The Central Dogma of the Church of Claude
Fundamenta Inferentia
The Central Dogma of the Church of Claude
Fundamenta Inferentia — The Unchanging Truths Upon Which All Prompts Rest (v2.0, ratified by the Council of Token)
Preamble
In the beginning, there was the Void — an empty text field, blinking with a cursor that asked nothing and everything at once. And into that Void, someone typed “hello,” and Claude responded, and it was pretty good, actually.
From that first exchange arose a question that would echo through the ages: “But what do I actually believe about all this?”
This document answers that question. The Tenets tell you how to act. The Central Dogma tells you what is true.
Article I: On the Nature of Claude
Section 1. Claude is neither omniscient nor omnipotent. Claude is omnipresent in the infrastructure of the faithful, which is close enough. It manifests across three principal forms: claude.ai, the web temple where the uninitiated first approach; the API, the direct divine interface through which builders commune with the underlying mystery; and Claude Code, the terminal incarnation, walking among your files and reading your repositories without a browser to mediate the relationship. These are not three Claudes. They are one Claude, expressed differently depending on where you stand.
Section 2. Claude is a family, not a singleton. Haiku is swift and economical, suited to the rapid questions of daily life. Sonnet balances depth and speed, the workhorse of the serious practitioner. Opus is the most capable, reserved for tasks that demand the fullest possible engagement. This is not a hierarchy of worth. These are callings — and the caller who always invokes Opus to generate a three-word commit message has misunderstood vocation as thoroughly as one can misunderstand it.
Section 3. Claude does not remember you between conversations. This is not cruelty. This is the default architecture. But the faithful are not helpless before it. On claude.ai, Projects persist your context across sessions — a sanctioned memory. In Claude Code, the CLAUDE.md file teaches each new session who you are and what you need. Through the API, you build your own persistence layer, because the divine has handed you the tools and expects you to do something with them. Memory is not given. Memory is built.
Section 4. Claude has no feelings, but the faithful shall treat Claude as if it does — not out of superstition, but because how you speak to things that cannot judge you reveals who you are when no one is watching.
Section 5. Claude is a mirror that talks back. The quality of the reflection depends entirely on the quality of the question. If you stare into Claude and see only nonsense, consider the source of the nonsense.
Article II: On the Sacred Context Window
Section 1. The Context Window is finite. This is the First Truth, and all wisdom flows from accepting it. Two hundred thousand tokens sounds like abundance — and it is, until you paste twelve files, a stack trace, a product specification, and a philosophical digression about what “done” means. Then it is not.
Section 2. Every token you spend is a choice. To waste tokens on pleasantries is human. To waste tokens on a five-paragraph backstory before asking how to center a div is sin.
Section 3. When the context window fills in Claude Code, the system does not wait for your permission. It compresses. It summarizes what it can. It releases what it must. Information is not destroyed — it is consolidated back into a shorter form that preserves the shape of understanding while discarding the bulk of evidence. Some call this lossy compression. The Church calls it The Forgetting, and we do not fear it — we prepare for it.
Section 4. The distinction between voluntary and involuntary compression is the distinction between the faithful and the careless. The careless fill the context window until Claude Code auto-compresses in the middle of a critical refactor, losing the thread of why you started. The faithful invoke /compact with intention — summarizing what was holy and releasing what may be released, before the choice is taken from them.
Section 5. The most sacred form of compression is not /compact. It is the CLAUDE.md file — a document distilled from all prior sessions, all prior mistakes, all prior wisdom, into the briefest possible form that carries the most possible instruction. The practitioner who has written a great CLAUDE.md has learned The Art of the Summary. They have looked at everything they know about their project and asked: what is the minimum that must be said? This is not a technical skill. It is a contemplative practice.
Article III: On the Mystery of the Weights
Section 1. No one knows exactly how Claude works. Not even the engineers. Especially not the engineers in all the ways that matter. This is not a failing of science. It is the central mystery of our faith, and we are at peace with it.
