LLM Instructions

These apply in every session, with no exceptions:

• UNIX line endings — all new files use LF, never CRLF.
• New org docs via compass only — never write .org frontmatter by hand. Always use ./projects/ores.compass/compass.sh add <type>. See How do I create a new doc?.
• Check recipes first — before searching for how to do anything, open Recipes and find the matching topic.
• Every new .org file is scaffolded via compass add — this ensures correct frontmatter (:ID:, #+type:, #+description:, etc.).

Read these first on a new contact:

• Compass — top-level map of doc/; the entry point.
• Orientation — navigation guide by reader intent.
• Glossary — definitions for task, story, sprint, version, function, recipe.
• Agile — current version and sprint at a glance.

• CMake setup — preset names, output layout, build types.
• How do I set up a development environment? — fresh-install bootstrap.
• How do I configure the project? — cmake --preset step.
• How do I build the system? — day-to-day build command.
• How do I run the tests? — cmake --build ... --target rat.

• Code review checklist — naming (snake_case / CamelCase-Qt), include order, comment quality, platform isolation, and all review categories.

• System model — four-layer architecture and component inventory.
• Knowledge — cross-cutting reference: entity patterns, SQL schema, Qt/Wt patterns, CLI/shell patterns, unit-test conventions.

• Document types — frontmatter contract.
• How do I validate component documentation before committing? — validate_docs.sh before every push that touches modeling/.

• ores.sql — component overview; the strict service-table isolation invariant lives under "Service Table Isolation" in this doc.
• Rule: each service user holds DML only on its own ores_<service>_* tables. Do not add cross-service grants without documented justification. Trigger functions that SELECT across services must be SECURITY DEFINER SET search_path = public, pg_temp.

• Unit Test Writer Skill — Catch2 conventions, file structure, naming, table-driven patterns.
• How do I run the tests? — test command.

• How do I create a PR? — title format ([COMPONENT] Description), body structure, and the no-test-plan rule.
• How do I generate a PR summary? — assemble the body before gh pr create.

• Claude Code Settings — the literate source for .claude/settings.json. Lists every permission entry with its security rationale and the concrete command variations it covers. Edit this file (not the JSON) to add or remove permissions; regenerate with How do I deploy the settings?.
• How do I deploy all org-mode artefacts? — the deploy_org (dorg) target rebuilds site, skills, settings, and diagrams in one shot.

• Project memory — catalogue of session-feedback memories; read and write here to persist durable facts across conversations.
• doc-add-memory skill — LLM-side procedure for adding memories.

Non-negotiable: all memories go in doc/llm/memory/ via compass add memory. Never write to the Claude Code harness memory system (~/.claude/projects/*/memory/) — it is disabled via autoMemoryEnabled: false in .claude/settings.json and its files are not version-controlled or reviewed.

• Execution model — how LLMs and agents operate within the cybernetic information architecture during a sprint.
• Recipes — full index of operational how-tos by topic.