How we run an OpenClaw bot safely: skills, long runs, and GitHub access

Running an agent against a real repository is where it gets useful — and where small security mistakes turn into big problems.

This post is the playbook we use to run an OpenClaw bot reliably:

• skills that route correctly
• long-running tasks that don’t fall apart
• secure GitHub access without leaking credentials

If you’re building AI workflows for growth + operations, start here: https://zeiko.io.

1) Skills are routing logic (not marketing)

A skill’s description is the model’s decision boundary. Write it like a filter:

• Use when: concrete triggers (inputs, file types, tools used)
• Do NOT use when: negative examples (common misroutes)
• Output: what the skill should produce

Example:

• “Use when the user pasted a WhatsApp chat and wants a reply. Do NOT use for reminders or gateway troubleshooting.”

Why it matters: when you add multiple skills, routing gets harder unless each skill has a sharp edge.

2) Add negative examples and edge cases

This is the fastest way to improve accuracy.

Add 3–6 bullets like:

• “Don’t call this skill when the user is asking for X. Use <other-skill> instead.”
• “If the time is ambiguous (9am vs 9pm), ask one question instead of guessing.”

3) Put templates inside skills

Templates inside skills are effectively free until invoked.

Examples:

• WhatsApp reply templates (reschedule, confirmation, apology, boundary)
• Research brief template (summary, key points, risks, sources)
• Diagnostic checklist (what to capture before restarting services)

4) Determinism lever: explicitly call the skill

When you want reliability over creativity, don’t rely on routing:

“Use the <skill-name> skill.”

This turns “maybe” into a contract.

5) Long-run durability: artifacts + commits

Agents work best when there’s a clean handoff boundary.

Artifacts

Write real outputs to disk:

• plans, briefs, reports
• exports (CSV/PDF)

Commits

Keep changes reviewable:

• small commits
• clear messages
• prefer PRs for user-facing changes

If you’re building productized automation and want a repeatable workflow, Zeiko can help: https://zeiko.io.

6) Security posture: skills + networking is high risk