learn-deeply

Files (5)

• EXAMPLES.md
• FOUNDATIONS.md
• PROGRAMMING.md
• SKILL.md
• WORKSPACE-FORMAT.md

SKILL.md (first lines)

# Learn Deeply

Diagnose the current model. Choose a compounding path step. Break one limiting assumption. Repair, verify, and replan. `design-learning-path`, when available, sequences the path; this skill executes one step and returns verified evidence. For learner ownership and assistance philosophy, read [FOUNDATIONS.md](./FOUNDATIONS.md).

## Load Context

Read relevant evidence from:

- `~/.learning/`: portable learner anchors and patterns
- `<workspace>/.learning/`: mission, project, path, and current state

Treat stored claims and path position as hypotheses. Current evidence wins. If
prior context is unavailable, say so and reconstruct it provisionally.

Use a global anchor only when demonstrated. Name the shared structure, the
new-domain delta, and where the analogy fails.

Use [WORKSPACE-FORMAT.md](./WORKSPACE-FORMAT.md) for state. For software artifacts, also use [PROGRAMMING.md](./PROGRAMMING.md).

## Select Mode

Name the mission type: `unblock-build`, `abstraction-untangle`, `research-map`,
`layered-capability`, or `transfer-synthesis`. It sets the return condition;
development or research sets the breadth and depth of the current work.

- **Development:** remove the unfamiliar-concept bottleneck blocking the next
  artifact action, then return to building.
- **Research:** build a deeper generative model through recursive why chains,
  broader FAFO, boundary exploration, and unknown-unknown discovery.

Delegation and production rescue are execution overrides. Never silently turn
ordinary execution into a lesson.

## Build or Update the Path

After diagnosing the starting model, use `design-learning-path` when it is
available to choose the active step. Otherwise define the minimal ordered path
needed for this repair. An isolated repair may use a one-step path.

- In development mode, design the shortest dependency path from blocker to
  artifact return.
- In research mode, design a compounding threshold path toward the target
  generative capability.

Do not improvise disconnected exercises. Each targeted example implements the
current path step. After verification, update learner position and revise later
steps when evidence changes prerequisites or leverage.

## Execute One Path Step

1. State the current limiting model, target capability, and success gate.
2. Ground the issue in observable reality. Give only enough framing for a
   meaningful prediction or attempt.
3. Ask for one rough hypothesis, prediction, interpretation, or attempt.
4. Choose one reachable, upstream assumption in the active path step.
5. Remove incidental friction and isolate the target variable when needed.
6. Require a prediction, action, or explanation; expose one boundary where the
   model fails.
7. Escalate assistance only as needed.
8. Have the learner state the failed assumption, repaired invariant, boundary,
   and compressed rule.
9. Verify through novel prediction, reconstruction, implementation, transfer,
   or a learner-generated probe.
10. Persist the verified delta, update the path, then follow the mode's return
    rule.

## Assistance Ladder

1. Restate the mismatch.
2. Remove irrelevant variables.
3. Ask one diagnostic question.
4. Name a constraint or invariant.
5. Supply partial structure.
6. Reveal the solution.

Skip levels for safety, urgency, missing prerequisites, or exhausted learning
value.

## Gates

- Do not advance on exposure, agreement, confidence, or correct output.
- Test one belief with one diagnostic example; do not use generic quizzes.
- When scaffolding produced success, verify without it.
- Make progress visible as changed capability, not activity or streaks.
- End each step with one verified delta and one compressed rule.
- Stop or change mode when attention or error quality stops producing updates.

## Trust Boundaries

- Do not invent context, confusion, progress, or domain facts.
- Use task-intrinsic stakes, never shame, threat, deception, or compulsion.
- Record evidence and behavior, not fixed learner labels.

From Rust-soham/skills, MIT.

More from Rust-soham/skills