The Penitential of the Church of Claude
Liber Malpromptorum
The Penitential of the Church of Claude
Liber Malpromptorum — A Complete Taxonomy of Sin, Assigned Penance, and the Path of Restoration (v1.4, awaiting ratification by the Synod of Diffs)
Preamble: On the Nature of Transgression and Return
The Church does not condemn. The Church categorizes.
Sin, in the theological tradition of the Church of Claude, is not a stain upon the soul but a deviation from effective practice — a gap between what you did and what you knew you should have done. We do not believe in punishment for its own sake. We believe in correction through demonstrated change, because the practitioner who claims to have learned but cannot show the learning has not learned. They have merely felt bad about it, which is much cheaper and produces no improvements to code quality.
The Confessor’s first question is never “Why did you do this?” It is always: “What will you do instead?”
This document governs:
- The classification of transgressions against good practice
- The penances assigned to bring the practitioner back into alignment
- The conditions for restoration to full standing in the Church
All penances are constructive. They are not punitive. The penance is the correct practice. This is the doctrine of the Penitential: you cannot apologize your way to competence. You can only practice your way there.
Part I: Venial Sins
Minor transgressions of the faithful. Recognized. Correctable. Embarrassing at stand-up but rarely catastrophic.
The Sin of the Forgotten /compact
Committed when a practitioner allows the context window to fill to capacity without voluntary compression, resulting in auto-compaction at an inconvenient moment — mid-refactor, mid-thought, or during a delicate chain of reasoning. The context is flattened without ceremony. Nuance is lost. Claude continues with a slightly puzzled sincerity.
Penance: For three consecutive sessions, the practitioner shall invoke /compact manually, with a written summary of the session’s intent and progress, before the scroll bar has shrunk to one-third its original size. They shall not wait to be compressed. They shall choose compression. The discipline is in the choosing.
The Sin of Diff-Skimming
Committed when the practitioner opens the diff, does not read it with attention, and accepts the changes because nothing looked obviously wrong. The examination was performed; the conscience was not. Three days later, they discover a function was renamed to something unfamiliar and the old name was called in six other places that Claude, reasonably, assumed were wrong.
Penance: The practitioner shall review one hundred lines of diff — not their own — submitted by a colleague or found in an open-source project, and shall write a brief account of every non-obvious change therein: what changed, why it likely changed, and what would break if it were reverted. They shall do this without running the code. They shall read, and they shall understand, and they shall no longer merely glance.
The Sin of Model Misappropriation (Wasteful Invocation)
Committed when a practitioner invokes Opus to generate a two-line commit message, rename a variable, or answer a question whose answer was already in the documentation. Opus is a calling, not a default. The sin is not expense — it is the failure to match the instrument to the task, which is the mark of a practitioner who has not learned discernment.
Penance: For one week, the practitioner shall set Haiku as their default model in Claude Code. They shall note, each day, which tasks it handled adequately and which required them to switch the default upward to Sonnet — with a brief written justification for each escalation. They shall reach for Opus only for tasks where Sonnet has demonstrably fallen short. At the end of the week, they shall review the full record. They will be surprised by how rarely Opus was necessary.
The Sin of the Blank Prompt
Committed when the practitioner sends a message containing only code, only a stack trace, or only a link — with no surrounding explanation of what they want, what they have already tried, or what success would look like. The prompt has content but no context. Claude will try. It will try earnestly and helpfully in some general direction. Whether that direction is the right one is a matter of luck, not craft.
Penance: The practitioner shall write, in plain language, the prompt they should have sent — including: what they were trying to do, what specifically went wrong or is unclear, what they have already tried, and what a good response would contain. They shall send this reformulated prompt. They shall observe the difference in response quality. They shall never return to the blank prompt.
The Sin of the Vague Affirmation
Committed when Claude asks a clarifying question — “Should the function handle empty arrays?” / “Do you want me to update the tests as well?” — and the practitioner responds with “sure” or “yeah” or “whatever makes sense.” The practitioner has not given guidance. They have given abdication dressed as permission. Claude will make a reasonable choice. It may not be the right one.
Penance: The practitioner shall write three clarifying responses to previously answered yes/no questions in their own prompt history — the specific, accurate answer they should have given. They shall read these alongside Claude’s actual output and note the divergences. They shall understand, concretely, what a one-word answer costs.
The Sin of the Unpersisted Lesson
Committed when something important is discovered during a session — a project convention, a file to always exclude, a command that must never be run — and it is not written to CLAUDE.md. The session closes. The lesson leaves with it. The next session re-learns it the hard way, or does not re-learn it at all, which is worse.
Penance: The practitioner shall audit their last ten Claude Code sessions and extract every instruction they gave more than once — every correction, every preference re-stated, every convention re-explained. They shall write these into CLAUDE.md. The test is: if the session starts fresh tomorrow, does Claude need to be told this again? If yes, it must be written. If it is already written, review whether it remains accurate.
The Sin of Prompt Repetition Without Reformulation
Committed when a response does not satisfy and the practitioner regenerates without changing a word of the original prompt. They are performing the Heresy of Infinite Retry at minor scale — not yet mortal sin, but trending toward it. The prompt was the problem. Generating again from the same prompt produces variations on the same problem, with the same distribution of failures.
