agent-scripts

Peter Steinberger‘s personal hub for the agent instructions, skills, and helper scripts he shares across all his repositories — one canonical copy that downstream projects point at rather than duplicate. It’s the individual-developer end of the skill-infrastructure spectrum: not a marketplace (pm-skills, ClawHub) or a distributed methodology (superpowers), but one engineer keeping his own tooling DRY.

What’s in it

• skills/<name>/SKILL.md — reusable workflow skills with YAML front matter, in the same agentskills.io shape the rest of the spoke tracks. Some are symlinked in from other repos (agent-skills, discrawl), with naming conflicts resolved hierarchically.
• AGENTS.MD — the shared “hard rules” for agents, linked from downstream repos instead of copied. A concrete instance of the AGENTS.md convention used as a single source of truth across a developer’s whole footprint.
• scripts/ helpers — dependency-light portables: committer (stages files, validates commits), sync-skills (builds per-machine skill mirrors, symlinking into both Codex and Claude Code), validate-skills (checks SKILL.md formatting), browser-tools.ts (Chrome DevTools automation).
• hooks/ validation guardrails, docs/, and release/ + .github/workflows/ deploy automation.
• Mostly Shell / JavaScript / Python.

Why it matters here

• Cross-harness by construction. sync-skills mirrors one skill set into Codex and Claude Code, the same portability bet as superpowers and pm-skills — but scoped to one person’s machines, so it’s the smallest useful unit of the portable-skill thesis.
• The maintenance ops, human-run. sync-skills / validate-skills are the discovery / dedup / validation operations autoharness performs automatically by distilling sessions — here they’re authored and validated by hand. Same library-maintenance problem, opposite side of the manual→autonomous authoring spectrum (hermes-learn-skill-capture).
• Same author as openclaw. The person who built the gateway-first OpenClaw harness also keeps a private cross-harness skill toolkit — evidence the skill/AGENTS.md conventions are load-bearing enough that a harness author standardizes his own work on them, not just ships them to others.

Tier

T1 — first-party repo, canonical for its author and the primary source on itself. No benchmarks or external claims to weigh; it’s personal infrastructure, not a product pitch. freshness: volatile (actively churned).

Related

peter-steinberger · openclaw · superpowers · agents-md · agentskills-spec · autoharness · pm-skills · synthesis