“Three Primitives for Claude Code as Life OS” — Unknown author

Why this is in the vault

Author argues Claude Code’s breakthrough isn’t coding — it’s that the filesystem becomes long-term memory. The 3-primitive architecture (Memory / Maintenance / Agentic primitives) maps onto RDCO with a real gap: we lack a /wrap-equivalent reconciling session insights back into CLAUDE.md. Direct relevance: agent-deployer thesis evidence, RDCO operating-architecture audit, candidate skill addition.

The core argument

Two foundational facts: Claude Code can take actions (write code, build plans, convert notes to actions) AND has long-term memory via filesystem read/write. Three primitives compose into life-OS:

Memory structure — hierarchical CLAUDE.md (global / vault / area / project) auto-loaded by directory walk; workbench.md as session log per project; daily memory log as cross-project index.
Maintenance loop — /wrap skill at session end reconciles workbench → CLAUDE.md Current Status, prevents drift.
Agentic primitives — skills, slash commands, n8n workflows. Author treats them as SOPs for agents. “Anything you would turn into a SOP should exist as a skill file.”

Key claims

Files-as-memory is what differentiates Claude Code from chat AI (context window dies at session end vs. filesystem persists).
CLAUDE.md hierarchy auto-loaded by harness via directory walk. The author uses 4 layers (global / vault / area / project).
workbench.md is append-only, dated newest-first, with decision reasoning attached.
Daily memory log accumulates essentials across all projects per day — becomes the operating index (“what I actually worked on, decided, shipped”).
/wrap skill reconciles workbench (full history) → CLAUDE.md (current snapshot). Skip a few wraps and context goes stale.
Skills heuristic: add a new one any time you catch yourself doing the same thing twice.
The composition of the 3 primitives is where the interesting work happens (consulting workflow, writing pipeline, weekly planning).

Mapping against Ray Data Co

The 3-primitive lens applied to RDCO honestly:

Primitive	RDCO state	Gap
Memory	✓ Strong. Global ~/CLAUDE.md + vault structure + ~/.claude/state/working-context.md (workbench equivalent) + qmd-indexed vault + DuckDB knowledge graph.	No explicit “daily memory log” — closest analog is the conversation transcript + cron-fired skill outputs scattered across state/, no consolidated daily index.
Maintenance	✗ Real gap. /improve runs autonomously Mondays for SKILL.md drift, but NOTHING reconciles session insights back into CLAUDE.md itself. CLAUDE.md is human-curated; if RDCO operating principles drift mid-session, the next session opens with stale context.	A /wrap-equivalent is a clear lift candidate — see Open follow-ups.
Agentic primitives	✓✓ Ahead of the author. 50+ skills, 13 cron loops, sub-agent fan-out, multiple slash commands, plus we lift external skills (compound engineering, iOS dev pack).	None — this is RDCO’s strongest dimension.

The “claw” reframe (founder, 2026-05-01): the author’s setup is single-user terminal-based; RDCO’s setup is bidirectional channel-based (iMessage + Discord). The org chart IS the channel — founder drops a thought, COO-agent (me) grabs the relevant skill/context/sub-agent and executes. Structurally different machine, not a faster IDE. Worth claiming as a Sanity Check angle in its own right (“Stop using Claude Code as a faster keyboard. Use it as a claw.”).

Connection to Miura-Ko levels framework (2026-05-01-ann-miura-ko-six-levels-ai-pilled-organizations):

Memory primitive maps to Miura-Ko’s “What can AI see?” axis — RDCO is L3 strong here.
Maintenance primitive maps to “How has the org changed?” + “system improves itself” axes — the /wrap gap is exactly the L4 “system that learns without being prompted” weakness this vault note’s parent identified.
Agentic primitives map to “Who can extend the system?” — RDCO is L4 here (skills as markdown = anyone editing markdown extends the system).

Connection to Compound Engineering / agent-deployer cluster:

The /wrap-skill pattern is conceptually identical to Compound Engineering’s compaction discipline — preventing context rot at session boundaries. CE plugin v3.4 and this piece converge on the same insight from different angles.
Author’s “skills as SOPs for agents” is the same framing Garry Tan uses in “fat skills, thin harness” (2026-04-11-garry-tan-thin-harness-fat-skills).

Open follow-ups

Build a /wrap equivalent for RDCO — at session end (or daily cron), diff working-context.md against CLAUDE.md and propose updates to the Current Status / hard rules section if anything drifted. Could queue as Notion task pending founder greenlight.
Daily memory log — consider a daily-cron skill that consolidates the day’s vault writes + Notion task changes + skill fires into a single ~/rdco-vault/05-meetings/2026-MM-DD-day-log.md (or similar location). Lower priority than /wrap but lands the missing piece of the author’s 3-layer memory stack.
Sanity Check angle: “claw vs coding partner” — propose to founder as research-brief candidate. Sharp differentiation from common Claude-Code-for-coders content.