SDLC Agent-Skill Wiki

agent-os

5 min read

Agent OS

Workflow — the primary command per SDLC stage this framework runs, plus the cross-cutting standards loop that is its real center of gravity. It owns the front of the lifecycle and a persistent guide layer; it does not implement, review, validate, or ship.

flowchart TB
    subgraph SL["Standards loop — cross-cutting guide layer (off-stage)"]
      D[discover-standards] --> IX[index-standards] --> IN[inject-standards]
      IN -. refine → sync-to-profile .-> D
    end
    P["plan-product<br/><i>Align</i>"] --> SS["shape-spec<br/><i>Specify</i>"]
    IN -. injected into .-> SS
    SS -. build handed to the agent tool .-> B([agent tool: Plan Mode + todo lists])

Agent OS (Brian Casel / Builder Methods, MIT, ~5.2K★) is “a system for injecting your codebase standards and writing better specs for spec-driven development.” Its tagline is “Agents that build the way you would.” The problem it names: “Every time you prompt an AI coding agent, you’re re-teaching it context that should already be known.” Like Casel’s other project bm-skills — and like nano-spec — it is deliberately not a full-SDLC framework. It owns a durable guide layer (project standards + product docs) and the front of the lifecycle (align + specify), and hands build, review, validate, ship to the coding agent itself.

This is the wiki’s purest instance of the guide (feedforward) layer in topic-harness-engineering: almost the entire framework is machinery for externalizing conventions and feeding the right ones into context at the right moment.

The v3 refocus (this page documents v3.0.0)

v3.0.0 (2026-01-20) is a major refocus — a rare case of a framework deliberately shrinking. Earlier versions shipped create-spec / create-tasks / execute-tasks commands and five sub-agents (context-fetcher, file-creator, git-workflow, test-runner, project-manager) that drove implementation. v3 retired all of that: “Claude Code’s plan mode, extended thinking, and improved models now handle much of the scaffolding that earlier versions of Agent OS provided.” What remains is the part Casel argues the model still does poorly — capturing an organization’s conventions. It is a live illustration of the “as models improve, the authored layer thins toward the irreducible project-specific kernel” thesis: the framework itself deleted its derivable scaffolding and kept only the non-derivable guide layer.

The three-layer context model

All plain Markdown/YAML in a project-local agent-os/ folder:

  • Standards (agent-os/standards/) — declarative conventions, one file per rule, nested by domain (api/response-format.md, database/migrations.md), with an index.yml matcher. “They describe patterns and rules … without prescribing procedures.” The defining artifact. → artifact-standards
  • Product (agent-os/product/) — mission.md + roadmap.md + tech-stack.md; the durable product-direction context. → artifact-product-docs
  • Specs (agent-os/specs/YYYY-MM-DD-HHMM-{slug}/) — a timestamped pack (plan.md, shape.md, standards.md, references.md, visuals/) that persists beyond the conversation. → artifact-spec-md

Capabilities

Agent OS v3 ships five slash-commands and no sub-agents (“you can create your own”). One page each for full coverage:

Standards loop (cross-cutting guide layer — no single canonical stage):

Lifecycle front:

Distinctive contributions

  • Standards as a first-class, token-budgeted artifact. The index.yml matcher loads only relevant standards by one-line description — “Every word costs tokens. Keep them concise.” No other framework here treats externalized conventions as a matched, injected resource with an explicit reference-vs-embed payload choice.
  • Standards vs Skills, named explicitly. “Standards describe what conventions to follow. Skills describe how to do tasks.” — declarative + explicit-invocation vs procedural + auto-detected; Skills may reference standards. This is the cleanest articulation in the wiki of the guide/skill boundary.
  • Bidirectional profile sync as the refinement loop. Refine standards in a project, then sync-to-profile.sh promotes them to a reusable base profile (with inheritance chains). This is pattern-knowledge-compounding applied to standards — the steering loop of topic-harness-engineering, baked in.
  • Deliberate subtraction. With nano-spec and bm-skills, one of three wiki frameworks that consciously refuse to own the whole lifecycle — but the only one whose primary value is the persistent convention layer rather than a spec.

Patterns applied

See Also

Graph View

Backlinks