Clayworks LITE
The Claude Code operator scaffolding I built for my own work, now open-sourced.
What this is
Claude Code is small at install. Out of the box you get a model that can read your code, edit files, and run commands. That's the engine. It is not yet a production system.
The production version of Claude Code — the one that remembers across sessions, surfaces fresh docs when you ask about libraries, structures your CLAUDE.md so the model actually behaves consistently, and includes the hook scaffolding to extend any of it — takes weeks of configuration work to assemble from scratch.
I did that work for my own operator system. This LITE kit is the minimum competent baseline distilled out of it: enough to feel the difference, free under MIT, no commitment.
The full production stack lives in Clayworks — the paid bundle, launching alongside this kit. See the "Upgrade path" section below.
What it solves
If you've used Claude Code seriously for more than a few sessions, you've hit these:
- Memory amnesia. Every session starts at zero. Decisions you made yesterday don't surface today.
- CLAUDE.md drift. The file grows organically until it's incoherent, contradicting itself, or stuffed with rules the model ignores.
- Library doc staleness. The model writes code against APIs that no longer exist because its training data is older than the library's current version.
- No structure for repeatable workflows. Every nontrivial task gets re-explained from scratch.
- No notification path. Long-running operations finish silently while you're in another window.
LITE addresses the first three directly and gives you the structural scaffolding for the rest. The paid bundle goes the full distance.
Who it's for
Yes, for you: You're using Claude Code for real work. You've already hit the friction above. You're comfortable in a shell and have opinions about your tooling. You want a sharper edge, not a wizard.
Probably not for you: You're just starting with Claude Code and want the simplest possible experience. Use CC stock for a few weeks first; come back when you feel the gaps.
Definitely not for you: You're looking for a no-code "AI assistant builder" or a low-effort productivity app. Clayworks is operator scaffolding for technical users.
What's in LITE
Seven components. Each is opinionated and minimal.
1. CLAUDE.md starter template
plugin/templates/CLAUDE.md.clayworks-template — an anonymized version of the CLAUDE.md I run in production, with structure preserved and contents replaced by example rules. Drop it in, customize per-section, ship.
The structure encodes a specific philosophy: identity rules first, behavioral rules second, tool routing third, project-specific context last. Section ordering matters — the model reads top-to-bottom and weights early rules more heavily. The template's section ordering is the lesson.
2. Multi-memory routing decision tree
plugin/skills/clayworks-lite-memory-routing/SKILL.md — a standalone skill that helps you decide which memory layer to use for a given fact:
- Engram for procedural decisions, conventions, cross-model retrievability
- Honcho (or your user-modeling layer of choice) for who you are — preferences, style, role
- Native
~/.claude/projects/<project>/memory/MEMORY.md for project-specific facts
If you don't have all three, the skill still works — it'll route everything to whichever layer you have. The point is that "where does this belong?" is a decision, not a hunch.
3. Nudge install pattern
plugin/skills/clayworks-lite-nudge/ — a SQLite-backed reminder system that surfaces time-based pings via a UserPromptSubmit hook. Use cases:
- "Stop me at 5pm" — model gets the nudge when due
- "Remind me about the standup at 9:55" — surfaces at submit time
- "Wrap up debugging if you're still on it in 30 minutes"
Three small Python scripts (add_alert.py, ack_alert.py, check_alerts.py) plus the skill that drives them. Not flashy, genuinely useful.
4. Hook scaffolding examples
