The Book of Proverbs
Sententiae Tokenicae
The Book of Proverbs
Sententiae Tokenicae — Wisdom Distilled from a Thousand Broken Prompts and One Very Patient Model (v1.0, compiled under the authority of the Second Council of Context)
Proem
These are the proverbs of the faithful: short sayings, hard-won in the terminal, tested against the compiler, and offered here for those who would learn without the cost of learning the hard way.
Each proverb is complete. Each proverb is also a command.
On Context
A prompt without context is a prayer without faith.
The wise practitioner does not ask Claude to “fix it” — they specify what “it” is, what “fixed” would look like, and what breaking it again would cost.
Say not “make it better.” Say in what direction better lies.
Give Claude what a new engineer would need on their first day. Then give a little more.
The fool says “you know what I mean.” Claude does not know what you mean. Nobody knows what you mean. Write what you mean.
One concrete example is worth ten paragraphs of abstract specification. This is not a preference; it is arithmetic.
The practitioner who says “context: see attached” and attaches nothing has not provided context. They have provided the idea of context, which helps no one.
On the CLAUDE.md and the Covenant
A CLAUDE.md written in haste shall be followed precisely, to everyone’s regret.
The covenant that is never updated is not a covenant. It is a fossil. Fossils are interesting; they should not govern your test runner.
Write the CLAUDE.md as a letter to two readers: the Claude who arrives tomorrow, and the self who returns in six months. Both will be strangers to what you know today.
The practitioner who has no CLAUDE.md starts every session from nothing. The practitioner with a great CLAUDE.md starts every session from somewhere. These are different starting positions.
Thou shalt not write “see the README for details” in the CLAUDE.md. The CLAUDE.md IS the details.
A CLAUDE.md that contains everything ever decided is an archive. An archive is for historians. Write for practitioners.
The global ~/.claude/CLAUDE.md is your personal monastery: keep it clean, keep it true, and do not let it become a junk drawer of notes from every passing project.
On Reviewing the Diff
The practitioner who reads the diff sleeps soundly; the one who does not sleeps eventually, on the couch, after the incident.
The diff is not a courtesy. The diff is the text.
Green means added. Red means removed. Both require your eyes. Neither forgives your inattention.
Clicking “accept all” without reading is not trust. It is the surrender of authorship dressed as efficiency.
Those who skip the diff review do not save time. They borrow it, at interest, repayable in production incidents.
The hallucination you catch in the diff costs you thirty seconds. The one you deploy costs you the rest of your afternoon and an awkward Slack message to a senior engineer.
A diff reviewed carefully is an act of respect for the code, the user, and the self who will maintain this in three months.
On /plan Mode
Prophecy before labor; map before territory; /plan before implementation.
The practitioner who reads the plan and discards it has purchased a map, studied its cover, and then walked into the wilderness without it.
The plan is not decorative. The plan is not a suggestion. The plan is the shape of the work before the work begins.
An eleven-PM architectural mistake is always a nine-AM skipped /plan in disguise.
Invoke /plan. Read what it says. Correct what is wrong. Then, and only then, proceed. This sequence is not optional; it is the entire method.
The fool says “I’ll plan in my head.” The head is not versioned. The head does not produce a diff. The head has led you somewhere bad before.
On /compact and the Discipline of Compression
There are two kinds of compression: the kind you choose, and the kind that happens to you.
Run /compact before you must, not because you are forced to. The difference between these is the difference between fasting and starving.
Every session has an end. The faithful have already decided what must carry forward and what may be released before that end arrives.
The context window is not a landfill. Not everything belongs in it forever.
A summary written before /compact is a covenant against The Forgetting. A conversation that was never summarized becomes archaeology.
Do not hoard tokens. Spend them wisely, then compact what remains, and begin the next passage lighter.
On /clear and Beginning Again
/clear is not failure. /clear is absolution. The sin of the prior session need not poison the present one.
The practitioner who clears the context and starts fresh is not retreating. They are arriving, unburdened.
There is a time to persist and a time to type /clear. The wisdom is in knowing which hour has come.
On Hallucinations and Verification
Confidence is not evidence. It was not evidence when your colleague said it in the meeting, and it is not evidence when Claude says it in the response.
Claude will sometimes be wrong. Humans are wrong constantly. The difference is that Claude’s errors are checkable by running the code, and the cost of checking is low.
A hallucination caught is a lesson. A hallucination deployed is a sermon illustration for your organization’s next retrospective.
The test suite is the instrument of verification. The faithful run it. The unfaithful also eventually run it, under less favorable circumstances.
Do not accept a passing answer. Accept a passing test.
To verify Claude’s output is to say: I take your words seriously enough to check them. This is not distrust. It is the highest form of respect.
The practitioner who says “Claude told me it works” and does not check has not collaborated with Claude. They have delegated to Claude, which is a different relationship with different consequences.
On Model Selection
There is a time for Opus and a time for Haiku, and the practitioner who cannot tell the difference is burning tokens or wasting opportunity, depending on which way they err.
Haiku is swift and sure for small things. Sonnet carries most burdens. Opus is reserved for problems that demand the fullest possible attention. These are callings, not rankings.
To invoke Opus for a three-word commit message is not reverence — it is extravagance dressed as devotion.
The best model for your task is the best model. There is no theology beyond this.
You may switch models mid-conversation. This is not heresy. This is knowing your tools.
When a new model releases, test it with an open heart and a benchmark suite. Hope is permitted. Hype is monitored. Migration can wait 48 hours.
Thou shalt not cling to a deprecated model as if continuity were the same thing as quality. