Curriculum Draft: AI-Assisted Problem Decomposition
Working document — structure, not final copy
Framework Summary
Teaches problem construction and decomposition using AI as the working medium — not AI tools, not prompt syntax. Organized around three functions:
- Constrain the search space
- Build a model before extracting conclusions
- Introduce adversarial pressure on your own thinking
Modules (draft — five categories, mapped to observable objectives)
Module A — Problem Space Construction
Before substantive AI interaction begins.
Objective: student can front-load full context, pin scope, and define a success condition before asking for solutions.
- Context dump before questions (invert the default ask-first pattern)
- Scope pinning (what's in/out of this session)
- Outcome contract (what a good output looks like, stated upfront)
- Multi-factorial cause splitting (decompose independent variables before solutions)
Module B — Steering Operations
Navigating mid-session.
Objective: student can recognize when a line of inquiry has stalled or narrowed prematurely, and correct without derailing the process.
- Fractal branching (let answers generate new questions, expand before converging)
- Constraint tightening loop (binary-search the solution space)
- Stall detection (wrong input vs. genuinely exhausted problem space)
- Graceful contradiction (correct direction, not just negate)
Module C — Adversarial Pressure
Stress-testing the model and yourself.
Objective: student can generate a genuine counter-case against their own position before committing to it.
- Steelmanning on demand
- Role inversion (AI questions the student, not the reverse)
- Cross-provider verification
- Differential diagnosis ("what else could explain this?")
Open modality question: does this require a live facilitator, or can AI itself hold the adversarial role in a self-paced format? Untested, worth piloting.
Module D — Perspective Architecture
Borrowed from mediation/mentalization practice.
Objective: student can represent a conflicting stakeholder's internal logic accurately enough to examine a conflict from outside their own position.
- Perspective embodiment (AI represents multiple stakeholders faithfully)
- Handoff framing ("explain this to someone else" as a forcing function)
- Third-person self-framing
Module E — Output Engineering
What's demanded from the output.
Objective: student ends a session with a decision, next step, or falsifiable claim — not a summary.
- Forcing actionable outcomes
- Model interrogation (surfacing hidden assumptions behind a recommendation)
- Delta prompting (difference analysis between two states/framings)
Sequencing Logic — Two Distinct Questions
Application order (expert use): flexible and associative — a practitioner reaches for whichever technique the moment calls for. Not a fixed pipeline.
Teaching order (novice acquisition): likely does need scaffolding. Proposed default, open to ID revision:
Module A → Module B → Module C → Module D → Module E
Rationale: can't constrain/build a model without first controlling scope (A); steering (B) needs a model to steer; adversarial pressure (C) needs something built to pressure-test; perspective work (D) is a more advanced application of the same pressure; output engineering (E) closes every session regardless of which modules were used.
This ordering is a hypothesis, not a validated sequence — an explicit area for ID input.
Delivery Modality — Draft Options
| Mode | Status | Notes |
|---|---|---|
| 1:1 | Proven | Original case study; highest personalization, lowest scale |
| Cohort / workshop, live-facilitated | Proposed | Modules A, B, E likely translate cleanly; C, D need group-adapted variants (generic rather than personalized blindspots) |
| Self-paced / async, AI as facilitator | Untested, high-interest | A, B, E likely scriptable as guided exercises; C plausible with AI as sparring partner; D likely needs a live facilitator or is the last module to attempt async |
Assessment — Draft Rubric
Derived from a single case study so far — treat as hypothesis pending validation against more students.
A student has successfully decomposed a problem if they can:
- Restate the real question in one sentence, distinct from their initial framing
- Separate explicit knowns from assumptions
- Name 3+ independent variables/factors rather than a single narrative
- State a success condition before solving
- Hold two competing explanations without collapsing to the first
Outcome measure across a full session/course: entry restatement of the problem vs. exit restatement — the delta between the two is the primary evidence of learning, independent of whether the original problem was fully solved.
Scalability Notes
- Modules are designed to be separable — could license as standalone units (e.g., Module C alone as an "adversarial thinking" add-on to an existing curriculum) rather than only as a full package
- Full transferability to instructors other than the original practitioner is a stated goal, not yet achieved — v1 assumes founder-delivered or co-delivered; v2 target is instructor-independent delivery
This is a working draft intended to be revised jointly with instructional design input, particularly on sequencing validation and async modality design.