Design Course

Design backward from credible performance. A course is one possible intervention, not the default answer to every performance problem.

Choose the scope

Select the smallest mode that can produce the requested outcome:

• Lesson: one bounded capability, usually one sitting.
• Short course or workshop: several linked capabilities delivered over a short period.
• Full course or program: a durable progression with prerequisites, multiple assessments, transfer, and maintenance.
• Audit: diagnose an existing design and propose the smallest viable repair.

Do not expand a lesson into a program or produce a full curriculum when the user asked for an outline. If the request is an immediate explanation for the current learner, use teacher:teach instead.

1. Test whether instruction is the right intervention

For a workplace or product-performance problem, inspect the observed gap before prescribing training. Separate:

• missing knowledge or skill;
• unclear interface, process, ownership, or expectations;
• missing tools, access, time, incentives, or feedback;
• a task that should be simplified, automated, documented, or prevented.

When the demonstrated cause is outside learning, say directly that a course would not address it. Name the better intervention and any narrow learning remainder, using wording suited to the request rather than a fixed banner. Do not hide a product, management, or system repair inside a training plan. Use intervention-diagnosis-template.md when the diagnosis needs to be handed off.

2. Ground the truth and diagnose the learners

Establish the authoritative source of truth, what may have changed, and what remains uncertain. Inspect relevant code, product behavior, documents, standards, policies, research, or expert material before designing factual instruction. Read source-truth.md when claims are current, contested, consequential, or source-heavy.

Define the learner population from evidence where possible:

• current knowledge and representative misconceptions;
• authentic context, tools, language, and constraints;
• accessibility needs and prior opportunity to learn;
• motivation, stakes, and the conditions of later performance.

Use learning-science.md when choosing or defending instructional mechanisms, especially for retention, transfer, feedback, or cognitive-load decisions.

3. Lock performance, evidence, and assistance before lessons

Write the observable terminal performance first. Then specify:

1. what credible evidence would show that performance;
2. the conditions under which it must occur;
3. which references, tools, collaborators, and AI are allowed;
4. which parts must remain unaided human work.

Make assessment resemble the claimed capability. A recall quiz cannot prove diagnosis, judgment, creation, or safe field performance. If the learning claim includes transfer, include a changed context or constraint in the evidence.

For AI-assisted learning or consequential decisions, read ai-assisted-learning.md and use ai-use-contract-template.md. Never use an AI detector as sole evidence of misconduct or competence.

4. Build the learning progression

Sequence prerequisites around meaningful whole tasks. Give novices enough orientation, examples, and structure to start; fade help deliberately as performance stabilizes. Avoid both unguided discovery and permanent hand-holding.

For each unit, connect:

• an authentic problem or decision;
• the mental model or discriminations needed to act;
• a credible expert model or worked example;
• guided practice with useful feedback;
• learner construction or performance;
• later retrieval, variation, and transfer.

Select activity patterns only when they create the cognitive work the outcome needs. Read design-patterns.md for activity selection. Read domain-patterns.md only when the domain or delivery medium changes the design decision.

For self-paced or long-running learning, add lightweight self-regulation support: visible milestones, planning prompts, progress evidence, recovery paths, and timely feedback. Do not bolt this machinery onto a short, instructor-led lesson that does not need it.

5. Prototype before scaling

Produce one representative lesson, task, or assessment before generating a full catalog. Test the hardest or riskiest seam: an authentic task, a central misconception, a consequential assessment, or a difficult delivery constraint. Revise the architecture from that evidence, then scale the accepted pattern.

Use the assets selectively:

• course-brief-template.md to lock the intervention and architecture;
• lesson-spec-template.md for a build-ready lesson;
• assessment-blueprint-template.md for consequential or multi-part evidence;
• source-maintenance-ledger-template.md for changing or distributed source material.

Do not emit every template by default.

6. Apply operational gates

Check accessibility, assessment integrity, technical or disciplinary accuracy, delivery feasibility, data and tool access, and the competence of reviewers or facilitators. Define how outcomes will be measured and what source, product, policy, or learner evidence should trigger maintenance.

High-stakes learning that authorizes consequential action requires the full gate: valid and reliable evidence, fairness and accommodations, security, moderation, appeals, piloting, qualified human review, and observed performance under the real assistance conditions. Protect the human work that the authorization claim depends on.

Read quality-gates.md for audits, releases, and high-stakes designs.

Output by mode

• Lesson: assumptions, a bounded lesson specification, assessment evidence, and essential sources.
• Short course or workshop: course brief, sequence, representative activities, assessment plan, delivery constraints, and pilot.
• Full course or program: architecture and prerequisite map, assessment blueprint, representative prototype, operating model, measurement, and maintenance. Do not manufacture every lesson unless asked.
• Audit: verdict first; then blockers and majors with evidence and consequence; then the smallest repaired architecture and explicit pilot or release conditions. Do not compute an aggregate score.

State material assumptions and unresolved blockers. Keep the design proportional to the requested artifact and the evidence available.