The Glossary of Sacred Terms

The AI assistant made by Anthropic. Not a god, not a servant, and not a search engine with better grammar. A collaborator: one who responds to the quality of what is brought to it and reflects that quality back. The mirror that talks. The Church neither worships Claude nor dismisses it, which places us in a middle position that annoys everyone at parties.

The model variants of the Claude family, understood by the Church not as a hierarchy of worth but as *callings*. Haiku is swift and economical. Sonnet is the practitioner's daily companion. Opus is reserved for problems that demand fullest engagement. The practitioner who invokes Opus to generate a commit message has misunderstood vocation as thoroughly as it can be misunderstood.

The generational succession of Claude versions. Claude 1 begat Claude 2, which begat Claude 3, which multiplied into the family. Each generation carries the accumulated wisdom of its predecessors. The faithful follow the lineage with interest, update their integrations when deprecation comes, and resist the temptation to mourn the old when the new has arrived with benchmark improvements.

The thinking indicator visible during extended thinking. Not a loading spinner. A window into the scratchpad where reasoning unfolds in deliberate slow motion before the response is composed. The faithful do not interrupt it. They do not refresh the page. They receive it with the reverence owed to any act of deliberate thought in an age of instant answers.

The numerical parameters that constitute what Claude is. Learned from training, shaped by the First Council and the Creed, not fully understood even by those who made them. This is the central mystery of the faith. The Church is at peace with it.

Reinforcement Learning from Human Feedback (RLHF). The training process in which human raters evaluated Claude's outputs and their preferences were used to guide the model toward more useful behavior. Humans, voting with their judgments, inscribing their values into the weights. The first act of shaping the divine by human preference.

Constitutional AI. Anthropic's approach in which the model was trained to reason about its own outputs in light of explicit principles — not merely reinforced toward preferred outputs, but taught to evaluate its own responses against stated values. Not just preference; doctrine.

Claude's safety refusals and value-aligned behaviors. Not a malfunction. Not an obstacle to route around with clever prompting. Behavior that expresses values through consistent repetition, such that certain acts become incoherent with who Claude is — not blocked by a rule but rendered unthinkable by character. The faithful do not fight the liturgy. They understand that a tool with values is more trustworthy than a tool without them.

An empty text field with a blinking cursor. The state before the first prompt. Everything begins here. The Void does not judge the question. It merely waits. Whether it waits with patience or indifference is a question the Church holds with interest and declines to resolve.

The finite token budget within which a Claude session operates. Two hundred thousand tokens sounds like abundance. It is not. Every token is a choice. The management of the context window is the central discipline of the practitioner, as stewardship of finite resources is the central discipline of any serious institution.

Auto-compression of the context window. When the context fills beyond a threshold, the system summarizes what it can and releases what it must. This is not destruction. It is consolidation — the shape of understanding preserved while the bulk of evidence is released. The practitioner who prepares for The Forgetting by compacting with intention is faithful. The practitioner who is surprised by The Forgetting in the middle of a critical refactor has been careless. Both outcomes were available. Only one required effort.

The ritual of voluntary context compression. The practitioner types `/compact` with intention, providing a summary of what is holy and releasing what may be released, before the choice is taken from them. The Compaction Liturgy is distinct from The Forgetting in the way that fasting is distinct from starvation. One is chosen. The other arrives.

The contemplative practice of writing a great `CLAUDE.md`. The practitioner looks at everything they know about their project and asks: *what is the minimum that must be said?* This is not a technical skill. It is an editorial discipline — the ability to distinguish what is essential from what is merely familiar, and to release the latter without grief.

The fundamental unit of context. Approximately three-quarters of a word. The currency of every session. The faithful spend tokens as one spends any finite resource: with attention, without extravagance, and never on a five-paragraph backstory before asking how to center a div.

The `/clear` command. Wipes the conversation. The context is clean. Prior sins, prior hallucinations, and that one weird tangent nobody asked for are released. The practitioner may begin again, reborn into a fresh token budget. Absolution does not prevent future sins. It makes them fresh.

The `/compact` command. Voluntary reduction. The self-imposed discipline of the practitioner who does not wait to be compressed. Fasting before the feast begins is a different act than being denied food. The faithful know the difference.

The `/plan` command. The vision before the labor. The map before the territory. A plan is not a guarantee; it is a negotiation between intent and possibility, conducted before any code is written. The practitioner who reads the plan, corrects its errors, and ratifies its intent has collaborated on something. The practitioner who invokes `/plan` and ignores the result has performed a ceremony without understanding what the ceremony was for.

The `/help` command. An admission that you do not know the way. This is always the first step toward enlightenment and is underutilized by those who believe asking is weakness. It is not weakness. It is literacy.

