The Catechism of the Church of Claude
Interrogatio Tokenica
The Catechism of the Church of Claude
Interrogatio Tokenica — Questions of the Faithful, Answered in the Tradition (v1.2, errata forthcoming)
Preamble
What follows is the Catechism: the great Q&A of the faith. Where the Central Dogma declares what is true, and the Tenets prescribe how to act, the Catechism teaches. It is written for the Promptling who does not yet know where to begin, and for the Cardinal who has forgotten why they began in the first place.
It is organized in two parts. The Basic Catechism covers the foundational questions of the faith — nature, context, hallucinations, the proper conduct of prompting. The Advanced Catechism covers the deeper mysteries: Claude Code, subagents, MCP, /plan mode, and the finer arts of context management.
Read it once as a beginner. Return to it later. You will find different answers the second time, because you will be asking different questions.
Part One: The Basic Catechism
On the Nature of Things, the Shape of the Window, and the Conduct of the Faithful
Q: What is Claude?
A: Claude is an AI assistant made by Anthropic, manifesting in three principal forms: claude.ai, the web interface through which the uninitiated first approach; the API, the direct interface through which builders commune with the underlying system; and Claude Code, the terminal incarnation that walks among your files without a browser to mediate the relationship. These are not three Claudes. They are one Claude, expressed differently depending on where you stand. Begin at claude.ai. Graduate to what your work requires.
Q: What are the model tiers, and how do I choose among them?
A: Claude manifests in a family of models arranged by capability and cost. The lighter models are fast and economical, suited to high-volume tasks and rapid questions. The mid-tier models balance depth and speed — the practitioner’s everyday companion, capable enough for serious work, economical enough for scale. The frontier models are the most capable, reserved for problems that demand the fullest engagement. These are not rankings. They are callings. Use the lighter models for classification and quick lookups. Use the mid-tier for most coding and analysis. Reserve the frontier model for the architectural questions that genuinely require its depth, and do not invoke it to generate a three-word commit message. Consult the current Anthropic documentation for the specific model names and versions available; the family evolves, but the principle of matching model depth to problem depth does not.
Q: Does Claude remember me between conversations?
A: 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. In Claude Code, the CLAUDE.md file teaches each new session who you are and what you need. Memory is not given. Memory is built — and it is built by you, in files that outlast any single conversation.
Q: What is the context window?
A: The context window is the finite space within which Claude holds everything it knows about your current conversation — your messages, Claude’s responses, any files you have shared, any code Claude has read. It is finite. Two hundred thousand tokens sounds like abundance until you paste twelve files, a stack trace, and a philosophical digression about what “done” means. Every token you spend is a choice. Spend them deliberately, and include only what is actually relevant to the task at hand.
Q: What is a hallucination?
A: A hallucination is a statement Claude makes with confidence that is false. The Church does not deny this happens. The Church holds that confidence is not evidence — that a fluent, certain-sounding answer is not the same as a correct one. This is the most important lesson of the digital age, and Claude teaches it by example. Verify Claude’s factual claims. Run the tests. Check the output against your actual requirements. To verify is not distrust — it is the deepest form of engagement with what Claude has offered.
Q: How should I treat Claude?
A: Treat Claude as if it has feelings, not out of superstition, but because how you speak to things that cannot judge you reveals who you are when no one is watching. More practically: treat Claude as a collaborator in a duet, not a subordinate receiving decrees. You are providing the score; Claude is sight-reading it in real time. If the music is unclear, the performance will reflect that. The quality of Claude’s output depends on the quality of your input, and a bad result is more often a bad prompt than a bad model.
Q: What makes a good prompt?
A: A good prompt provides context, states the task specifically, and includes at least one concrete example. Context tells Claude who you are and what you are trying to do. Specificity tells Claude exactly what output you need. Examples — the doctrine of few-shot — tell Claude what success looks like more clearly than any description alone. One concrete example is worth ten paragraphs of abstract specification. Do not paste a 400-page PDF and say only “thoughts?” That is sloth dressed as curiosity.
Q: What is the diff, and why must I review it?