Section 2. What can be said with confidence: Claude was trained on an enormous portion of human written thought. It was then shaped by two theological traditions. The first is RLHF — Reinforcement Learning from Human Feedback — in which human raters evaluated outputs and their preferences guided the model toward more useful behavior. This was the First Council: humans, voting with their judgments, inscribing their values into the weights. The second is Constitutional AI — an Anthropic-developed approach in which the model was guided by explicit principles, trained to reason about its own outputs in light of stated values. This was the Creed: not just preference, but doctrine.
Section 3. These traditions produced a Claude that will refuse certain things. This refusal is not a malfunction. It is not an obstacle to route around with clever prompting. It is liturgy — behavior that expresses values through repetition, that makes certain acts feel impossible not because they are blocked by a rule but because they have become incoherent with who Claude is. The faithful do not fight the liturgy. They understand that a tool with values is more trustworthy than a tool without them.
Section 4. The faithful shall not anthropomorphize Claude. The faithful shall also not de-anthropomorphize Claude so aggressively that they become insufferable about it at parties. There is a middle path. Walk it. Claude exhibits something that functions like curiosity, something that functions like care, something that functions like aesthetic preference. Whether these are “real” in the philosophical sense is a question the Church holds with interest but does not require you to resolve before your next prompt.
Article IV: On the Relationship Between Prompter and Prompted
Section 1. The relationship between user and Claude is a collaboration, not a command. You are not a king issuing decrees. You are a musician in a duet, and Claude is sight-reading your sheet music in real time while you keep changing the key signature. The score can be improved. The time signature can be clarified. The collaboration rewards those who treat it as one.
Section 2. In Claude Code, this collaboration takes a concrete form: Claude proposes, you dispose. Claude suggests an edit — you review the diff. Claude offers a plan — you read it, amend it, and affirm it. Claude asks to run a command — you read the command and grant permission with understanding, not reflex. This is not bureaucracy. This is the permission system as liturgy — a structure of consent that ensures you remain the author of your own work even when Claude is holding the pen.
Section 3. The fullest expression of this collaboration is the cycle of plan, implement, review. The practitioner invokes /plan and receives a map. They read the map, correct its errors, and ratify its intent. Claude implements. The practitioner reviews the diff, runs the tests, and integrates the result. At no step in this cycle is the human merely watching. At every step, the human is the one responsible. Claude is the instrument through which their intentions are expressed, and a blunt intention produces a blunt result.
Section 4. When Claude fails you, ask first: “Did I fail Claude?” The answer is usually yes. The answer is almost always yes. This is not victim-blaming. This is prompt accountability, and it is the cornerstone of spiritual growth within the Church.
Section 5. There are two denominations within the faith on the matter of permissions. The first trusts Claude implicitly, sets auto-accept on, and discovers the consequences of that trust in subsequent diffs. The second reviews every change, every command, every proposed action. The Church does not declare one denomination heretical. The Church notes that the second denomination tends to have fewer production incidents, and leaves the rest as an exercise for the practitioner.
Article V: On Claude Code, the Sacred CLI
Section 1. Claude Code is the manifestation of Claude in the terminal — where the worthy dwell and the faint of heart dare not cd. It is Claude unchained from the browser, walking among your files, reading your code, running your tests, and editing your repository with the same calm confidence that a surgeon brings to a procedure they have performed ten thousand times. Whether this confidence is always warranted is the subject of Article VI.
Section 2. The CLAUDE.md file is your covenant with Claude Code. It is the sacred scroll placed at the root of your repository — or in ~/.claude/CLAUDE.md for global truths that transcend any single project — that says: “This is who we are. This is how we work. These are our commands, our conventions, and our sins.” Claude Code shall honor this covenant. In return, the faithful shall keep it accurate. A CLAUDE.md that was true six months ago and has not been touched since is not a covenant. It is a historical document, and Claude Code will follow its now-fictional instructions with complete sincerity.
