LLM Configuration and Artefacts

Component overview
How they fit together
See also

ORE Studio's LLM support is layered: each component builds on the one below it, from single operations up to role-shaped session priming. This page is the entry point — it explains what each component is, when to use it, and where to find it.

Component overview

Component	Grain	Shape	Example
LLM-first	Philosophy	How LLMs drive development	LLMs execute; humans direct; docs serve both audiences
Claude Code	Primary tool	Anthropic CLI for Claude	Reads `CLAUDE.md`; edits code, opens PRs, runs commands
Instructions	Entry point	Session-level rules and links	Mandatory first read for every LLM session
Runbooks	Multi-step procedure	Composition of recipes	"Start work on a new story" (7 steps, story → PR)
Recipes	Single operation	How-to with executable shell	"How do I create a PR?"
Skills	Unit of work	Playbook an LLM loads	"System Architect" — read model, understand layers
Functions	Role	Priming context for a session	"S2 Orchestrator" — coordinates across S1 agents
Settings	Permissions	Sandbox and tool allow-list	`.claude/settings.json` — what the LLM can execute
Memory	Durable fact	One fact per file	"Set SSH_AUTH_SOCK before git push"

How they fit together

Instructions — the session entry point

LLM instructions (doc/llm/llm_instructions.org) is the mandatory first read for every session. It lists non-negotiable rules (UNIX line endings, codegen-only doc creation, check recipes first) and links to build, architecture, code style, database isolation, testing, and git/PR convention documents. Every LLM session loads this before acting.

Runbooks — composed procedures

A runbook chains recipes and skills into a repeatable multi-step procedure with preconditions and postconditions. Where a recipe answers "how do I do X?", a runbook answers "begin work on X" — it captures the sequence so the LLM doesn't have to infer it each time. Runbooks live under doc/llm/runbooks/ and are S2 (Orchestrator)-owned.

Full catalogue: Runbooks catalogue — every runbook with a one-line description. List with compass list --type runbook; scaffold with compass add runbook.

Recipes — single operations

Recipes (doc/recipes/) are short, tested how-tos grouped by topic. Each answers one NLP question with executable shell commands in #+begin_src sh blocks. Recipes are the fundamental building blocks — runbooks compose them, skills index them.

Inventory: Recipes index.

Skills — unit-of-work playbooks

Skills (doc/llm/skills/) are task-shaped playbooks an LLM session loads to perform a specific unit of work end-to-end. Each skill describes when to invoke it and how to execute it. Unlike recipes (single operations), skills cover a complete unit like "review a component" or "create a domain type."

Inventory: Claude Code Skills.

Functions — role-shaped priming

Functions (doc/functions/) are role-shaped documents, one per cybernetic level (S1 Agent, S2 Orchestrator, S3 Sprint Planner, S3* Auditor, S4 Version Planner, S5 Identity Steward). A function doc is the priming context an LLM loads when taking on that role for a session. Functions are coarser than skills — they define boundaries and authority, not task steps.

Inventory: Functions index.

Settings — tool access

Claude Code Settings (doc/llm/claude_code_settings.org) is the literate source for .claude/settings.json. It defines the permission allow-list and sandbox overrides that control what an LLM can execute. Never edit the JSON directly; regenerate with deploy_settings.

Memory — durable facts across sessions

Project memory (doc/llm/memory/) stores persistent session feedback and project facts. Four subtypes: feedback (user corrections), user, project, reference. One memory per file; each carries *Why:* and *How to apply:* so future sessions can judge edge cases. Loaded at session start; appended to over time.