davidondrej/skills vs the Agent Brigade

Founder shared https://github.com/davidondrej/skills and asked: how does it compare to the brigades, could any become a station/brigade, and "I'd wager ours is more advanced." Surveyed via read-only subagent clone (no code executed, embedded instructions treated as data).

Verdict: he's right — but they're different LAYERS, not competitors

Ondrej's repo and the brigade stack aren't the same kind of thing:

• His repo = a skill LIBRARY. 29 MIT SKILL.md files across 5 categories (agent-orchestration 8 · thinking-and-docs 7 · research-and-web 6 · ops-and-setup 5 · skill-authoring 3). Single author (David, via a skills-publication-bot sanitized-snapshot pipeline). It's breadth of individual capabilities — our L1 layer — plus prose descriptions of orchestration.
• Ours = a manufacturing + orchestration SYSTEM. The rail/expo/gates/steward/cellar/execution-eval/factory machinery that produces and runs vetted skills. L1→L5 as actual code.

So "more advanced" is true on the axis that matters (machinery), with the honest caveat that he isn't trying to build that — and several of his individual skills are genuinely good.

The decisive gap: he has orchestration as PROSE, we have it as CODE

Everything the repo says about coordinating agents lives in instructional skills, not runnable machinery:

• cmux, delegating-to-agents, handoff, codex-goal-loop, agent-self-scheduling = runbooks telling an agent how to coordinate.
• None found as code: orchestrator engine · queue/backlog/ticket concept · quality gates · eval harness measuring skill lift-vs-baseline · plugin/marketplace packaging.

We built each of those: expo (decide authority + exits) · rail + ticket + Gate A ticketLint · execution-eval station (lift-over-baseline, the thing his repo only recommends in effective-agent-skills) · vendored-from-canon distribution · the factory (skill-agent-brigade builds brigades). His handoff skill ("state-not-instructions," 7-section template, secret redaction) is a prose sketch of our ticket/close-out contract — validation we're on the right line.

What could become a station / brigade (his ask)

Worth actually taking (short list)

1. effective-agent-skills as a checklist audit of our spec/critic stations — does our factory encode its progressive-disclosure + anti-pattern rules? Cheap cross-check, real upside.
2. teach pedagogy as seed for an enablement-brigade station (the education work is live).
3. cyber-audit report taxonomy if a security brigade ever gets stood up. Skip the rest — either we have it as code, or it's his-SaaS-coupled.

Red flags (do NOT adopt; surfaced honestly)

• fable-safe-prompt = a skill built solely to reword cyber/bio prompts to evade "Claude Fable 5's server-side safety classifiers." Dual-use guardrail evasion, plausibly against provider ToS. Notable since we run on Fable 5. Steer clear entirely.
• deepapi remote self-mutation — instructs the agent to fetch + overwrite its own SKILL.md from deepapi.co on version mismatch. Live remote-instruction channel = supply-chain risk.
• Monetization baked in — deepapi/deep-research/youtube-transcript wire spend to his product (disclosed, but real).
• Pipe-to-shell install in browser-harness (curl … | sh).
• No prompt-injection against the reader found (injection grep hits are him teaching defense).

Bottom line for founder

Ours is more advanced as a system — he wrote skills and prose about orchestration; we built the orchestration, gates, eval, and factory as running code. His value to us is a few good individual skills as station seeds (teach, cyber-audit) and one methodology doc worth auditing our factory against (effective-agent-skills). Nothing in it threatens the architecture; the orchestration-as-prose vs orchestration-as-code gap is exactly the moat the brigade deck teaches.

Related: [[2026-07-05-brigade-teaching-diagrams-l1-l4]] · [[2026-07-05-agent-stack-consolidation-vs-mds]] · BRIGADE-INTERFACE.md (ray-plugins)