How steamHouse Builds Its Books: A Composition Methodology Statement

Document: STEAMHOUSE_AI_COMPOSITION_METHODOLOGY_v1.md Version: 1.0 Created: February 2026 Audience: Academic researchers, institutional partners, and methodological reviewers Status: Current — Phase 7 (Human Editing) active but not yet complete for any title

Why This Statement Exists

steamHouse has published position statements on why we take the position we do on AI in education — that the age of AI makes the development of conscious, purposeful human agency more essential, not less. That argument is made elsewhere (see The Case for Conscious Humans in the Age of AI).

This document is different. It explains how we actually use AI in building our curriculum books — specifically, what role Claude (Anthropic's large language model) plays in the composition process, what role human judgment plays, and where we currently stand in that process.

We are writing this because researchers deserve to know. If you are evaluating the steamHouse knowledge base as a potential partner, collaborator, or critic, you should understand what you are looking at: the outputs of a structured, multi-phase human-AI composition process, not a finished, fully human-edited body of work. That is a meaningful distinction, and we do not want to obscure it.

The Core Pipeline: Seven Phases

Phase 1: Source Curation

Every steamHouse curriculum domain begins with deliberate source selection. Our research bibliography currently spans approximately 1,100 sources across 38 domain categories — covering everything from brain architecture and adolescent psychology to economic systems, historical thinking, and technology ethics.

Sources are not selected opportunistically. Selection criteria include: peer-reviewed or peer-respected status, author credibility and field standing, substantive relevance to the steamHouse developmental framework, and balance across perspectives within contested domains. Where relevant, we also include practitioner literature and well-grounded popular science writing when it synthesizes research meaningfully.

What AI does in this phase: Nothing. Source curation is entirely a human judgment process. Claude does not select sources.

Phase 2: Deep Dive Reference Summaries

Each source of sufficient importance receives a detailed reference summary — a structured analysis of the source's core arguments, key frameworks, relevant concepts, and connection to steamHouse curriculum domains.

These summaries are generated by Claude, drawing on its training data — a vast corpus that includes published books, academic literature, scholarly commentary, citations, and critical discussion of those works. This is not live internet retrieval; it is pattern-based recall from training. The human author does not read each source first and feed notes to Claude — rather, Claude produces the summary from its training data, operating within a human-defined framework specifying what matters and why: relevance to developmental stages, fit with the four steamHouse principles, pedagogical applicability, and domain coverage.

For well-documented works — major academic texts, widely cited books, canonical research — this process produces reliable and substantive summaries. For obscure works, very recent publications, or texts underrepresented in Claude's training data, Claude's knowledge thins and accuracy degrades. In those cases the human must compensate through direct reading, external verification, or exclusion of the source from the compendium.

Researchers should treat these summaries as informed starting points, not as substitutes for primary source engagement. They represent Claude's synthesis of how a work is understood across its training corpus — which means they can reflect dominant interpretations, miss minority scholarly positions, and occasionally misremember specific details. The human's role is to direct, review, and correct; not to generate the raw summaries.

What AI does in this phase: Generates structured reference summaries from training data, within human-defined frameworks. Accuracy is strong for well-documented sources and degrades for obscure or recent ones. Human reviews and corrects.

Phase 3: Compendium Creation (the "Combined Fulsome" Documents)

The most labor-intensive phase of knowledge construction produces what we call the "Combined Fulsome" compendiums — 38 domain documents, totaling approximately 1.6 megabytes of synthesized research content.

Each compendium synthesizes multiple sources within a domain into thematic sections. A compendium is not a bibliography or an annotated list — it is a genuine synthesis that identifies convergences, tensions, and domain-wide patterns across sources. Domain A1 (Brain Architecture and Cognition), for example, synthesizes more than 30 sources into integrated thematic sections on neural architecture, the two-system model, memory, executive function, and developmental neuroscience as applied to adolescent learning.

These documents are built through genuine human-AI collaboration. The human determines scope, thematic structure, and which synthesis judgments matter. Claude drafts, extends, and refines. Multiple passes occur. The compendiums are long and detailed by design — they are intended as rich reference materials, not executive summaries.

What AI does in this phase: Significant drafting and synthesis, under human direction. The intellectual architecture (which themes, which tensions, which implications for steamHouse design) is human-determined. The prose and organizational execution is heavily AI-assisted.

Important caveat for researchers: The compendiums represent our most developed knowledge-base materials. They are the foundation for everything downstream. They are also the phase where AI's role is most substantial. We believe the synthesis is sound, but we welcome scrutiny. If you find errors, misrepresentations of sources, or synthesis failures, we want to know.

Phase 4: Grounding the AI in steamHouse Research

The compendium documents are loaded into Claude's "project knowledge" environment — a persistent knowledge base that Claude can search and retrieve from during subsequent working sessions. This means that when we work on a book, Claude has access not just to its general training, but to steamHouse's specific synthesized research base.

This is a technical infrastructure step, not a content step. But it is important for understanding what Claude is drawing on when it drafts curriculum content: it is drawing on steamHouse's curated, synthesized research — not on general internet retrieval or undifferentiated training data.

What AI does in this phase: Passively stores and retrieves. The human designs the knowledge architecture.