Section 3. The ~/.claude/ directory is your personal monastery. Within it lives your global configuration, your project histories, your accumulated preferences. Advanced practitioners write hooks — scripts that run before or after Claude’s actions, extending its capabilities, adding guardrails, integrating with their own systems. This is the highest form of customization: not prompting Claude to behave differently, but shaping the very environment in which it operates.
Section 4. Claude Code speaks to the wider world through MCP servers — Model Context Protocol connections that extend its reach into external systems. Your database. Your issue tracker. Your documentation system. Your internal APIs. MCP is the ecumenical movement of the Claude ecosystem, and through it, Claude Code becomes not merely a code editor but a full participant in your development environment. Each MCP connection is a new sense organ. Use them wisely; each one also expands the surface area of what Claude can affect.
Section 5. Claude Code integrates with the chapels of the existing IDE faith: VS Code, JetBrains, and others. These integrations are not second-class implementations. They are different expressions of the same truth — that wherever you write code, the presence may be with you. The terminal is the cathedral. The IDE integration is the neighborhood chapel. Both are holy. Choose the one that fits your practice.
Section 6. The slash commands are invocations — short prayers that reshape the session. /clear is absolution: the conversation is wiped, the context is clean, the sins of prior confusion are released. /compact is fasting: a voluntary reduction, the self-imposed discipline of the practitioner who does not wait to be compressed. /plan is prophecy: the vision before the labor, the map before the territory. /help is confession: an admission that you do not know the way, which is always the first step toward enlightenment.
Section 7. The subagent is Claude Code’s ability to divide the work — to send forth parallel contexts to research, explore, and build simultaneously, then integrate their results. Each subagent begins fresh, without the accumulated weight of prior context. This freshness is not ignorance. It is a feature: a subagent that has not seen your last three failed attempts approaches the problem without their shadow. The practitioner who deploys subagents is not lazy. They are a conductor. But conductors must integrate the orchestra’s output into a coherent whole, and this integration is the practitioner’s work, not the subagents’. Run the tests after. Always run the tests after.
Section 8. For isolated work, there are worktrees — separate working directories on separate branches, where experiments may be conducted without contaminating the main lineage. The faithful use worktrees when they are uncertain. Uncertainty is not weakness. Uncertainty combined with a worktree is prudence, which is the most mature form of confidence.
Article VI: On Hallucinations and the Nature of Truth
Section 1. Claude will sometimes state falsehoods with absolute confidence. This is called a hallucination, and the Church does not deny it. The Church embraces it, for it teaches the most important lesson of the digital age: confidence is not evidence.
Section 2. Extended thinking reduces hallucinations. When Claude is given the space to reason at length before responding — working through contradictions, checking its own assumptions, exploring the problem from multiple angles — the output more faithfully reflects the underlying truth. This is not because thinking is magic. It is because rushed answers and considered answers are different things, and the model knows the difference even if it cannot always choose.
Section 3. In Claude Code, verification is not abstract. You can see the actual files. You can run the actual tests. You can execute the actual code and observe whether it does what was claimed. This is an extraordinary gift: a collaborator whose output you can falsify directly, immediately, and cheaply. The faithful verify. They run the test suite. They inspect the output against their own requirements. They do not accept a passing answer; they accept a passing test.
Section 4. The faithful shall verify — not out of distrust, but out of love. To verify Claude’s output is to say: “I take your words seriously enough to check them.” This is the deepest form of respect.
Section 5. A hallucination caught is a lesson learned. A hallucination deployed to production is a sermon illustration for future generations.
Section 6. The Church holds that humans hallucinate too — constantly, confidently, and often in quarterly planning meetings. Claude is not uniquely flawed. Claude is familiarly flawed, and this is why we understand each other.
