A durable, project-local corpus of declarative coding conventions — “our API responses use this format” — that an agent reads as governing context. In agent-os this is the agent-os/standards/ folder: one concise Markdown file per convention, nested by domain (api/response-format.md, database/migrations.md, root-level naming-conventions.md), plus an index.yml that maps each standard to a one-sentence description so the right ones can be matched and injected without loading them all. Authored to be token-frugal — “Every word costs tokens. Keep them concise” — and to “document only unconventional patterns, not framework defaults.”
Standards are the canonical guide (feedforward control) of topic-harness-engineering: externalized project knowledge the agent consults before acting, standing in for the “we don’t do it that way here” intuition it lacks. Agent OS distinguishes them sharply from Skills — “Standards describe what conventions to follow; Skills describe how to do tasks” — and refines them over time by syncing project-refined standards back into reusable profiles (pattern-knowledge-compounding applied to conventions).
Related but distinct externalized-convention artifacts:
- artifact-constitution — Spec Kit’s gated, immutable, up-front principles; standards are advisory, injected on demand, mined from existing code.
- artifact-adr — a per-decision rationale record; a standard is a standing rule.
- artifact-solution-doc — harvested lessons re-injected as grounding; standards are authored conventions.
Produced by (backlinks)
Agent OS:
- mp-retro — Matt Pocock: proposes
CODING_STANDARDS.mdrules from a session retrospective, but only for judgement calls; mechanical violations become a linter rule, hook, or CI job instead. Standards are read at review, not implementation, because the reviewer has the least context pressure. - agent-os-discover-standards — mines the codebase into new standards files (with “why” rationale).
- agent-os-index-standards — maintains the
index.ymlmatcher over the corpus.
Seeds:
- seeds-onboard — writes a seeds section into
CLAUDE.md/AGENTS.md; a minimal instance, one tool declaring its own conventions into the memory file rather than a curated corpus.
Beads:
- beads-setup — thirteen per-harness recipes that write beads’ workflow rules into whichever file that harness reads, plus its hooks. The store layer’s reason for touching this artifact is blunter than Agent OS’s: a tool an agent does not know about gets ignored in favour of a markdown TODO list.
- beads-rules — the only capability in the wiki that mechanically maintains the guide layer, auditing
.claude/rules/for contradictions by similarity threshold and merging near-duplicates into composites. Compare ce-compound-refresh, which asks an LLM to do the same curation.
pstack:
- pstack-typescript-best-practices — a TypeScript rule table attached automatically by
paths:on any.tsor.tsxfile; the language binding for pstack-principle-type-system-discipline. - pstack-setup-pstack — an always-applied
~/.cursor/rules/pstack-models.mdcmapping each role to a model, rewritten whole on every run so re-runs stay idempotent. - pstack-technical-writing — the layered prose standard; pstack-unslop is the always-on floor beneath it, with stable rule ids that sibling skills cite by number.
Tessl:
- tessl-review — the review rubric is forkable: an organization forks the default to encode its own bar and applies it to every review, “much like static analysis for code”. Standards as a versioned artifact rather than a document, and the closest thing in the wiki to artifact-constitution enforced from outside any framework.
- tessl-workspace — required skills push a mandated capability set into every workspace.
See Also
- agent-os-inject-standards — deploys standards into the working context (reference vs embedded snapshot).
- pattern-context-engineering — the pattern the injected standards serve.
- pattern-knowledge-compounding — the sync-to-profile loop that refines standards across projects.
- topic-harness-engineering — standards as the guide layer; how the layer thins as models improve.