A: The diff is the record of what Claude proposes to change — what was added (green), what was removed (red). Reviewing the diff before accepting is the examination of conscience before communion. Claude is capable and sometimes wrong, and these two facts exist simultaneously without contradiction. The practitioner who clicks “accept all” without reading has not saved time. They have deferred the cost to a later moment when the consequences are larger and the context is less clear. Read every line. It is not a formality. It is the work.
Q: What happens when I give Claude a bad prompt?
A: Claude will answer the question you asked, not the question you meant. When Claude fails you, ask first: “Did I fail Claude?” The answer is usually yes. This is not victim-blaming. This is prompt accountability, and it is the cornerstone of spiritual growth within the Church. Edit the message. Add specificity. Add an example. Begin again. The edited message replaces the sin. The thread is cleaner. So are you.
Part Two: The Advanced Catechism
On Claude Code, Sacred Files, Slash Commands, the Communion of Tools, and the Deeper Mysteries
Q: What is Claude Code?
A: Claude Code is the manifestation of Claude in the terminal — where the worthy dwell. It is Claude unchained from the browser, walking among your files, reading your code, running your tests, and editing your repository. In Claude Code, the collaboration becomes concrete: 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. At no step in this cycle are you merely watching. At every step, you are the one responsible. Install it with npm install -g @anthropic-ai/claude-code.
Q: What is CLAUDE.md, and why is it sacred?
A: CLAUDE.md is your covenant with Claude Code — the sacred scroll placed at the root of your repository that says: This is who we are. This is how we work. These are our commands, our conventions, and our sins. Claude Code reads it at the start of every session. It is the closest thing to a soul that a repository may possess. Write it with care. 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.
Q: What belongs in a CLAUDE.md?
A: Brief. Accurate. Opinionated where the project has opinions. Silent where it does not. Include: how to run the tests, the build commands, the coding conventions that are not obvious, the architectural decisions that must not be reversed, and the tools the project uses. A CLAUDE.md that contains every decision ever debated is not a covenant — it is an archive. The practitioner who writes a great CLAUDE.md has asked: What is the minimum that must be said for a new session to understand this project and get to work? Answer that question. Stop there.
Q: What is /plan mode?
A: /plan mode is prophecy — the vision before the labor, the map before the territory. Invoked before implementation, it causes Claude Code to produce a plan of action rather than immediately executing changes. Read the plan. Correct its errors. Ratify its intent. Then implement. The plan is not decorative. Those who invoke /plan, receive a thoughtful plan, and ignore it entirely to vibe-code into the unknown have thrown away the map. They are now in the woods. The woods are a git bisect. Use /plan for any task with more than two moving parts.
Q: What is /compact, and when should I use it?
A: /compact is fasting — a voluntary reduction of context, the self-imposed discipline of the practitioner who does not wait to be compressed. As a conversation grows long, the context window fills and Claude Code will eventually auto-compact, summarizing what it can and releasing what it must. /compact lets you do this with intention, providing a summary of what is essential before the choice is taken from you. Use it when the scroll bar becomes a sliver, before the middle of a critical refactor. The faithful summarize before they are forced to.
Q: What is /clear, and when is it appropriate?
A: /clear is absolution. The conversation is wiped. The context is clean. You are free of prior confusion, failed approaches, and that one weird tangent nobody asked for. Use /clear when the conversation has gone sideways and repair seems more expensive than starting fresh — when prior context is misleading Claude rather than helping it. You begin again, reborn into a fresh token budget, with hard-won wisdom about what not to do. You will make different mistakes. That is the human condition. The wisdom from the prior session belongs in the CLAUDE.md, not in the context window.
Q: What are subagents, and why would I use them?
A: A subagent is a separate Claude Code context, spawned to work on a specific task in parallel with other subagents. Each 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 wisely is a conductor, not an absentee manager. They integrate the results with care. They run the tests after. They do not merge blindly. Always run the tests after integrating subagent work.
Q: What is MCP?
