llm-constitution

An operating constitution for AI coding agents, packaged as a Claude Code plugin.

At the top: the Ten Base Laws — invariant-grounded, precedence-ordered, model-agnostic. Below them: dispatch discipline (the commander stays off the grunt work), never-self-verify acceptance, escalation ladders, judgment rubrics, and a perishable bindings layer that absorbs model-name rot. Written to keep binding any LLM — any vendor, any capability tier — for years.

Why a constitution

• Durable by construction. Every law rests on an invariant about agents acting in the world ("a claim is not evidence", "some actions cannot be undone") — facts that do not change as models improve. Perishable specifics (today's model names, budgets, tool mechanics) are quarantined in BINDINGS.md, expected to rot, and paired with re-verification triggers.
• Quality floors that never flex. Work is accepted only by a context that did not produce it; "done" is a checkable state, not a feeling; two failures of one approach end that approach. Verification is never skipped for savings.
• Token economy by design. Progressive disclosure: only skill descriptions sit in a session's context by default. Invoking a skill loads a ~40-line card — the operative core of its doc — not the doc itself; the card names the one section to read when a situation exceeds it. Subagent prompts embed the relevant rules inline, so subagents never load the constitution at all. Small or non-Claude models get only the ≤80-line AGENTS.md floor. Precedence when summaries drift: full doc wins over card, Ten Base Laws win over everything.

Install

/plugin marketplace add a9650615/LLM_constitution
/plugin install llm-constitution

Updating, other harnesses (OpenCode / Codex / Hermes), and the no-plugin bare-checkout path: INSTALL.md.

What you get

Kind	Name	Purpose
Skill	ten-laws	constitutional questions — what is absolute, precedence, how laws are edited
Skill	dispatching	delegation discipline: delegate-vs-DIY, tier selection, escalation ladder
Skill	judgment	rubrics R1–R6: when it's done, when to ask, when to change approach
Command	/verify	fresh-context acceptance check before claiming a task done
Command	/lesson	record a machine/environment pitfall into LESSONS.md (format-checked, deduplicated)
Command	/rebind	re-verify the perishable facts in BINDINGS.md
Command	/handoff	write a session handoff file before context runs out
Agent	scout	cheap read-only locate searches (fixed cheap tier)
Agent	verifier	fresh-context acceptance checker — no file edits, read-only test commands

How it's structured

Precedence, top to bottom — a lower layer never overrides a higher one:

1. docs/05-ten-laws.md — the Ten Base Laws, the supreme layer. Editable only through an explicit protocol (proposal → user approval → fresh-context read-back).
2. CLAUDE.md — operational entry point: routing table and three of the laws restated for fast loading. Universal by design — machine facts live in BINDINGS.md, a per-machine file generated from BINDINGS.template.md via /rebind (not shipped; may not exist yet on a fresh checkout) that names the one machine it binds.
3. docs/10–40 — operational statutes: dispatch, token economy, judgment, templates, maintenance. Defaults, overridable with a stated one-line reason — except the sections marked as law-elaborations, which may only be tightened.
4. BINDINGS.md / LESSONS.md — perishable facts and pitfall log; models may update these themselves (tighten-only, evidence required).

Models may tighten the institution on their own; only the human may loosen it.

File map

File	Content
docs/05-ten-laws.md	the Ten Base Laws — the supreme layer; everything else is subordinate
CLAUDE.md	master operational constitution (auto-loaded every Claude session)
BINDINGS.md	per-machine, generated (gitignored, not shipped): tier→model names, effort mechanics, budget, local model roster, agent wiring. Run /rebind to create it
BINDINGS.template.md	shipped skeleton for BINDINGS.md — copy and fill every placeholder locally
AGENTS.md	the eight-rule floor constitution for non-Claude agents (aider/OpenCode/goose and local models — Qwen/Gemma/DeepSeek, see BINDINGS.md)
docs/10-dispatch.md	dispatch rules: delegate-vs-DIY, tiers, escalation ladder, verify-never-self-verify, spend discipline
docs/15-token-economy.md	context hygiene for the main conversation: reading/output/tool-call discipline, long-context checkpointing, what never to economize
docs/20-judgment.md	judgment rubrics R1–R6 with positive/counter examples
docs/30-templates.md	delegation prompt templates ×5 (search/implement/refactor/research/review)
docs/40-maintenance.md	maintenance protocol: git snapshots, permission levels, lesson format, compaction, language policy, version stamps
INSTALL.md	install details: plugin updates, OpenCode/Codex/Hermes wiring, bare checkout
LESSONS.md	machine-specific pitfall log (appendable; see docs/40 §3)
archive/	frozen founding-era records (中文, append-only): diagnosis + founding letter. Era-bound machine facts stay here — never in docs/
backups/	pre-edit copies for out-of-repo files and git-less contexts (docs/40 §1)
zh/	Chinese mirrors of CLAUDE.md and docs/05, docs/10–40 (headers name the mirrored version)
agents/, commands/, skills/	the plugin surface (see "What you get")
.agents/skills/	per-skill symlinks into skills/ for OpenCode/Codex discovery (agentskills.io standard location)
.claude-plugin/	plugin manifest + self-hosted marketplace listing

English files are canonical (cheaper tokens); zh/ holds Chinese mirrors for the user. The repo's git history is the primary audit trail; the remote (https://github.com/a9650615/LLM_constitution.git) is the disaster-recovery copy — push after every verified commit.

Provenance

Founded 2026-07-03 by a one-off Fable 5 session. Purpose: externalize a strong model's judgment into files, so that every future session — any model tier, Claude or not — runs stronger. Rationale: archive/00-diagnosis.md (中文).

License

MIT — see LICENSE.