Verifiedon 2026.7.1
View this course’s maintenance history
Practical outcome
What this gives you
- Choose a skill, native tool, MCP server, or browser workflow by its smallest sufficient authority
- Define typed inputs, bounded outputs, failure contracts, and an explicit side-effect boundary
- Review MCP and tool access for provenance, permissions, secrets, and retirement
- Prove a browser flow against a disposable fixture before allowing an irreversible step
- Ship a read-only research capability with citations, logs, tests, and a kill switch
Preflight
Before you start
- OpenClaw CLI Essentials and Operator Diagnostics
- Security Hardening and Threat Boundaries
- JSON and configuration literacy
- A disposable browser test target
Syllabus
Work through the map
- Choose the smallest extension surface35 min
- Make inputs, outputs, and failure modes explicit45 min
- Review MCP and tool permissions before connection55 min
- Automate a browser with observable boundaries60 min
- Prove dry runs, denials, and recovery paths50 min
- Operate, retire, and defend the research-skill capstone95 min
- Final assessmentAfter coursework
Course details and supporting evidence
Extend capability without expanding trust by accident
This course teaches a deliberately narrow outcome: build one read-only research capability that accepts bounded input, reads one approved public source, returns cited structured findings, and can be disabled and removed. It does not teach multi-agent orchestration, production messaging, personal browser automation, or permission escalation. Those boundaries are part of the curriculum, not disclaimers added after the implementation.
The course follows the pinned OpenClaw 2026.7.1 stable documentation revision 2d2ddc43d0dcf71f31283d780f9fe9ff4cc04fe4. The primary receipts are the skills guide, MCP guide, browser guide, and tool-policy guidance. The links are direct, immutable revision links so a future editor can re-check claims.
Learning path
- Choose the smallest extension surface. Compare a reusable skill, native tool, external MCP server, and browser workflow by reads, writes, credentials, state, and retirement cost.
- Make inputs, outputs, and failure modes explicit. Design a request and evidence response that can reject malformed input, unavailable evidence, and side effects before invocation.
- Review MCP and tool permissions. Treat a connection as a supply-chain decision with provenance, containment, owner, expiry, and a tested revoke path.
- Automate a browser with observable boundaries. Use a disposable fixture and separate observation, proposal, and fresh confirmation for destructive transitions.
- Prove dry runs, denials, and recovery. Exercise valid, invalid, unavailable, blocked, timeout, and kill-switch outcomes with redacted evidence.
- Operate, retire, and defend the capstone. Assemble the delivery pack and prove that post-removal requests are denied and disposable access is gone.
What you will produce
Your final artifact is a trust-review worksheet and fixture-backed read-only research skill. It includes a decision record, contract, permission inventory, four deterministic fixtures, redacted operation notes, source citations, a kill-switch drill, and retirement proof. A successful capstone is evidence that another operator can reproduce and remove the capability; it is not a credential or a production approval.
Every lesson has independent editorial and technical review, source receipts bound to the current content commit, and reproducible browser/fixture evidence. Never copy a video transcript or treat a supplementary video as canonical instruction. The text, source receipts, and original fixture artifacts are the authoritative learning path.