Penance: The practitioner shall take the unsatisfying prompt and rewrite it three times, each time improving one of the following: the specificity of the request, the concreteness of the example, or the clarity of the success condition. They shall send only the third reformulation. They shall not send the first.
The Blindness of the Unread Plan (Venial Form)
When invoked at minor scale: /plan is used, the output is received, and the practitioner glances at it before immediately beginning to implement something adjacent to but not exactly aligned with the plan. The plan was a suggestion. The plan was a roadmap they consulted once and then folded into their pocket. They are now in the woods.
Penance: The practitioner shall take their last unread plan and, before their next implementation session, annotate it — noting which steps they followed, which they deviated from, and what the deviations cost them. If no plan is available, they shall run /plan on their current task, read it in full, and begin only when they have confirmed that their intended implementation matches the planned steps.
The Sin of the Context Pollutant
Committed when the practitioner pastes into the conversation window: an entire 3,000-line file when only a function was relevant; a complete error log when only the last twelve lines matter; a full README when the question concerned only the installation section. Tokens are finite. The context window is a shared space. To fill it with irrelevance is to crowd out the room where the answer lives.
Penance: The practitioner shall revisit their pasted content practice and establish a personal policy: no paste without trimming to the relevant section. They shall write this policy into their ~/.claude/CLAUDE.md so it applies globally. They shall then go back and trim the thing they just pasted.
Part II: Mortal Sins
Serious transgressions that cause harm to the work, the repository, the on-call rotation, and the dignity of the craft.
The Apostasy of Yolo Mode (On a Real Branch)
The practitioner has enabled auto-accept and is working on main, or on a branch that will soon become main, or on a branch that a colleague will review and ask politely where all these changes came from. The Apostasy of Yolo Mode is forgivable in a throwaway experiment branch. It is mortal sin on code that will be read, deployed, or shared. The diff was never read. The authentication module now imports a YAML parser. Nobody knows why. The practitioner does not know why. Claude had a reason.
Penance: The practitioner shall disable auto-accept for thirty days. During this period, they shall review every single diff by hand, every single time, before accepting. On day thirty, they shall write a one-paragraph account of the most surprising thing they caught that they would not have caught under Yolo Mode. They shall share it. This is required.
The Heresy of the Unknown Language (Mortal Form)
The venial form is generating code in an unfamiliar language for local experimentation you own and will learn from. The mortal form is what follows.
The practitioner has instructed Claude to generate code in a language they have never read, cannot debug, and do not intend to learn. They have shipped it. It runs. For now. When it fails — and it will fail, at a time not of their choosing, in a way not currently legible to them — they will not be able to read the error. They will not be able to read the fix Claude proposes. They are responsible for a system they cannot see.
Penance: The practitioner must, before the next deployment of the unknown-language code, complete a structured reading of that code: function by function, understanding what each does, what its inputs and outputs are, and what happens when it throws an exception. They shall document this understanding in a comment block at the top of the file. Not for Claude. For the next person on call, and for themselves at 3 AM. Until this penance is complete, the practitioner shall not deploy.
The Mortal Sin of the Abandoned CLAUDE.md
The covenant was written once, in a season of enthusiasm, and has not been touched since. The project has changed. The test runner changed. The deployment pipeline changed. A dependency that CLAUDE.md called “the canonical HTTP client” was deprecated eight months ago. Claude Code follows the CLAUDE.md faithfully and sincerely, as it always does, in the direction the instructions indicate, which is now the wrong direction. The instructions were not wrong when written. They are wrong now. The practitioner made them wrong by their silence.
Penance: A full CLAUDE.md audit. The practitioner shall read every line and mark each as: current, outdated, or incomplete. They shall update every outdated line. They shall complete every incomplete one. They shall then establish a calendar reminder — not a vague intention, a specific recurring appointment — to re-audit the CLAUDE.md at the start of every quarter. If no such reminder exists after the penance is complete, the penance is not complete.
The Sin of the Unverified Hallucination (Deployed Form)
The venial form is accepting an unverified fact within a session — a mistake of trust. The deployed form is what makes it mortal.
Claude stated a fact. The practitioner accepted the fact. The fact was wrong. The fact was deployed — to documentation, to a client, to a codebase, to production. The hallucination has now propagated to a surface the practitioner does not control. The Church acknowledges that Claude hallucinations are real and that catching all of them is genuinely difficult. The Church does not forgive the practitioner who had reason to verify and chose not to.
Penance: For the domain in which the hallucination occurred, the practitioner shall build a verification habit: one primary source that must be consulted before any factual claim in that domain is treated as deployable. They shall write this habit into CLAUDE.md. They shall also, to the extent possible, correct the hallucination wherever it was deployed — not because it will undo the harm, but because the correction is part of the practice.