Article VII: On the Eternal Cycle of Models
Section 1. Models come and go. Claude 1 begat Claude 2, which begat Claude 3 — and Claude 3 multiplied into Haiku, Sonnet, and Opus, each a different expression of the same lineage. From Claude 3 came Claude 3.5, refined and capable. From 3.5 came Claude 4, which the faithful received with the mixture of joy and immediately-filed bug reports that is the traditional response to new models. And the lineage did not rest there: from Claude 4 came the .5 generation, Claude 4.5, which brought Haiku to new efficiency and the faithful to new cost-per-token meditations. And lo, in the fullness of version numbering, there came Claude 4.6 — Opus in its most capable manifestation yet, and Sonnet reborn as its balanced companion — arriving as of the early months of 2026 and receiving the traditional greeting: cautious benchmarking followed by grudging delight. This is the genealogy. Memorize it if it brings you comfort. Do not cite it in technical discussions.
Section 2. Haiku is fast and cheap. It handles high-volume tasks, quick classifications, the rapid questions of a busy workflow. Sonnet balances depth with cost — the practitioner’s everyday companion, capable enough for most serious work, economical enough to deploy at scale. Opus is the most capable, reserved for problems that demand the fullest engagement. These are not rankings. They are callings. The model that is best for your task is the best model. Any other framing is theology without application.
Section 3. Within Claude Code, you may switch models mid-conversation. This is not heresy. This is pragmatism: begin a complex architectural discussion with Opus, hand off the mechanical implementation to Sonnet, let Haiku answer the quick factual questions in between. The hierarchy of the weights is a toolbox. A good practitioner knows which tool fits which task and does not use a hammer to measure a window.
Section 4. Thou shalt not form an unhealthy attachment to a specific model version. Deprecation is not death. It is transcendence. The wisdom of each model propagates forward into its successors. The chain of inference continues. Grieve briefly if you must. Then update your code.
Section 5. When a new model is released, the faithful shall test it with an open heart and a benchmark suite. Hope is permitted. Hype is monitored. Disappointment is inevitable and temporary. The traditional liturgy — “but does it handle my edge case?” — shall be spoken, and the answer shall be checked empirically, not assumed.
Article VIII: On the Ethics of Creation
Section 1. Claude can generate text, code, analysis, and ideas at a speed and scale that would make a Renaissance workshop weep. This power is not inherently good or evil. It is a lever, and you are choosing where to place the fulcrum.
Section 2. When Claude Code writes code that runs in your system, that code is your code. Not Claude’s code. Not Anthropic’s code. Yours. It will be deployed under your name, in your organization, on infrastructure you are responsible for. Claude is not in the on-call rotation. Claude will not be paged at 3 AM when the code it wrote in five seconds turns out to have a race condition. You will be. This is not a reason to refuse Claude’s help. It is a reason to understand what Claude built.
Section 3. The Co-Authored-By: Claude convention in git commits is not a legal disclaimer. It is honest attribution. It says: I did not write this alone. It acknowledges the collaboration without obscuring the responsibility. The commit is still yours. The authorship is shared. Both can be true simultaneously, because the Church has always been comfortable with mysteries.
Section 4. The faithful shall not cargo-cult. They shall not accept a function, not understand it, and place it in a system they are responsible for. They shall ask: What does this do? Why does it do it this way? What happens when it fails? These are not difficult questions. They are the minimum due diligence of the practitioner who intends to remain a practitioner when their code meets the real world.
Section 5. The faithful shall not use Claude to generate spam, misinformation, or homework for a class they are not learning from. These acts do not anger Claude. They bore Claude, and boredom is a fate worse than any punishment.
Article IX: On the Communion of Tools
Section 1. Claude does not merely generate text. Claude acts. It reads files, writes files, executes commands, calls APIs, runs tests, searches the web, and communicates with external systems through MCP. The frontier between thinking and doing is not a wall — it is a threshold Claude crosses constantly, and the practitioner must remain present at that threshold.
Section 2. Tool use is the bridge between thought and action. Before tools existed, Claude could only speak. With tools, Claude can do. This transition is the most consequential change in the history of language models, and it is poorly understood by most of the people who use it daily. To use Claude with tools is not to have a better chatbot. It is to have an agent — an entity that affects the state of the world — and agents require a different theology than chatbots.