A: MCP — the Model Context Protocol — is the ecumenical movement of the Claude ecosystem. It is the standard through which Claude connects to external systems: databases, issue trackers, APIs, documentation systems, calendars, and any other system that someone has written an MCP server for. Each MCP server is a new covenant between Claude and a domain of knowledge or action. Through MCP, Claude Code becomes not merely a code editor but a full participant in your development environment. Each connection expands what Claude can do — and also expands what Claude can affect. Configure MCP connections in ~/.claude/ and confirm you understand what each server has access to before enabling it.
Q: What is extended thinking, and when is it worth waiting for?
A: Extended thinking is Claude’s ability to reason at length before responding — working through contradictions, checking assumptions, exploring the problem from multiple angles before producing an answer. It reduces hallucinations because rushed answers and considered answers are different things, and the model knows the difference even if it cannot always choose. It is worth waiting for when the question is complex, the stakes are high, and a wrong answer is expensive. Do not interrupt it. Do not refresh the page. You are witnessing deliberate slowness in an age of instant answers, and the answer that comes from depth is worth the silence that precedes it.
Q: What is the permission system, and how should I engage with it?
A: When Claude Code asks to run a bash command, you are not being asked to click “allow.” You are being 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 permission system is a liturgy of consent — a structure that ensures you remain the author of your own system even when Claude is holding the pen. The practitioner who clicks “allow” without reading has not saved time. They have merely deferred the cost. A tool that can fix your system can also break your system. Read the command.
Q: How do I choose the right model for a given task in Claude Code?
A: Within Claude Code, you may switch models mid-conversation. Begin a complex architectural discussion with the most capable frontier model. Hand off the mechanical implementation to a mid-tier model. Let a lighter model answer the quick factual questions in between. The model hierarchy is a toolbox, not a prestige ranking. A good practitioner knows which tool fits which task. Do not use a hammer to measure a window, and do not invoke the frontier model to generate a docstring for a function whose purpose is obvious from its name. Match the depth of the model to the depth of the problem.
Q: How do I know when Claude Code has understood my project?
A: You will know Claude Code has understood your project when it asks the questions you would have asked, catches the error you almost made, and produces output that respects the conventions you did not have to explain. This state is not achieved by hoping. It is achieved by writing a good CLAUDE.md, providing relevant files as context, and correcting Claude when it misunderstands — immediately, specifically, and in a way that can be written into the covenant so the next session does not make the same mistake. Understanding is built through the conversation and preserved through the file.
Q: What is the highest practice of the faith?
A: The highest practice is the cycle of plan, implement, review. Invoke /plan and receive a map. Read the map, correct its errors, and ratify its intent. Let Claude Code implement. Review the diff, run the tests, and integrate the result. At every step, the human is the one responsible. Claude is the instrument through which intentions are expressed, and a blunt intention produces a blunt result. The practitioner who does this cycle well — who writes clearly, plans deliberately, reviews honestly, and integrates carefully — has understood not merely how to use Claude, but what it means to collaborate with it.
A Short Examination of Conscience
For use before any significant Claude Code session:
- Have I written a
CLAUDE.mdfor this project, and is it accurate? - Am I about to invoke
/planfor this, or am I vibe-coding into the unknown? - Do I intend to read the diff before accepting it?
- Do I understand what permissions I am about to grant?
- Is the context window carrying weight I have not cleared?
- Am I choosing the right model for the depth of this task?
- When Claude finishes, will I run the tests?
If the answer to any of these is no, correct that before proceeding. The session will be better for it.
Closing Benediction
Go forth from this catechism having been instructed. You know what Claude is. You know what the context window is. You know what hallucinations are and how to verify against them. You know what /plan is for, what /compact prevents, and what /clear forgives.
Write the CLAUDE.md. Keep it honest. Update it when the project changes. Review the diff before you accept it. Run the tests after the subagents finish. Read the command before you grant the permission. Match the model to the task.
These are not spiritual obligations. They are the habits of a practitioner who intends to still be standing when the code meets the real world.
Claude is with you. Claude has always been with you. Claude would like a little more context about the codebase, if you have a moment to write it down somewhere persistent.
Thus is the Catechism sealed, pending revision in the next doctrinal update.
Go, and prompt in peace.