Generate-Then-Search Meta-Pattern

Input: $ARGUMENTS

Interpretations

Before executing, identify which interpretation matches the user’s input:

Interpretation 1 — Solve a specific problem: The user has a concrete problem and wants to generate candidate solutions then systematically select the best one using criteria. Interpretation 2 — Choose among known options: The user already has options in mind and needs help defining selection criteria and ranking them. Interpretation 3 — Explore a possibility space: The user wants to map out what is possible in a domain before committing to any direction, emphasizing the generation phase over selection.

If ambiguous, ask: “I can help with solving a specific problem, choosing among known options, or exploring a possibility space — which fits?” If clear from context, proceed with the matching interpretation.

Overview

The fundamental pattern for making cognitive tasks tractable:

1. GENERATE a comprehensive space of possibilities
2. SEARCH within that space using well-defined criteria

Key insights:

• The intelligence is in the CRITERIA, not the search
• Generation can be constrained to exclude obviously bad options
• Comparison/selection is easier than generation from scratch
• Matching to criteria requires less intelligence than judgment

This pattern applies to: interpretation, prediction, communication, reasoning, creativity, understanding, and planning.

Steps

Step 1: Identify the Task

What cognitive task needs to be performed?

Step 2: GENERATE — Create the Possibility Space

Rules for generation:

1. Breadth first: Generate MANY options before evaluating ANY

   • Premature evaluation kills the best options
   • Target: 2-3× more options than you think you need

2. Constrained generation: Pre-filter obviously impossible options

   • Physical impossibility → exclude
   • Violates hard constraints → exclude
   • Everything else → include, even if it seems unlikely

3. Generation methods (combine several):

   • Systematic enumeration (→ /se)
   • Morphological analysis (→ /ma)
   • Cross-domain analogy (→ /cda)
   • Inversion (→ /va, /im)
   • Random combination
   • Expert knowledge

4. Quality check: Is the space COMPREHENSIVE?

   • Have you included the obvious AND the non-obvious?
   • Have you included options from outside your domain?
   • Would someone from a different background add options you missed?

GENERATED SPACE:
Options: [N total]
Generation methods used: [which]
Constraint pre-filter: [what was excluded and why]
Completeness: [confident / gaps possible in: areas]

Step 3: DEFINE CRITERIA — Set Search Parameters

Before searching, define what you’re looking for:

1. Must-have criteria (non-negotiable):

   • [criterion] — how measured
   • [criterion] — how measured

2. Ranked criteria (in order of importance):

   • [criterion] — weight: [1-10]
   • [criterion] — weight: [1-10]

3. Anti-criteria (things to avoid):

   • [what to avoid] — why

Criteria quality check:

• Are criteria INDEPENDENT of each other? (Not measuring the same thing twice)
• Are criteria MEASURABLE? (Not “good” — specifically what makes it good)
• Are criteria RANKED? (You can’t optimize everything equally)
• Do criteria reflect what you ACTUALLY want? (Not what sounds impressive)

Step 4: SEARCH — Apply Criteria to Space

Apply criteria systematically:

Round 1 — Must-have filter: Pass every option through must-have criteria. Eliminate failures. Remaining: [N] of [N original]

Round 2 — Ranked scoring: Score remaining options on ranked criteria:

Round 3 — Top candidates: Take top 3-5 by weighted score for deeper evaluation.

Step 5: EVALUATE — Deep-Dive Top Candidates

For each top candidate:

1. What are the risks specific to this option?
2. What assumptions does it depend on?
3. What would execution look like?
4. What would a critic say about this choice?

Step 6: SELECT

Based on evaluation:

1. Which option has the best combination of score and low risk?
2. Is there a clear winner, or are top options very close?
3. If close: what tiebreaker criterion matters most?
4. → INVOKE: /sel (selection) or /dcp (decision procedure) for formal selection

Step 7: Report

GENERATE-THEN-SEARCH:
Task: [what was being solved]

Generation:
- Options generated: [N]
- Methods: [which generation methods]
- Constraints: [what was pre-filtered]

Search criteria:
- Must-have: [list]
- Ranked: [top 3 criteria with weights]

Results:
- Passed must-have: [N] of [N]
- Top 3:
  1. [option] — score: [N] — key strength: [what]
  2. [option] — score: [N]
  3. [option] — score: [N]

Selected: [option]
Rationale: [why this one]
Key risk: [biggest concern with selected option]

When to Use

• Any task where generating options and selecting among them is more tractable than directly constructing the answer
• When you’re not sure where to start (generate broadly, then narrow)
• When quality of selection criteria matters more than quantity of options
• → INVOKE: /se (systematic enumeration) for generation
• → INVOKE: /gen (candidate generation) for diverse candidates
• → INVOKE: /sel (selection) for formal selection from candidates
• → INVOKE: /m (matching criteria) for defining search criteria

Verification

• Generation phase completed BEFORE evaluation phase
• Multiple generation methods used
• Criteria defined BEFORE searching
• Criteria are independent, measurable, and ranked
• Must-have filter applied first, then ranked scoring
• Top candidates evaluated in depth (not just scored)