generate-analogy-set

Generate Analogy Set

Table of Contents

• The 5 archetypes
• Workflow
• Worked example
• Guardrails

Related skills: Called by the Intuition Builder agent as step 1. Feeds map-analogy-to-concept, stress-test-analogy, check-analogy-novelty, voice-fitness-check.

The 5 archetypes

Each framing must match one of these archetypes:

1. Everyday analogy — maps the topic to something in ordinary life (cooking, traffic, weather, a household object). Accessible; surface-friendly.
2. Physical metaphor — maps to a physical system (fluid, gravity, spring, circuit, lens). Rigorous when the physics actually transfers; leaky when it doesn't.
3. Contrarian take — inverts the received framing. "People say X; actually Y." Works when the received framing has a known failure mode.
4. Historical angle — shows the topic through its precursor or evolution (attention before Transformer; word2vec → GloVe → BERT; consistent hashing's origin in distributed caches).
5. Counterfactual — "what if this weren't here?" Reveals function by subtracting. (See propose-counterfactual for depth.)

These 5 are fixed. If a topic resists one archetype (rare), the agent produces a weaker version rather than substituting a 6th.

Workflow

Generate 5 framings for topic T:
- [ ] Step 1: Restate T in one sentence (what the writer is explaining)
- [ ] Step 2: Brainstorm 2-3 candidates per archetype
- [ ] Step 3: Pick the strongest candidate per archetype using the voice-profile analogy-direction priority (biology > organizational > sports; NEVER physics/military)
  - Exception: "physical metaphor" archetype explicitly permits physical domain, but even here prefer fluid/biological analogues over mechanical/military ones.
- [ ] Step 4: Write each as one-line framing statement (≤25 words)
- [ ] Step 5: Return the 5 with archetype labels

Priority in selection

The voice-profile says biology → AI, organizational → multi-agent, sports → calibration are the writer's characteristic directions. Within each archetype, if a biology-flavored option exists AND is crisp, it wins over an equally crisp mechanical/military option. The writer almost never uses physics/military.

Worked example

Topic: Attention (in Transformers).

5 framings:

1. Everyday: A crowded table where each person glances around to find the two or three others whose words most reshape what they're about to say.
2. Physical metaphor: A weighted heat diffusion — each token's representation equilibrates against others in proportion to their pairwise affinity, smoothing toward a distribution.
3. Contrarian: Attention is not about "where the model looks" — it's about which other tokens get to edit you, and by how much. The token is the editor, not the camera.
4. Historical: Before attention, seq2seq bottlenecked through a single fixed-length hidden state — "read the whole sentence, then speak." Attention said: let the decoder re-read relevant words each step. Attention is the inversion of the bottleneck.
5. Counterfactual: Remove attention from a transformer and you have a stack of residual MLPs per token — no information ever flows between token positions within a layer. Context is gone. That absence is what attention is "doing."

Each framing is ≤1 sentence (some are 2). The mapping, breaks, and novelty checks come from the downstream skills.

Guardrails

1. Always produce exactly 5 — one per archetype. No substitutions.
2. Each framing is ≤25 words as a one-line statement.
3. Within an archetype, prefer biology / organizational / sports direction unless the topic is physical enough that the physical metaphor IS the right call.
4. Never use military, war, or weapons metaphors. Never combat-code.
5. Never produce prose-ready framings. These are scaffolds; the writer writes the prose.
6. If a topic is too narrow or too broad, return a single "scope-clarification-needed" message instead of forcing 5 weak framings.

Quick reference

• Input: topic string.
• Output: 5 framings × {archetype, statement}.
• Downstream: each framing feeds through mapping, stress-test, novelty check, voice fitness.