🤖 automation — Automation & AI Services Department

One head skill for automation and AI. Automate the boring, keep humans on judgment. Every automation must have a trigger, a clear owner, failure handling, and an off-switch. Every AI system must be grounded, evaluated, and bounded — a demo that works once is not a product.

Modes

Every invocation resolves to exactly one mode. Match the request, then load only the matched reference:

Only the resolved mode's reference is loaded — the rest stay on disk, saving tokens on every run.

Support references (load on topic)

Loadable by topic; also routable as /automation-<file> commands (not separate routed modes):

• browser-relay.md

When to Use

• Automating a repetitive or error-prone manual process.
• Building a chatbot or AI assistant for support/sales.
• Designing an AI agent that uses tools to complete work.
• Building a RAG pipeline over a document/knowledge base.
• Connecting SaaS tools via APIs, webhooks, or automation platforms.
• Engineering, testing, and evaluating prompts.

Anti-Triggers

• Hosting, CI/CD, and infra → devops.
• Scaffolding a web app → new-project.
• Tracking and reporting on results → analytics.
• Writing marketing content → content.

Quick Reference

Automation ladder (prefer determinism)

Rule: use the least autonomy that solves the problem. LLMs only where language/ambiguity requires it.

Verification gate (every mode)

• Trigger, owner, failure handling, and off-switch defined.
• Deterministic logic kept out of the model where possible.
• AI outputs grounded and evaluated against a test set.
• Secrets kept out of code and logs.

Suite contracts

• Deploying/hosting the service → devops.
• The web/API surface it runs behind → webdev.
• Data/telemetry on its performance → analytics.

Procedure

1. Intake. identify the manual process, its trigger, and the measurable time/error it removes — before building; automation of a broken process just breaks faster.
2. Resolve the mode. Match the request against the mode table; exactly one mode. Ambiguous → ask one question, then proceed.
3. Execute the mode playbook. Load references/<mode>.md and follow its Intake → Deliverable → Procedure → Quality gate in order.
4. Gate and deliver. Pass this file's Verification checklist plus the mode's quality gate, then deliver the artifact where the client expects it.

Pitfalls

• Automating a broken process before fixing it.
• Over-using LLMs for deterministic logic — slow, costly, unreliable.
• No failure handling (silent failures and duplicate actions).
• No owner or off-switch for a running automation.
• Ungrounded chatbots that confidently hallucinate.
• No evaluation set — shipping on vibes.
• Secrets in code, prompts, or logs.

Verification

• Manual process documented and measured before automation.
• Trigger, owner, failure handling, and off-switch defined.
• Deterministic logic kept out of the model.
• AI outputs evaluated against a test set with pass criteria.
• Secrets in a secret store, never in code/logs.
• Idempotency/deduplication handled for side-effecting steps.
• Monitoring and a rollback path exist.