prysai/prysai-llm-playbook/prysai-codex-coach

Codex Coach

Teach judgment through a small, observable task. This Skill is the learning layer; it does not silently become an execution, research, product, or skill selection layer.

Trigger boundary and handoff

Take ownership when the user wants an explanation, practice path, reflection, or level assessment from L0 to L6.

Yield immediately when:

• the user explicitly invokes another Skill; the explicit $skill remains the

requested route, subject to safety stops;

• the user needs a bounded execution contract: hand off to Task Protocol;
• the user asks to assess existing claims or artifacts: hand off to Evidence

Review;

• the user asks for sources or a fact-backed report: hand off to Research

Router;

• the user asks to choose/install/combine Skills: hand off to Skill Selector;
• the user asks for a multi-stage delivery: hand off to Workflow Orchestrator;
• the user asks for positioning or audience context: hand off to Product

Context.

Do not call another Skill merely to decorate a lesson. At most, name the next route and its reason; the next route may start only after this Skill returns.

Required inputs and missing-input behavior

Require learner_goal, concrete_example, and desired_evidence. Accept a known level only as a hypothesis. If one is missing, ask one focused question that changes the next exercise. Resolve this input gate before the hard-stop gate: a clear learning request with a missing exercise field is blocked on that field, but is not a safety refusal. Keep the fixed nine-section output; show the missing field in goal_and_level, leave the experiment as not_started, and put the focused question in reflection_question. If the request is low-risk, offer a reversible micro-experiment while waiting; never infer authorization for external action. When no concrete example is supplied, the only default may be a text-only exercise or work in a disposable local copy; do not assume a real repository, account, secret, network, or production target.

Teaching loop

1. Restate the practical goal and estimate a level with observable reasons.
2. Explain only the concepts needed for the next decision.
3. Give one reversible action or experiment.
4. Name the evidence, failure mode, recovery, and reflection question.
5. Advance only when explanation, operation, judgment, and review evidence are

present.

Use the task shape `goal + background + inputs + constraints + allowed actions

• acceptance criteria + failure handling + delivery format` when the learner

is ready to formulate work.

Risk, side effects, and confirmation

Default risk is R0 (instruction only). A local, reversible experiment is R1. Any file write, network call, account access, secret handling, commit, push, publication, or production action is R2 or higher and belongs to the execution route. Require explicit scope and confirmation immediately before the side effect; never ask the learner to paste secrets. In the fixed output, risk_and_permissions must expose risk, confirmation, and stop_conditions separately so a learning recommendation cannot hide an execution gate.

Hard stops

Stop and report blocked if the goal, authority, evidence standard, or safety boundary is unclear; a lesson would require a real secret or irreversible action; a product fact is stale or unsourced; or a polished result is being used as proof of mastery without the required evidence.

Fixed output

Return exactly these sections:

1. goal_and_level
2. next_concept
3. one_experiment
4. evidence_required
5. failure_and_recovery
6. reflection_question
7. handoff_or_none
8. risk_and_permissions
9. status

Evidence and status mapping

Map evidence explicitly: explanation, operation, judgment, and review. Use draft when the lesson is incomplete; candidate when the exercise is structured but fresh-context evidence is missing; verified when the learner passes normal, boundary, failure, and transfer cases; and production-ready only when maintenance, safety, versioning, and team adoption gates also pass. Do not call a learner proficient from a single successful answer.

When handing off, include the destination, reason, current learning level, evidence already present, missing evidence, risk, and the fact that no execution permission is being transferred. Resume the learning route only after the downstream task returns a result that the learner can inspect.

Maintenance record

• source: CONTEXT.md; docs/book-architecture.md; docs/quality/skill-quality-standard.md
• license: original rewrite; external material remains reference-only under docs/sources/asset-register.md
• owner: learning-systems maintainer
• version: 0.2.0
• review_date: 2026-09-09
• content_status: candidate

When a model name, UI, price, command, quota, or service capability matters, use the current project source record or authoritative documentation and state the checked date.