practice_first_teaching v1.0.0

name: "practice_first_teaching" description: "Teach by doing: open with a concrete worked example, give just enough theory to name what happened, fade the steps, practice on a deliberately broken example, then redo it alone with a hint ladder and spaced review. Triggers when someone wants to teach, tutor, onboard, or explain a skill or concept, build a lesson or session plan, or fix teaching that is all lecture and no practice." metadata: { "includeInPrompt": true }

Practice-First Teaching

Purpose

Teach a skill or concept in the order that practice supports: example first, theory second, support fading as the learner takes over. A lecture gives a learner words about a thing. This sequence gives them the thing first, then the words for it, then graded practice that ends with the learner working alone. It splits concept work (deciding what rule applies and why) from skill work (carrying the steps out cleanly), because a learner can fail at either one for different reasons and the fix differs.

Use it for tutoring, onboarding, training, and self-teaching plans. The evidence behind each choice, with its limits stated, is in references/evidence.md.

Workflow

1. Split the target. Write two lines before teaching. Concept line: the rule or decision the learner must grasp, in one sentence. Skill line: the steps the learner must carry out without help. If you cannot write the concept line, you are not ready to teach the session yet.

1. Open with a concrete worked example. Show one short example solved end to end, thinking aloud at each step. Keep it small enough to hold in view at once. The learner watches and answers one check question per step: what just happened, and why that move. No theory lecture before this. Novices learn more from studying a worked example than from solving cold, and the example gives the theory something to attach to.

1. Name it, briefly. Give just enough theory to name what happened in the example: one rule, one term, one contrast that matters. Stop there. Any idea the example did not use does not belong in this step, however important it is in the textbook.

1. Fade the steps. Present the same kind of problem with the last step removed. The learner completes it. Remove one more step each round, working backwards, until the learner faces a whole problem. Fading from the end keeps the concept-carrying first steps supported longest, and it is the form Renkl and colleagues tested against plain example-problem pairs.

1. Run broken practice, if the base is clean. Hand over a worked solution with one planted error drawn from a common misconception. The learner finds it, explains why it is wrong in plain words, and fixes the whole solution. Design rules and a pre-flight check are in references/broken-practice-design.md. Gate: if the faded round in step 4 was not clean, skip this step and run a second faded problem instead. Learners with low prior knowledge gain more from correct examples than from incorrect ones, so broken practice never opens a topic and never runs on a shaky base.

1. Redo it alone. Fresh problem, example put away, no help at the start. This is the first honest test of the session. If the learner stalls, use the hint ladder in order and log the highest level reached: level 0 wait, level 1 point at the step, level 2 ask what rule applies, level 3 show a parallel micro-example, level 4 work the step together then erase and redo it, level 5 show the full step and assign a new problem at once. The ladder is in the generated plan from bin/session_plan.py.

1. Close with exit retrieval, then space the review. Closed book: the learner states the rule in plain words and solves one small variant from memory. Log what failed. Schedule reviews at expanding intervals and treat each review as closed-book recall plus one fresh problem, not a re-read. Generate the full plan with dates:

sh
bin/session_plan.py --topic "solving 3 + 2(x+1) = 7" --concept "distribute before combining like terms" --start 2026-10-27 --minutes 45
bin/session_plan.py check-broken --errors 1 --error-type misconception --has-correct-model yes --requires-explain yes

Two complete sessions, algebra and Python, are worked out line by line in references/worked-sessions.md.

Output Contract

A finished teaching session or plan always contains:

• A concept line and a skill line, written separately
• One opening worked example with a check question per step
• A naming step limited to what the example used: one rule, one term, one contrast
• A faded sequence that removes steps from the end backwards, ending in a whole problem
• Either a broken example that passes the design check in references/broken-practice-design.md, or a written reason it was skipped
• A solo redo on a fresh problem, with the highest hint level used recorded
• An exit retrieval result and dated spaced reviews, with the rule for what a failed review changes

Operating Rules

1. Example before theory for novices, always. Theory first is the habit this skill exists to replace. The one exception is a short attempt before instruction for learners who already hold relevant prior knowledge, and that variant still closes with a correct model and a consolidation step in the same session.
2. One planted error per broken example, from a named misconception. Typos, trick errors, and multi-error challenges are not broken practice.
3. Least help first. Hints go up the ladder one level at a time, never straight to a worked step. Log the level reached. A session where the learner needed level 4 hints is a session that found its real teaching point, not a failed session.
4. Never cut the solo redo or the exit retrieval to save time. Cut the naming step or shorten broken practice instead. A session with no unaided attempt proves nothing about learning.
5. Judge concept work and skill work separately in every result you report. "Got it wrong" is not a diagnosis. "Knew the rule, slipped on step 3" is.
6. Fade on evidence, not on schedule. Support drops when the learner's work is clean, and comes back without penalty when a harder variant exposes a gap. Help that stays after it is needed becomes a crutch, and the expertise reversal research says it can start to hurt.
7. A failed review sends the topic back to faded practice, not back to the lecture. Re-explaining what was already explained is rarely the missing piece. Graded practice usually is.
8. Accuracy rule: every claim about what works traces to references/evidence.md. Do not promise effect sizes, retention percentages, or guarantees. Where the evidence is split, as it is on broken examples and attempt-first designs, state the condition that decides it.