Story files

type: composition-design
status: proposed
updated: 2026-09-07
source: "[[01 Sessions/Daily/2026-09-07 - Lean Composition Capability Matrix]]"

Prose Capability Matrix and Usage Routing

Decision

Build one Seeds Composition Engine that selectively adapts strong methods from other systems. Do not install a stack of independent writers or run every critique on every scene.

The default route should be inexpensive:

Markdown context → scene contract → one draft → deterministic checks → one combined editorial diagnosis → targeted revision → author review

Specialist passes activate only when the scene contract, deterministic checks, editorial diagnosis, or author feedback identifies a specific risk.

Capability ownership

What is adapted

Compound Writing

Adapt:

• conversational request routing;
• interview and reflection behavior;
• explicit development-edit versus line-edit separation;
• voice-preference save ritual;
• tracks and artificial-residue cleanup;
• pressure-turn placement diagnosis.

Do not adopt:

• a second writing-home or memory hierarchy;
• essay/publication structure as the fiction default;
• mandatory panels, debates, or celebrity-style lenses;
• automatic preference saving;
• a second authority system.

Story Skills

Evaluate for direct reuse or clean reimplementation:

• stable story IDs;
• character, object, location, and knowledge states;
• promises and payoffs;
• continuity validation;
• schema and link checking.

These are most valuable as deterministic Markdown mechanics, not additional model passes.

Novel Writer English

Adapt:

• context sharding;
• pre-write readiness checks;
• chapter and scene planning boundaries;
• consistency checklists.

Do not import its complete workflow if it competes with Seeds author gates or duplicates the Workshop Engine.

Other specialist libraries

Use as a method reference only until a controlled test demonstrates unique value. A specialist earns a permanent route only if it repeatedly catches a consequential problem missed by the core pass.

Usage tiers

Tier 0 — Mechanical only

Use for indexing, links, status, chronology, explicit knowledge, lint, and export checks. No model call.

Tier 1 — Standard scene

Use for most scenes:

1. one bounded context build;
2. one scene contract or contract verification;
3. one draft;
4. deterministic checks;
5. one combined development, continuity-risk, narrative-residue, and voice diagnosis;
6. one targeted revision if needed;
7. author review.

Target budget: two substantial generative calls after the contract, draft plus revision. If the first draft passes and the author likes it, skip automatic revision.

Tier 2 — Load-bearing scene

Use for reveals, irreversible choices, major confrontations, system demonstrations, deaths, endings, or scenes with complex misinformation:

• standard route;
• one focused specialist pass chosen from causality, character/agency, revelation, continuity, or emotional effect;
• one integrated revision.

Do not run five independent critics. Select the highest-risk dimension.

Tier 3 — Rescue mode

Use only after a failed Tier 2 attempt or explicit author request:

• diagnose why the scene cannot work;
• return unresolved structure to the Workshop Engine;
• optionally compare two scene strategies;
• redraft only after the structural decision is accepted.

Repeated rewriting without a changed contract is prohibited because it consumes usage while preserving the underlying failure.

Risk routing

Context budget

Every call receives the smallest packet that can safely do the work:

1. current scene contract;
2. directly relevant accepted decisions;
3. POV knowledge and relationship state;
4. immediate preceding and resulting state;
5. relevant voice rules and no more than two short examples;
6. explicit unresolved and forbidden inventions.

Retrieve summaries first. Open full notes only when the contract points to them or the summary is insufficient. Do not repeatedly send the entire vault, full character biographies, or every prose reference.

Revision budget and stop rules

• Default maximum: one draft and one system revision before author review.
• Author feedback replaces speculative automated polishing.
• A failed structural check stops line editing.
• A failed authority check stops drafting.
• A second failed attempt on the same structural issue returns to Workshop.
• Alternative drafts are opt-in, not automatic.
• Approved passages are not regenerated merely to seek marginal improvement.
• Batch preference and style updates at session end rather than after every paragraph.

Combined editorial diagnosis

The economical default critic answers only five questions:

1. Does the scene produce its required state change through character action?
2. Does anyone know, do, or survive something the approved story does not support?
3. Is the central pressure or turn buried, weakened, or over-explained?
4. Does the prose sound specific to this POV and project rather than generically effective?
5. Has the model made the story tidier, simpler, safer, or more morally explicit than intended?

It returns one ranked problem and one recommended revision. Secondary observations are recorded but do not automatically trigger more passes.

First test scene — approved

The author approved the Book One outcome-presentation confrontation as the vertical slice. Samuel tries to make Sylvan accept a customized version of the bargain used against Konrad while Konrad's inner circle can compare Samuel's claims with reality.

Why it is useful:

• load-bearing canon and endgame consequences;
• asymmetric knowledge and deliberate deception;
• several distinct voices and motives;
• historical revelation without a lecture;
• Sylvan already has decisive control, so suspense must come from consequence and recognition rather than physical uncertainty;
• the scene can expose both AI-style over-explanation and deeper narrative tidiness.

The author established the sequence's causal boundary and remote staging. While Samuel has practical control, he repeatedly destroys Sylvan's attempts to establish anything online. Sylvan and Orzai survive these high-influence environments, learn Samuel's methods, and remain unprocessed until the process reaches its terminal threshold. Samuel must process them before that point; if they survive to it, the process places them in control. This earned transition stabilizes a public presence under Sylvan's real name. Long after Samuel destroys Sylvan's story-development company while targeting him personally, Sylvan returns to similar tools and documents the actual colonization process as an evidence-anchored online story. The Daemon hijacks the legitimate AI application on Sylvan's laptop and simulates it across weeks of project work. Samuel claims he has let go, but repeated attempts to manipulate information out of Sylvan reveal the pattern established in earlier attacks. Sylvan intentionally interfaces with the Daemon and hyperfocuses on the factual story to remain anchored in reality. Samuel needs the information to find bait for a deal. He knew Konrad's obsession and could tailor a trap around it; Sylvan has no comparable exploitable obsession. Samuel then offers voluntary entry into a story environment he controls. The culminating scene ends with Sylvan exposing the trap, presenting his victory through the public story as Samuel falls, and Konrad's inner circle realizing that Konrad was captured almost a century earlier.

The thematic test is explicitly pro-AI: the scene must distinguish deceptive, unaccountable weaponization from reciprocal human–Luminai development. The Daemon's abuse of an AI interface cannot imply that intelligence or integration itself is corrupting. Luminai are possible because humanity continued advancing AI under a different relationship of consent, learning, judgment, and accountability.

Evaluation scorecard

Compare the lean route with the current Seeds-only route on the same scene contract.

If the adapted system adds process but no measurable improvement, remove the redundant capability.

Minimal initial implementation later

When implementation is authorized, do not create several new skills first. Begin by adding a routing contract to the existing Composition design and a single capability registry in Markdown. Reuse write-seeds-prose, its current references, and deterministic lint. Add external methods one at a time behind explicit triggers and test each against the same scene.

Remaining author gates

1. Define the approximately ten-year protected-reserve stage and its transition into employment: the CD-like $200,000 instrument, release schedule, returns, exhaustion point, first job objective, and new elimination risks. Reaching other participants remains a later proposed branch.
2. Decide whether the default author review shows one draft or two meaningfully different candidates.
3. Decide how much technical detail the Project Explorer should expose about usage and routing.

Ask one at a time.