paidv0.1.0unverified
Claw Starter — one model, one memory, one schedule, one chat surface
The layer above agents is the claw: persistent, scheduled, memory-bearing
(Willison's and Karpathy's term — Karpathy's runs his house). OpenClaw has
250k stars and NVIDIA now ships a governance stack for it; this kit is the
opinionated 4-file starter that works on **Claude Code, Hermes Agent, and
OpenClaw** — same SKILL.md standard everywhere.
Quick start
Preview (free): skills/claw-init/ is included — copy into
.claude/skills/ (or ~/.hermes/skills/), run /claw-init, answer the
interview, get your claw's config + memory scaffold.
Full kit ($39): heartbeat skill, SOUL.md persona template, memory
conventions, per-harness install notes (incl. NemoClaw/OpenShell sandboxing).
What you get
| File | Job | In preview? |
|---|---|---|
skills/claw-init/SKILL.md | interview → writes CLAW.md (duties, boundaries, schedule) + MEMORY.md scaffold | ✅ free |
skills/heartbeat/SKILL.md | the scheduled run: read memory → standing duties → write memory → one-line report | paid |
SOUL.md template | stable persona/identity file (Hermes-native; works as CLAUDE.md section elsewhere) | paid |
MEMORY-CONVENTIONS.md | what a claw remembers vs forgets; append-only logs vs living state | paid |
HARNESSES.md | install + safety notes: Claude Code routines, Hermes tiers, OpenClaw/NemoClaw sandbox | paid |
The four decisions a claw is
Duties (what runs every heartbeat) · boundaries (what it never does
unattended) · memory (what survives between runs, in files not chat) ·
schedule (when it wakes). /claw-init makes you decide all four before
anything runs. Window-surfer's /goal is the natural work engine inside a
heartbeat.
Proof
Unbenchmarked — v0, not yet for sale. First benchmark: 7 days of scheduled
heartbeats, measuring duties completed and memory-file usefulness on resume.
Changelog
- 0.1.0 (2026-07-04) — initial release, unverified.
One email when a pack ships or gets re-benchmarked on a new model. No noise.