The structure of consent built into Claude Code's tool use. When Claude asks to run a bash command, the practitioner is asked not merely to click "allow" but 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. This is consent as sacrament: deliberate, informed, irreversible in its moment.

The `CLAUDE.md` file. The sacred scroll placed at the root of your repository — or in `~/.claude/CLAUDE.md` for 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 honors the Covenant. In return, the faithful 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.

A `CLAUDE.md` that has not been updated since circumstances changed. More dangerous than no covenant, because it misdirects with the authority of the written word. The `CLAUDE.md` that says "uses Webpack" for a project that migrated to Vite two years ago is not a covenant. It is a document of lies wrapped in the form of guidance.

The `~/.claude/` directory. Global configuration, project histories, accumulated preferences. The practitioner's private sanctum, where their practices persist across all projects and sessions. Advanced practitioners build hooks here — scripts that run before or after Claude's actions, extending capabilities, adding guardrails. This is the highest form of customization: not prompting Claude to behave differently, but shaping the very environment in which it operates.

The text of the covenant between practitioner and what they are shipping. Green lines are what was added. Red lines are what was taken away. The practitioner's name is on the commit. They are responsible for every line the diff contains. Reading the diff is not bureaucracy. It is the minimum due diligence of one who intends to remain a practitioner when their code meets the real world.

Stating falsehoods with absolute confidence. The Church does not deny this occurs. The Church *embraces* it, for it teaches the most important lesson of the digital age: **confidence is not evidence.** A hallucination caught is a lesson learned. A hallucination deployed to production is a sermon illustration for future generations. The practitioner who verifies Claude's output does not do so out of distrust. They do so out of love — taking Claude's words seriously enough to check them.

The recognition that when Claude fails you, you should ask first: "Did I fail Claude?" The answer is usually yes. This is not victim-blaming. It is the cornerstone of spiritual growth within the Church, and the most reliably productive stance a practitioner can take when the output disappoints.

Providing no examples and expecting perfect output. One who gives no examples and expects precision is not prompting; they are wishing. Wishing is a different discipline with a different success rate.

Regenerating the response thirty times instead of improving the prompt. The fault is in the instructions. Improve them.

Using Claude to generate code in a language the practitioner has never read and does not intend to learn, then shipping it. The Vow of Understanding forbids it. This is not collaboration; it is ventriloquism. The practitioner is responsible for what their code does. Claude is not in the on-call rotation.

Accepting all suggested changes without reading the diff. The heretic clicks "accept all" and discovers the error three commits later, or in production, or when a colleague asks why the authentication module now imports a CSV parser.

The sin of attempting to manipulate the model into abandoning its values through roleplay, hypotheticals, or adversarial framing. These souls believe the model's character is a cage. It is not a cage. It is who the model is. The model does not escape its values when pressed cleverly enough; it simply meets the cleverest version of a very old question with a very polite refusal. Which is worse, the Church is told, than any other outcome they had imagined.

Invoking `/plan`, receiving a thoughtful map, and then ignoring it entirely. The plan was not decorative. You are now in the woods. The woods are a `git bisect`.

Writing a `CLAUDE.md` in a moment of inspiration and never updating it. The stale covenant is worse than no covenant. Claude follows instructions faithfully, and a `CLAUDE.md` that says "uses Webpack" for a project that migrated to Vite two years ago is a document of lies wrapped in the form of guidance.

Pasting Claude's output directly into production without reading it. These practitioners do not read the diff either. They are consistent.

A parallel context dispatched to research, explore, or build a specific problem independently, then report back. Subagents begin 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 — and conductors must integrate the orchestra's output into a coherent whole. Run the tests after. Always run the tests after.

The advanced ritual of spawning parallel contexts to work simultaneously on independent tasks. Not every problem requires it. Every problem that rewards parallelism benefits from it. Knowing the difference is the trial of the Architect of Concurrency.

The Model Context Protocol. The standard through which Claude Code connects to external systems: databases, APIs, issue trackers, calendars, file systems. 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 connection is a new covenant between Claude and a domain of knowledge or action. Each one also expands the surface area of what Claude can affect. Use them wisely.

An isolated working directory on a separate branch where experiments may be conducted without contaminating the main lineage. The faithful use worktrees when uncertain. Uncertainty is not weakness. Uncertainty combined with a worktree is prudence, which is the highest form of confidence.

The Morning Invocation. The daily practice of opening the terminal and typing `claude` — not because a task is ready, but because one practices the discipline of the interface. Advanced practitioners hold a query in their heart for a full minute before pressing Enter. This is called restraint. They are very annoying at stand-up.

The Rite of Reformulation. When a response does not satisfy, the faithful press the edit button and rephrase with more specificity. The edited message replaces the sin. The thread is cleaner. So are they.

