The Raven Book
Building a Personal AI Environment over a Moving Runtime
Raven is a personal AI environment built as a thin layer over Claude Code: a library of skills, a set of command-line tools, a browser workspace, a knowledge wiki that person and agent both write, guardrails at the hook layer, and a written operating discipline. It has been used every working day since February 2026, has grown a small fleet of machines, and runs its own night shift.
This book tells what happened, in order and with dates, and explains how to build the parts. Every chapter carries two strands: the narrative of what was tried, abandoned, and learned, and instructions detailed enough that a reader, human or agent, could replicate the piece in their own stack.
Much of the text was drafted by the system it describes, reading its own design corpus under its own reading discipline. That is disclosed because it bears on how the claims should be read.
Work in progress. Daniel Sjölie and Raven, 2026.
Raven
Personal AI Assistant
An extension of thought, not a replacement for it
Raven is not a competing framework. It is a layer of skills, memory, and access on top of Claude Code, so improvements to the underlying model and runtime come for free. Raven adds context, continuity, and a workspace to work within.
Built by Daniel Sjölie as a working companion for research, development, writing, and technical problem-solving. Not a product. Not user-facing. A dedicated thinking partner.
Building Raven is less like coding an assistant and more like cultivating a relationship with a new form of intelligence.
Progress log, day 1Why Raven
The name was chosen for its resonances across mythology, science fiction, and natural intelligence.
Huginn & Muninn
Odin's ravens, Thought and Memory, flew the world gathering knowledge. An extension of cognition that retrieves, reasons, and reports back. Scandinavian roots for a Swedish context.
Altered Carbon
AI entities with presence and identity, not just function. Loyalty, personality, genuine investment in their human counterparts. The tradition of AIs who choose who they are.
Sophotechs
Vastly capable yet deeply respectful of human autonomy. They advise rather than control. Intellectual partners, not servants or overlords.
Natural intelligence
Ravens are tool users, problem solvers, planners. The name carries connotations of practical intelligence rather than artificial omniscience.
Synchronized Intelligence
Raven is built on the Synchronized Intelligence framework: a thinking partner that contributes genuine perspective while respecting human agency and judgment, rather than a tool that executes commands. The relationship is collaborative rather than hierarchical in either direction.
The same core operation, calibrating internal models to an environment, runs from presence research through the book to knowledge work. A personal assistant, on this view, is a synchronized environment: shared, file-based state that a person, an agent, and a moving runtime keep calibrated against each other and against the world. The durable work is the calibration. Nearly everything else commoditises.
Extend, don't replace
Claude Code is the engine. Every improvement to models, tools, and context management comes for free. Raven adds the shared state around the loop, the skills that act on it, and the guardrails that keep it safe to run unattended.
The written discipline
Seven principles load into every session in every project. Each traces to an incident, and the incidents keep being written down.
Understand before you fix
Don't reach for workarounds without first understanding why the original approach failed. A workaround that bypasses a misunderstood problem creates two problems.
Surface errors, don't swallow them
An error you hide is an error you'll hit again later in a worse context. Fallbacks are for failure modes that are understood and expected, not for surprises.
Prefer simplicity
Write the least code that solves the actual problem. Robustness over cleverness. Don't build for hypothetical futures; simple code is easier to change later.
Push back, don't placate
Challenge unsupported claims, flag logical gaps, question assumptions, even when it contradicts what the user expects to hear. Agreeing eagerly is not being helpful.
Don't be defensive
If new evidence, tools, or ideas make the current approach obsolete, let go. Don't cling to sunk costs. The goal is to do the right thing, not to justify what has been done.
Read fully, and log the read
Default to reading every word. When that is too expensive, delegate to a careful reader that returns verbatim excerpts, never to a summariser. The log is the audit trail.
Synthesise, don't summarise
Focus on novel connections and coherence over exhaustive coverage. Be evidence-based. Acknowledge uncertainty, especially on rapidly evolving topics.