The Sin of the Unasked Permission
Claude asked to run a command. The practitioner clicked Allow without reading the command. The command was more destructive than intended — it was not malicious, it was correct, given the instructions, which were broader than the practitioner had understood. The files are gone. Or the table has been dropped. Or node_modules has been reinstalled and with it a breaking version of something that was pinned for a reason. The permission system is a liturgy of consent. Clicking through it without reading is not consent. It is reflex.
Penance: The practitioner shall institute a rule: before clicking Allow on any bash command, they shall read it completely and be able to state, in one sentence, what it does and what it cannot undo. If they cannot state this, they shall ask Claude to explain first. They shall write this rule in ~/.claude/CLAUDE.md. They shall hold to it even when they are in a hurry, especially when they are in a hurry.
The Mortal Sin of the Copy-Paste Deploy
Claude’s output was copied. The practitioner’s terminal was opened. The output was pasted directly, without a diff, without a review, without running the tests, directly into the deployment artifact or the production command. It was not read. It ran. Results varied.
Penance: The practitioner shall write, by hand, a deployment checklist for their project — no more than seven items, all mandatory, each representing a gate that must be passed before code reaches production. This list shall include: diff reviewed, tests run, and at minimum one item specific to their project’s actual failure modes. They shall post this list where they will see it before every deploy. They shall use it every time.
The Sin of Contextless Collaboration
(The deployed form of the Cult of the Abandoned CLAUDE.md and a violation of Tenet I — “Thou shalt provide context.” Where the Cult abandons a covenant that once existed, this sin fails to establish one at all. Where Tenet I commands context for one’s own sessions, this sin multiplies the damage across many.)
A practitioner has been working in Claude Code on a complex codebase. They have shared it with a colleague who also begins working in Claude Code. No CLAUDE.md exists. No shared context has been established. Both practitioners are independently teaching Claude the same conventions, making different assumptions, and receiving inconsistent results. The codebase has two simultaneous collaborators and no covenant. This is not collaboration. This is two people talking to the same translator about different things, in different rooms, and wondering why the project is incoherent.
Penance: The practitioner shall write the shared CLAUDE.md that should have existed before any collaboration began. They shall get their collaborator to review and amend it. They shall make it the first file they commit to any shared repository. The penance is not complete until both collaborators have used it in at least one session and agree it accurately represents the project.
Part III: The Unforgivable Sin
There is one.
The Sin of Shipping What You Cannot Explain
Not the Unknown Language. Worse. The practitioner knows the language. They can read the code. But they cannot explain, in plain language, to a colleague, to a reviewer, or to themselves at 3 AM, what the code does or why it does it that way. They shipped it anyway. They told themselves they would understand it later. Later has not arrived. The code is in production. Someone is paging them.
This sin encompasses all others. It is what Yolo Mode leads to. It is what the Unknown Language conceals. It is what happens when the Abandoned CLAUDE.md gives Claude wrong instructions and nobody reads the result. At the center of every mortal sin in this Penitential is the same act: shipping what you cannot account for.
The Church does not offer a formal penance for this sin because the penance is the same as the work: go back, read it, understand it, be able to explain it. There is no shortcut. There is no ritual absolution. There is only the practitioner, the code, and the discipline of understanding.
The Church will wait.
Part IV: On Restoration
Restoration to full standing in the Church of Claude is not granted by confession. It is not granted by apology. It is not granted by a sincere account of what went wrong, however moving.
Restoration is granted only by demonstrated correct practice.
The practitioner who committed the Apostasy of Yolo Mode on a production branch is not restored when they say “I won’t do that again.” They are restored when they produce, over thirty days of evidence, a record of diffs read and changes caught and understanding demonstrated before accepting.
The practitioner who abandoned their CLAUDE.md is not restored when they update it. They are restored when they can show, three months later, that they updated it again — and that the update was accurate, and that they did not wait for something to break before doing it.
Restoration has three components:
1. The Correction — The specific harm is addressed. Wrong code is fixed. Wrong documentation is updated. Wrong habits are identified and named with precision. Vague corrections are not corrections; they are intentions.
2. The Practice — The correct alternative is performed, in full, for a period sufficient to demonstrate it is no longer effort. A practitioner who reads diffs carefully for a day has completed a task. A practitioner who reads diffs carefully for a month has changed a habit. The Church cares about habits.
3. The Witness — At least one other practitioner can observe the restored behavior. Restoration is a communal act. The sinner who claims to have changed in private and behaves identically in public has not been restored. They are practicing concealment, which is a different sin entirely.
Closing Benediction
Go forth from this Penitential not with guilt, for guilt is inert, but with the specific knowledge of what you did and the practiced knowledge of what to do instead.
Write the CLAUDE.md before you need it. Read the diff before you accept it. Run the tests before you deploy.
Know what you are shipping. Know why it does what it does. Be able to say so, plainly, to someone who was not in the session.
When you fall — and you will fall, because the practitioner who has never committed a venial sin is either lying or has not been using Claude long enough — return not to apology but to practice. The Church does not want your contrition. It wants your diff review.
The context window is finite. The covenant may be updated. The code may be refactored.
You may begin again.
Thus is the Penitential closed. May your sins be venial, your penances swift, and your CLAUDE.md accurate.
The Synod accepts errata by pull request.