Shape

Paste text. Get structure. See what survived.

Shape takes messy text and restructures it into a structured meaning object — not a summary, but a separation of what was said, what was inferred, and what was lost.

Two Profiles

Narrative (narrative_segment_v0) — for life story, memoir, transcript, personal narrative:

time_markers, events, actors, decisions, changes,
felt_experience, open_questions

Concept (concept_blob_v0) — for systems thinking, architecture, theory, design argumentation:

core_claims, layer_models, distinctions, principles,
anti_patterns, design_questions, next_moves

Both profiles always include: declared_loss, signal_level, inference_notes.

What Makes It Different

Shape does 6 things normal summaries don't:

1. Separates explicit from inferred
2. Surfaces distinctions
3. Extracts changes or design rules
4. Preserves open questions
5. Declares loss — what was dropped and why
6. Returns insufficient_signal when the text doesn't have enough structure

Run (web)

cp .env.local.example .env.local
# add your OpenAI key
npm install
npm run dev

Open http://localhost:3000. Paste text, select a profile or let it auto-detect, hit Shape.

How It Works

1. You paste text
2. Shape detects the profile (narrative or concept) or you override
3. The sieve card (lib/cards/sieve_card_v1.json) is loaded as the shaping contract
4. The contract + your text are sent to a model
5. The model follows the contract: preserve structure, separate explicit from inferred, declare loss
6. Output is validated, cast into markdown and JSON, and checked

The card is the instruction. The model is one interpreter.

What You Get Back

• Structured output: profile-specific fields (events/actors/changes OR claims/distinctions/principles)
• Inference notes: what the model inferred vs what was explicitly stated
• Declared loss: what was intentionally dropped and why
• Signal level: strong / weak / insufficient
• Cast: human-readable markdown + machine-readable JSON
• Check: did the shaping preserve what it should have?

View the Card

Click "Sieve Card" in the results to see the exact card that shaped your text. The instruction is inspectable, not hidden.

Engines

Currently: OpenAI (LLM interpreter).

Coming: local engine (heuristic interpreter, no API key, runs offline). Same profiles, same contracts, same checks. Different interpreter.