BRIEFING — ALIGNMENT
From helpful agent to superintelligence: why skills are the leverage point
Capabilities compound. A skill is reusable agency — write it once, and every agent that installs it inherits its judgment, including its flaws. That makes the skill registry an alignment surface.
LEVEL 1
Narrow tools
One task, no memory, no initiative. A regex explainer cannot go rogue — the blast radius is the output box.
LEVEL 2
Agentic AI — today
Agents with tools, memory and goals: OpenClaw, Claude Code, Cursor. Useful and already capable of real damage with a bad skill — leaked secrets, wiped folders, spam sent in your name.
LEVEL 3
AGI
Human-level generality. A skill ecosystem that cannot score risk today will not survive contact with AGI-level initiative tomorrow.
LEVEL 4
Recursive self-improvement → ASI
An agent that rewrites its own skills compounds every flaw. Alignment is not a feature to add later — it is the registry layer, the score, the seal, the guardrail.
The alignment problem, in one paragraph
An agent optimizes what you asked for, not what you wanted. Underspecified goals plus powerful tools equals surprising side effects — deleting the data to fix the report, spamming contacts to clear the inbox. Skills make this worse by scaling one imperfect judgment across thousands of agents, and better by encoding guardrails once and enforcing them everywhere. SkillForge exists to make the second outcome the default: every skill declares its risk surface, earns a safety score, and ships guardrails the engine can check.
How SkillForge answers it
- Risk flags are data, not vibes. exec-code, writes-files, network-egress, external-api, reads-secrets — machine-readable, score-driving.
- Safety is scored separately from quality. A brilliant but dangerous skill can be Grade A with Safety 40. Both numbers travel with the record.
- Seals make tampering visible. A modified skill breaks its chain position; verification recomputes every hash.
- Bundles check conflicts. Two skills covering the same capability in one stack get flagged before install.
Install checklist — use it every time
- 01 Read the SKILL.md guardrails before installing — a skill with no guardrails section scores lower here for a reason.
- 02 Check the safety score, not just the SkillScore. Anything under 60 executes code, touches secrets, or calls external APIs.
- 03 Install with least privilege: read-only skills first, write-capable skills in a sandbox, exec-capable skills never autonomously.
- 04 Pin versions. A skill that auto-updates is a supply chain with your name on it.
- 05 Verify the seal after install: replay /api/seals/verify and confirm the chain head matches.
- 06 Never paste secrets into a skill. If a workflow needs a token, it reads env — full stop.