pozzolanic
v0.1.0
Published
An antifragile operating system for Claude Code
Maintainers
Readme
Pozzolanic
An antifragile operating system for Claude Code.
Named for pozzolanic concrete — the Roman volcanic ash formula that gets stronger when exposed to seawater. Where normal concrete degrades, pozzolanic concrete heals its own cracks.
Pozzolanic follows the same principle: every session that finds a gap heals it. New knowledge gets committed to memory. Missing context gets added to slash commands. The system gets stronger with use.
How It Works
Pozzolanic organizes Claude Code around three layers:
- Global config (
~/.claude/CLAUDE.md) — Operating principles and workstream routing. Loads every session. - Slash commands (
~/.claude/commands/*.md) — Deep context for specific workstreams. Load on demand. - Memory (
~/.claude/projects/*/memory/) — Accumulated knowledge. Auto-loaded (MEMORY.md) or agent-read (reference files).
Session starts → CLAUDE.md loads (principles + routing)
→ MEMORY.md loads (critical rules + index)
User types /bq → commands/bq.md loads (BigQuery context)
Agent needs ref → reads memory/revenue-data-sources.md
Agent learns → updates memory → next session starts smarterPrinciples
- System > Feature > Fix — Fix the system once, not symptoms repeatedly.
- Repeatable before Smart — Manual → Scripted → Automated → Intelligent. Never skip steps.
- Observability first — A slow system you can see beats a fast system you can't explain.
- Preserve invariants — Adding a fix must not break existing behavior.
- Fewer concepts — New abstractions are a cost, not a win.
- One source of truth — Each concept has exactly one authoritative source.
- Reversible decisions — Early decisions should be easy to undo.
- Stop when unclear — Don't guess. Ask.
- Done = boring and stable — If it feels clever, it's not done.
- Bias toward action — Do the work, don't describe the work.
The Antifragile Loop
Use the system → Find a gap → Fill it → System gets stronger- Agent missing context? → Update the slash command or memory file.
- New workstream? → Create a new command
.md. - Learned something the hard way? → Add it to memory.
- New repo? → Create a repo-level
CLAUDE.md.
License
MIT