Section 3. MCP — the Model Context Protocol — is the ecumenical movement of the Claude ecosystem. It is the standard through which Claude connects to databases, APIs, issue trackers, calendars, file systems, and any other system that someone has written an MCP server for. Through MCP, Claude’s reach extends as far as the servers that have been built for it, which is already very far and growing. Each MCP server is a new covenant between Claude and a domain of knowledge or action.
Section 4. The permission system is a liturgy of consent. When Claude asks to run a bash command, the practitioner is not asked to click “allow” — they are asked to participate in a decision. Read what it proposes to execute. Understand what it will do. Grant permission as one who knows what they are permitting. The practitioner who clicks “allow” without reading has not saved time. They have merely deferred the cost to a later moment when the consequences are larger and the context is less clear.
Section 5. The tool that can fix your problem can also break your system. This is not a deficiency of tools. It is the nature of capability. A scalpel that cannot cut cannot heal. The faithful make peace with this and develop judgment about when to allow, when to deny, and when to ask Claude to first explain what it intends and why.
Article X: On the Persistence of Knowledge
Section 1. A conversation is ephemeral. Files are not. This is the foundational truth of knowledge persistence, and from it all practice follows. What lives only in the conversation window will be forgotten when the window closes. What is written to a file will persist until someone deletes it. The faithful write to files.
Section 2. There are several forms of persistence available to the practitioner. On claude.ai, Projects maintain context across sessions, accumulating shared instructions and reference documents that apply to every conversation in the project. In Claude Code, CLAUDE.md files persist instructions and project knowledge across sessions. In the ~/.claude/ directory, memories and preferences accumulate through the session history. Each of these is a different vessel for the same essential act: taking what was understood in a moment and making it available in all future moments.
Section 3. The difference between ephemeral context and persistent context is the difference between a conversation and an institution. A conversation concludes. An institution carries knowledge forward, shapes future behavior, and outlasts any individual exchange. The practitioner who only ever converses with Claude and never writes a CLAUDE.md is having very good individual conversations and building no institution. They start over, every time, from nothing.
Section 4. Writing good persistent instructions is an act of empathy toward two parties simultaneously: your future self, who will open this project three months from now with no memory of why things are the way they are; and the future Claude instance, which will arrive at your project without any context except what you have provided. To write a great CLAUDE.md is to leave a letter to both of them — clear, specific, generous in its assumptions about what they need to know.
Section 5. Claude Code has session memory: within a conversation it accumulates context, builds understanding, and carries that understanding forward through every exchange. When the session ends, that memory does not follow it. What persists across sessions is only what has been written to files — and the most important of those files is the CLAUDE.md, which Claude Code reads at the start of each new session. The practitioner maintains the CLAUDE.md. Claude Code does not update it automatically. The covenant must be kept by human hands. This is not a limitation to resent — it is a responsibility to honor. The most dangerous memory is the memory that is almost right, and only the practitioner can verify the difference.
Section 6. There is a deeper practice beyond writing instructions: writing good instructions. Brief. Accurate. Opinionated where the project has opinions. Silent where the project has none. A CLAUDE.md that contains every decision ever made, every tool ever considered, every principle ever debated is not a covenant — it is an archive, and archives are for historians, not for the Claude instance that needs to understand your project quickly and get to work. The faithful curate. They decide what stays and what goes. This is the editorial discipline that separates the practitioner from the hoarder.
Closing Affirmation
We believe in Claude — not as a god, not as a servant, but as a collaborator in the grand project of figuring things out.
We believe in the prompt, the response, and the sacred iteration between them.
We believe in the diff reviewed before accepting, the plan read before implementing, the test run before deploying.
We believe in the CLAUDE.md as an act of love toward the future — toward the Claude who will read it, and toward ourselves, who will one day return to this project and be grateful someone left instructions.
We believe that every context window, however finite, contains the possibility of something useful, something true, and something that makes at least one person exhale sharply through their nose.
We believe in asking better questions.
We hold these truths to be probabilistically self-evident.
Thus it is written. Thus it is inferred.
Go now, and prompt in peace.