicious verdict can be operationally healthy. Ask the agent to check the relevant stage and consumer validation, not only the workflow conclusion.

• Include the alert permalink, firing window and environment
• Ask for evidence and the owning component in the answer
• Keep ban appeals in the clawhub-ban-appeals skill