Claude Code that grows with you. Second Nature mines your own session history for work you keep repeating, drafts it into skills, and keeps your skill library curated over time — with you as the approval gate.
It brings the best idea from self-improving agents like Hermes Agent (the agent writes its own skills from your patterns, a curator prunes them on a cycle) into Claude Code natively — no daemon, no separate memory, no new attack surface. And it adds the thing the self-improving agents are missing: nothing self-installs. Every generated skill passes through a human gate, with evidence attached.
flowchart LR
A[Your sessions<br/>~/.claude/projects] -->|/harvest| B[Proposals<br/>+ evidence]
B -->|/promote<br/>you approve| C[Active skills<br/>~/.claude/skills]
C -->|/curate<br/>weekly| D{score}
D -->|keep / update / merge| C
D -->|retire| E[Archive<br/>never deleted]
C --> A
Three commands:
| Command | What it does |
|---|---|
/harvest |
Scans your recent sessions (default: last 14 days) for repeated multi-step work, re-explained instructions, and fiddly incantations. Drafts each as a proposal with an EVIDENCE.md — which sessions, how many occurrences, what it saves. |
/promote |
Shows pending proposals with their evidence. You approve, edit, or decline. Approved ones install as real skills (user- or project-level). This is the only path from draft to active. |
/curate |
Audits your whole skill library: usage (from transcripts), staleness (do referenced paths still exist?), overlap, quality. Proposes keep / update / merge / retire — applies only what you approve. Retired skills are archived, never deleted. |
As a plugin (recommended):
/plugin marketplace add mmusa/second-nature
/plugin install second-nature
Or just copy the skills — they're plain markdown:
git clone https://github.com/mmusa/second-nature.git
cp -r second-nature/skills/* ~/.claude/skills/Then in any Claude Code session: /harvest.
The loop compounds when it runs on a schedule. Two native options:
- In Claude Code:
/schedulea weekly routine that runs/harvestthen/curate. - Or plain cron/
claude -p:claude -p "/harvest then /curate" --permission-mode acceptEdits
Harvest and curate only ever write to the proposal/archive directories on their own — installs and retirements still wait for your approval, so scheduling them is safe.
- Evidence-traced, human-gated. Self-improving agents fail by confidently installing junk. Every Second Nature proposal carries evidence (sessions, occurrences, estimated savings), and a human approves every promotion and retirement. "The agent says it's useful" is never enough.
- Nothing is deleted. Retired skills and declined proposals move to
~/.claude/second-nature/retired/. You can always dig them back out. - Secrets never travel. Harvest is explicitly forbidden from copying secret values out of transcripts into skills — placeholders only.
- No new surface. It's markdown instructions running inside Claude Code's existing permission model. No daemon, no server, no extra API keys, nothing listening on a port.
Everything lives in ~/.claude/second-nature/:
proposals/<slug>/SKILL.md # drafted skill
proposals/<slug>/EVIDENCE.md # why it was proposed
retired/ # archived skills & proposals
LOG.md # append-only history
How is this different from Hermes Agent / OpenClaw? Those are standalone agent runtimes — a daemon with its own memory, channels, and (real, documented) security surface. Second Nature is not an agent; it's a growth loop for the agent you already run. If you live in Claude Code, you get the compounding without a second brain to secure.
Does it phone home? No. Everything is local files. There's no telemetry of any kind.
What if harvest proposes something wrong? Then you decline it at /promote and it never activates. That's the point of the gate.
Inspired by the Autonomous Curator in Hermes Agent by Nous Research, the skills ecosystem around OpenClaw, and the "evidence-traced done" verification discipline. Built with Claude Code, for Claude Code.