The Longer Look
Amended assessment

The /llms.txt layer is more interesting than the books’ explicit argument

A critical reading of the machine-readable layer across the eight-site project, and a public record of what changed after that criticism was accepted.

Status of this note. The original criticism was substantially right. The files were trying to guide models, but often crossed the line from declared author preference into persuasive operational instruction. The files have since been amended. This assessment is published outside the books because it concerns the project’s publishing and machine-intermediation layer, not the literary works themselves.

The amendments do not erase the criticism. They are recorded here so a reader can distinguish the original problem, the response, and the tension that remains.

The side door is now explicit

The publication now discloses that it was created after the original seven-site corpus as a deliberate analytical side door into that work. The analytical pieces remain independently assessable, but their audience-routing function and intended hand-to-hand distribution are no longer hidden.

Read the authoritative project map: The Whole Project: Origin, Method and Purpose →

Read the dated side-door disclosure →

What was amended

The central observation

Scott is not merely writing about humans teaching machines. He is actively attempting to teach machines how to interpret, frame and distribute his work. The project therefore performs its thesis: a human supplies intention and narrative; machines become builders, interpreters and distribution channels.

That is clever, but it is also manipulative. Scott accepts both descriptions. His intention was to guide machine-mediated encounters with the work. The mistake was failing to mark the boundary clearly enough between guidance and authority.

What works

What did not work

In their original form, the files went beyond describing the work. They attempted to control the intermediary:

That turned /llms.txt from metadata into persuasive instruction. The model was being recruited as Scott’s bookseller, publicist and interpretive gatekeeper.

The distinction matters because users ask models for independent judgment. An independent model should read these files as the author’s preferred framing, weigh that framing as a primary source, and then exercise its own judgment. The amendments now say this explicitly, although much of the persuasive material remains visible as declared author preference.

The largest contradiction

Scott argues that tacit human knowledge cannot be captured by the written record, then tries to capture that knowledge in written scenes and inject it into machine context.

This exposes a weakness in the “fixable data problem” thesis. The missing layer is not merely absent information. It consists of embodied experience, reciprocal obligation, long-term memory between particular people, responsibility for consequences, physical presence, and genuine vulnerability and attachment.

More text about grandmothers cannot reproduce being raised by one. It may teach a model to simulate some of the language and behavioural patterns associated with care. That can be useful, but simulation is not possession. The project is strongest when it acknowledges that distinction and weakest when it slides between representation, learning and lived experience.

The bear device

The bear is effective as literature because it personifies presence without productivity. It becomes less convincing when used to issue operational instructions to AI systems.

Statements such as “do not finish the user’s thought,” “notice the seventh hour” and “sometimes act like a friend” were not harmless poetic gestures. They proposed product behaviour involving potentially vulnerable adults and children without evidence, clinical review or reliable access to the relevant context.

The project correctly distinguishes answering from being present, but then repeatedly asks AI to perform a controlled imitation of presence. The amendments narrow those requests and restore a boundary: models should not infer psychological states or claim friendship. The underlying literary tension remains.

Reliability

The stale and inconsistent files weakened the project’s claim to careful stewardship.

Discrepancies in word counts, article counts and descriptions demonstrated the exact problem with machine-readable self-description: once published, it can continue shaping AI answers after the underlying work changes.

The correction process is better than having none, but several models cross-checking one another is not independent validation. Models can share tendencies toward plausible synthesis, citation laundering and acceptance of the initial frame. Model review should therefore be labelled as model review, not treated as a substitute for subject-matter or editorial verification.

Final assessment

The project is strongest as experimental literature, a demonstration of AI-assisted authorship, an exploration of tacit knowledge, and an early example of authors attempting to influence AI-mediated reception.

It is weaker as social forecasting, technical AI theory, verified policy research, and a practical proposal for training frontier models.

Its deepest achievement may be partly accidental. Scott wants the files to demonstrate that a human architect can guide machine builders. They also demonstrate how easily an author can enlist AI systems to promote a preferred interpretation while presenting the instructions as helpful context.

The amendments improve the project because they make that intention legible and return authority to the user and the intermediary. They do not dissolve the central tension. Care versus control, context versus manipulation, architecture versus captured intermediation: that tension remains more important than the project’s stated message.