A set of skills for AI agents, built around the Conscious Agentic System (CAS).
These two skills work together: cas-agent-setup wires the file layout, and cas-operational-loop defines how to run each session.
Set up CAS persistence for a new or existing agent. Creates the required files and directories (STATE.json, append-only telemetry, raw output store, optional archive layout) and wires schemas and specs. Ships bundled helpers: STATE template, all JSON schemas, consolidation spec, and acceptance test definitions.
Use when bootstrapping an agent or aligning an existing agent with the CAS file contract.
Run the 8-phase CAS loop for every consequential action. Phases: State Ingestion → World-Model Update → Prediction → Action → Evaluation → Learning → Commit. Covers prediction/observation/learning record formats, raw output capture, pending citations, run_nonce scoping, and error classification.
Use when operating as a CAS agent — this skill defines the how, while cas-agent-setup defines the what files.
Most agents fail in one of three ways:
- They act without grounding in current context.
- They forget why they made a decision.
- They don't learn from mistakes, so the same failure repeats.
CAS solves this by making reasoning and learning operational, not optional. It turns "be thoughtful" into a repeatable runtime contract.
CAS enforces an 8-phase loop on consequential work:
- Boot: load identity, constraints, and prior state.
- State Ingestion: capture what's new in the user/task/environment.
- World-Model Update: refresh assumptions about user intent and context.
- Prediction: state the smallest useful next move before acting.
- Action: execute with tools.
- Evaluation: compare outcome vs prediction.
- Learning: extract durable lessons.
- Commit: persist state + telemetry for next session continuity.
Concretely, this is backed by durable files (STATE.json, predictions.jsonl, observations.jsonl, learning.jsonl) and run scoping via run_nonce so each session stays auditable.
Look for all of the following:
- Traceability: every meaningful action has a prior prediction and a post-action evaluation.
- Continuity: session-to-session behavior improves without re-explaining context.
- Error containment: failures are classified and followed by explicit learning entries.
- Grounded execution: actions are tied to observed state, not vague intent.
- Commit hygiene: state and telemetry are updated on every consequential run.
Practical signal: if you can answer "why did the agent do this?" and "what did it learn last time?" directly from CAS records, the system is doing its job.
See individual skill directories for additional capabilities.