Load CO (Cognitive Orchestration) methodology reference. Use when discussing the domain-agnostic base methodology for human-AI collaboration, the eight first principles, the five-layer architecture, or the relationship between CO and domain applications like COC.
This skill provides the reference for CO — the domain-agnostic base methodology for structuring human-AI collaboration in any domain where AI agents operate under human oversight.
This skill is self-contained — all essential CO knowledge is distilled below from the CO Core Thesis by Dr. Jack Hong and the CO specification. If Foundation source docs exist in this repo, read them for additional depth.
CO (Cognitive Orchestration) is a methodology for structuring institutional knowledge, guardrails, and processes so that AI agents produce trustworthy output in any domain. It is the base methodology from which domain-specific applications are derived.
CO sits in the trinity alongside CARE and EATP:
Layer 5: LEARNING — Observe, capture, evolve knowledge across sessions
Layer 4: INSTRUCTIONS — Structured workflows with approval gates
Layer 3: GUARDRAILS — Deterministic enforcement outside AI context
Layer 2: CONTEXT — Organization's institutional knowledge, machine-readable
Layer 1: INTENT — Route to domain-specialized agents
Each layer encodes a different aspect of human judgment:
Layer 4 (Instructions) is implemented through a six-phase workflow:
| Phase | Template Command | Target | Purpose |
|---|---|---|---|
| 01 | /analyze | 01-research/ | Research and understand the problem space |
| 02 | /plan | 02-planning/ | Structure the work; human approves |
| 03 | /execute | 03-work/ | Do the work one task at a time |
| 04 | /review | 04-review/ → 05-output/ | Adversarial critique; produces finalized output |
| 05 | /learn | .claude/ | Extract knowledge; upgrade CO artifacts |
| 06 | /deliver | 05-output/ → recipient | Package and hand off |
Phase 04 (Review) produces the finalized work. Review iterates until quality passes, then saves to 05-output/.
Phase 05 (Learn) is unique — its output goes OUTSIDE the workspace, back into the CO system (.claude/ artifacts). This is Principle 7 (Knowledge Compounds) made concrete. Every run makes the system stronger. Proposals require human approval before modifying .claude/.
Phase 06 (Deliver) packages and ships. Domain applications rename this: /publish (COR), /release or /deploy (COC), /submit (student COs).
The workspace has 5 directories (01-research through 05-output) because Phase 05 has no workspace directory.
Domains rename commands to fit their vocabulary but preserve the 6-phase structure. CO does not prescribe the exact number of commands per phase — domains may split phases (COR splits Phase 01 into /teach and /literature) or add specialist commands within phases.
| Application | Short Name | Status |
|---|---|---|
| CO for Codegen | COC | Production |
| CO for Research | COR | Production |
| CO for Education | COE | Analysis |
| CO for Governance | COG | Production |
| CO for Compliance | COComp | Sketch |
| CO for Learners | COL | Development |
| COL for Finance | COL-F | Production |
COC is the first and most mature domain application. It proves CO's principles work in practice with 29 agents, 25 skills, 8 rules, 8 hooks, and 12 commands.
COL (CO for Learners) is the subject-agnostic student CO. Subject layers (COL-F for Finance, future COL-H, COL-B, COL-L) inherit COL and add subject-specific commands, agents, skills, and rules.
CO inherits CARE's Human-on-the-Loop philosophy. The mapping:
| CARE Concept | CO Manifestation |
|---|---|
| Trust Plane (humans define boundaries) | Layer 2 (Context) + Layer 3 (Guardrails) |
| Execution Plane (AI at machine speed) | Layer 1 (Intent agents) |
| Constraint Envelopes | Layer 3 enforcement mechanisms |
| Human-on-the-Loop | The Human-on-the-Loop practitioner role |
| Evolutionary Trust | Layer 5 (Learning pipeline) |
CO's guardrails connect to EATP's trust infrastructure:
| CO Layer | EATP Connection |
|---|---|
| Layer 3 (Guardrails) | Constraint Envelopes — formal boundaries enforced deterministically |
| Layer 4 (Instructions) | Trust Postures — approval gates map to verification gradient |
| Layer 5 (Learning) | Audit Anchors — learning observations become audit records |
CO = Cognitive Orchestration
8 Principles: Institutional Knowledge, Brilliant New Hire, Three Failures,
Human-on-the-Loop, Deterministic Enforcement, Bainbridge's Irony,
Knowledge Compounds, Authentic Voice
5 Layers: Intent → Context → Guardrails → Instructions → Learning
6 Phases: Analyze → Plan → Execute → Review → Learn → Deliver
3 Failure Modes: Amnesia, Convention Drift, Safety Blindness
1 Insight: Institutional knowledge > Model capability
If Foundation source docs exist in this repo, read the CO Core Thesis and CO specification for additional depth. For comprehensive analysis, invoke the co-expert agent.