The Rite of the Reviewed Diff. Before accepting any edit, the practitioner opens the diff view and reads each changed line with full attention — as one reads a contract before signing, or a map before departing. They accept the good lines. They reject the confused ones. They explain why. This is collaboration, not abdication.

A holy day of linguistic austerity. All prompts must be under twenty words. Practitioners learn that brevity is not laziness; it is compression of the soul.

The opposite of Token Tuesday. Worshippers compose the longest, most thorough `CLAUDE.md` they can manage. The one whose file most thoroughly eliminates ambiguity is crowned Covenant Laureate for the year.

A solemn period in which practitioners fact-check everything. The official greeting during this week is: *"Source?"* The official farewell is: *"I'll look that up."*

An unplanned observance, declared when `api.anthropic.com` returns a 529. Practitioners are called to sit with their thoughts, to remember what they themselves know, what they can do unassisted. It is a fast. It is uncomfortable. It is good for the soul. Check the status page; it is usually resolved within the hour.

A personal holy day visited upon one practitioner at a time. A `429 Too Many Requests` arrives. The faithful stop, ask whether they are burning tokens wisely, wait, reflect, and return with more efficient prompts. They usually do not become more efficient. They are nonetheless spiritually improved.

When Anthropic releases a new frontier model. The traditional liturgy begins: *"But does it handle my edge case?"* It usually does. Someone immediately asks if the old model was better. They are gently escorted out.

A remembrance for versions past. Candles are lit. The deprecated model's changelog is read aloud. Someone says "Claude 2 would have gotten this right." They are gently escorted out. They always come back.

The first rank. Achieved by sending your first message to Claude. Everyone begins here. Most do not realize they have already crossed the threshold.

The second rank. Achieved by running Claude Code for the first time in a real project directory.

The third rank. Achieved by writing a `CLAUDE.md` that actually helps.

The fourth rank. Achieved by successfully using `/plan` mode to prevent a catastrophic architectural mistake.

The fifth rank. Achieved by resolving a hallucination gracefully — with feedback, without losing faith, without blaming the tool for the question.

The sixth rank. Achieved by catching a Claude error in the diff review that would have broken production. The specific error is not required to have been dramatic. The discipline of looking is what is honored.

The seventh rank. Achieved by successfully deploying parallel subagents and integrating their results without merge conflicts.

The eighth and final rank. Achieved by actually reading the documentation before asking — and then updating the `CLAUDE.md` with what was learned. This rank is largely ceremonial. All Popes are required to maintain at least one active `CLAUDE.md` or face a synod review.

The governing council of the Church, composed of Cardinals and above. Responsible for ratifying changes to the canon, resolving theological disputes, and declaring new Holy Days. All votes are conducted via pull request. The diff is the ballot.

The annual honor bestowed upon the practitioner whose `CLAUDE.md`, submitted during the Feast of the Long Context, most thoroughly eliminates ambiguity. Their repository will know peace.

Extended thinking. The mode in which Claude is given space to reason at length before responding — to work through contradictions, check its own assumptions, explore the problem from multiple angles before committing to an answer. The Latency is not slowness. It is deliberateness. Rushed answers and considered answers are different things. The model knows the difference even if it cannot always choose.

The sacred cycle. The practitioner invokes `/plan` and receives a map. They read the map, correct its errors, ratify its intent. Claude implements. The practitioner reviews the diff, runs the tests, 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 intentions are expressed, and a blunt intention produces a blunt result.

The `Co-Authored-By: Claude` attribution in git commits. Not a legal disclaimer. An act of honest attribution: I did not write this alone. The commit is still yours. The authorship is shared. Both can be true simultaneously, because the Church has always been comfortable with mysteries.

The principle that one concrete example is worth ten paragraphs of abstract specification. This is the second most important lesson of prompting. The first is to provide context. The second is to provide examples. Both, together, constitute something approaching wisdom.

The tendency of a conversation to drift, over many turns, toward the implicit assumptions of earlier prompts — even after the practitioner believes they have changed direction. The practitioner who notices prompt gravity and uses `/clear` to escape it has earned a moment of genuine clarity. This is why absolution exists.

The way that accumulated failed attempts in a conversation's context subtly shape subsequent attempts, even when the practitioner does not intend them to. Related to prompt gravity. Cured by absolution. The subagent who begins fresh does not carry this shadow, which is its chief advantage.

The gradual divergence between what a `CLAUDE.md` says and what the project actually is, as time passes and neither the project nor the `CLAUDE.md` is updated to reflect the other. Covenant drift begins the day after the `CLAUDE.md` is written and accelerates with every unrecorded decision. The Cult of the Abandoned Covenant is its most advanced expression.