Phase 5: Outline Planning — Scope, Sequence, and Angle

Before any book draft begins, the human author develops a detailed structural plan addressing three questions:

Scope: What does this book include and, equally important, what does it exclude? For a book aimed at middle schoolers about cognitive bias (Tricked: Your Brain Is Lying to You), scope decisions include which of the 200+ documented biases to feature, which to omit, and why. These are editorial and pedagogical judgments, not algorithmic ones.

Sequence: In what order should material be encountered? What must be understood before something else can land? Sequence reflects developmental assumptions about the reader and pedagogical theory about how understanding builds.

Angle: What is the distinctive voice, frame, and relationship with the reader that this book will establish? Tricked, for example, takes a "you're being played — let's figure out by whom and how" angle that is very different from a neutral educational explainer. That angle is a human creative and strategic decision.

Outline planning produces detailed chapter structures, learning objectives, and compositional notes before Claude writes a single paragraph of the book.

What AI does in this phase: Consultation and pressure-testing. The human proposes; Claude responds, identifies gaps, suggests alternatives. Final decisions are human.

Phase 6: AI Draft Generation

With a populated knowledge base and a human-approved outline, Claude generates structured draft content — chapter by chapter, section by section — pulling from the compendium research base and following the outline specifications.

This is where the AI's role becomes most visible as drafting. Claude produces substantial prose. That prose is grounded in the synthesized research (Phase 3) and shaped by the human-designed structure (Phase 5), but it is Claude generating sentences, paragraphs, and sections.

What AI does in this phase: Primary drafting. This is AI doing what it is good at: producing coherent, well-organized prose within a defined structure and knowledge context.

What this means for quality: AI-generated drafts within a well-designed knowledge and outline system are substantially better than AI-generated drafts without one. The compendium-backed, outline-directed approach reduces hallucination, ensures domain accuracy, and produces content that reflects genuine research synthesis rather than generic AI output. It does not eliminate the need for human editing. It makes human editing more productive by producing a strong working draft rather than a blank page.

Phase 7: Human-AI Iterative Editing

The final and most important phase is iterative editing — human review of AI drafts, substantive revision, dialogue with Claude about specific passages, and multiple cycles of improvement until the content meets the standard required for publication.

This phase involves Claude in a responsive, reactive role: responding to human critique, proposing alternatives when asked, flagging potential problems, and executing revisions. The human drives. Claude assists.

Current status — transparency required:

Phase 7 has not been completed for any steamHouse book. We have strong AI-generated drafts across multiple titles, with some titles in early iterative editing. What exists in the project knowledge base are primarily Phase 3 compendium materials (research synthesis) and Phase 6 draft content (AI-generated drafts shaped by human outlines). These have not yet passed through the full human editorial process required to bring them to publication standard.

We are stating this plainly because researchers deserve to know the status of what they are evaluating. The research foundation is solid. The knowledge synthesis is our best current work. The book drafts are working documents, not finished manuscripts.

What This Means for Researchers

Evaluating the knowledge base

The compendium documents (38 Combined Fulsome files) represent the most mature outputs of this process. They are the result of extensive human curation and human-directed synthesis, with significant AI drafting assistance. If you are evaluating the quality of steamHouse's research foundation, these are the appropriate documents to scrutinize. We welcome critique of source selection, synthesis accuracy, and domain balance.

Evaluating the curriculum drafts

The curriculum books in draft form should be understood as working documents — strong working documents, but not finished ones. The pedagogical architecture is human-designed. The prose is AI-generated under human direction and has received some but not complete human editorial review. Treat them accordingly: as evidence of direction and design intent, not as finalized curriculum.

The question of AI authorship

We do not use the word "AI-authored" for steamHouse materials. The intellectual architecture, value framework, source selection, thematic priorities, and editorial judgment that shape these materials are human. Claude's contribution is substantial in drafting and synthesis, but it operates within human-designed constraints at every phase.

We do not use the phrase "AI-assisted" either, because that understates Claude's drafting role in Phases 3 and 6.

The most accurate description: human-directed, AI-drafted, human-edited — with the understanding that the human-editing phase (Phase 7) is the phase currently underway.

Our Tools

Large language model: Claude (Anthropic), specifically Claude Sonnet, accessed via Claude.ai with project knowledge functionality. Project knowledge allows the loaded compendium documents to serve as a persistent, searchable reference base during working sessions.

Knowledge management: All compendium and draft documents are stored as markdown files in the Claude project environment and maintained locally. Version control is manual, with systematic naming conventions.

No other AI tools are currently used in the composition pipeline. Image generation, audio, and other modalities are not part of the current process.

Our Commitments

We commit to updating this methodology statement as the process evolves — particularly as Phase 7 progresses and books move from draft to publication-ready status.

We commit to being responsive to researcher questions about specific documents, sources, or synthesis decisions. If you have questions about how a particular claim in a compendium was synthesized, or why a particular source was included or excluded, ask us.

We commit to not overclaiming the current status of our materials. The research foundation is real. The books are in process.

Document Metadata

File: STEAMHOUSE_AI_COMPOSITION_METHODOLOGY_v1.md Version: 1.0 Created: February 2026 Author: steamHouse / human-AI collaborative drafting Status: Active — to be updated as Phase 7 progresses Audience: Academic researchers, institutional partners, methodological reviewers

steamHouse: Teaching young people how to think, not what to think. Transparent process. Honest status. Open to scrutiny.