A style guide for prose written to agents and by agents.
Chicago tells one writer how to address one reader. TAMOS governs two directions at once: how humans write to agents (prompts, specs, instructions) and how agents write back (reviews, reports, commits, docs). Those are different registers with different failure modes, and most bad "agent writing" comes from conflating them.
A cascade. General rules at the top; specialized overrides below; the more specific rule wins on conflict.
Tier 0 — CORE universal invariants, both directions → core.md
Tier 1 — REGISTERS imperative (to agents) / declarative (from agents)
Tier 2 — ARTIFACTS per-output deltas (PR review, commit, report…)
A Tier-2 module states only its delta over the register it inherits. The PR-review module does not re-explain how to cite or hedge — it says only what is special about a PR review. ("Style" names the discipline as a whole; a single unit is always a "module".)
| Layer | Lives in | Loaded | Style |
|---|---|---|---|
| Enforced | CLAUDE.md / prompt | every turn | terse imperatives, no why |
| Reference | this repo | on demand | prose, rationale, examples |
Rationale never goes in the enforced layer. The why is for the human
maintaining the guide and lives in docs/.
TAMOS ships as a Claude Code plugin (the repo is its own marketplace):
/plugin marketplace add Battle-Creek-LLC/tamos
/plugin install tamos@battle-creek
Then opt the always-on layer into your global ~/.claude/CLAUDE.md by import.
Full steps, CI/CD, and release flow are in docs/user-guide.md.
docs/principles.md— the philosophy (two registers, token economy, dogfooding)docs/structure.md— the tier cascade, the fixed artifact shape, deliverydocs/runtime.md— how modules load (pull/skill) and get enforced (hook)docs/user-guide.md— install, the always-on layer, validation, CI/CD, releasesCHANGELOG.md— releases, and how to upgrade + verifydocs/contributing.md— how to add a new artifact moduleAGENT-STYLE.md— the index/registry of every moduledocs/use-cases.md— what each module governs and the words that engage itagents/— adversarial validators that attack this guide for defects