design-learning-path

Files (6)

• DURABILITY-GATES.md
• EXAMPLES.md
• FOUNDATIONS.md
• LEVERAGE-SELECTION.md
• PATH-FORMAT.md
• SKILL.md

SKILL.md (first lines)

# Design Learning Path

## Compounding Objective

Build a step-by-step structured path to the target state. This skill sequences leverage-ordered transitions; `learn-deeply`, when available, executes and verifies one selected step. For mission shape and agency, read [FOUNDATIONS.md](./FOUNDATIONS.md). For step selection and replanning, read [LEVERAGE-SELECTION.md](./LEVERAGE-SELECTION.md). Order it so each step
compounds competency, pushes the learner into a pit of success, and increases
the rate at which succeeding steps can be learned by reusing relevant prior
foundations.

Each model-transition step must:

- Increase current generative capability
- Build foundations reused by several later steps
- Reduce the friction and working memory required by later learning
- Recontextualize and compress prior knowledge
- Unlock better questions, probes, feedback, or artifacts
- Make the correct next action the easiest natural action: a pit of success

Treat "exponential" as multiplicative reuse and increasing learning leverage,
not a numerical promise. Each step is one model transition with a named causal
compounding mechanism.

## Establish the Target

1. Classify the mission: `unblock-build`, `abstraction-untangle`,
   `research-map`, `layered-capability`, or `transfer-synthesis`. It changes
   graph breadth, artifact, and return condition, not the curriculum invariant.
2. Define the target generative capability, use conditions, and why it matters.
3. Define observable independent success and explicit exclusions.
4. Read relevant global learner evidence and local mission state.
5. Verify claimed prior knowledge before using it to compress the path.

## Autopsy the Domain

Ground unstable claims in high-trust, version-matched sources. Identify:

- What experts perceive, predict, debug, or derive differently
- Core primitives, invariants, boundary conditions, and candidate coordinate shifts
- Prerequisite relations and high-leverage thresholds
- Common limiting models and where they fail
- Real use conditions and feedback loops

## Build the Graph

1. Map grounded thresholds. If absent, mine candidates by comparing varied
   cases, rotating representations, and testing which invariant survives.
2. Locate the learner using predictions, explanations, artifacts, or transfer.
3. Choose the highest-leverage threshold among currently reachable nodes.
4. Add only prerequisites needed for a meaningful collision and verification.
5. Order steps so each consumes foundations from earlier steps and lowers the
   cost of later ones.
6. Schedule recontextualization and replan after every major threshold.

Treat mined thresholds as hypotheses until domain evidence and learner behavior
support them. Leverage chooses the destination; capability chooses step size and scaffolding.

## Design Each Path Step

Define:

- Capability unlocked
- Prior foundations consumed
- Limiting model and useful boundary
- New invariant or generative model
- Engineered breaking experience
- Artifact and feedback loop
- Generative gate
- Future friction removed or later steps accelerated
- Pit-of-success environment

## Validate the Sequence

- Every threshold has a named compounding mechanism.
- Every foundation is consumed by a later step; no inert prerequisites.
- Every step produces evidence and makes at least one successor easier.
- The order beats plausible alternatives by leverage, not convention.
- Gates test generation, reconstruction, or transfer.
- Real use conditions appear early enough for feedback to shape the path.
- New learner evidence can reorder, split, insert, or remove steps.

For post-unlock retrieval and transfer without the original scaffold, use
[DURABILITY-GATES.md](./DURABILITY-GATES.md).

Write the result using [PATH-FORMAT.md](./PATH-FORMAT.md). In an active
workspace, update `.learning/CURRICULUM.md`.
See [EXAMPLES.md](./EXAMPLES.md) for concrete path and micro-path examples.

## Do Not

- Return a topic inventory or generic beginner-to-advanced ladder.
- Claim exponential growth without naming the reuse mechanism.
- Treat the graph as a fixed perfect DAG.
- Infer mastery from exposure, credentials, confidence, or correct output.

From Rust-soham/skills, MIT.

More from Rust-soham/skills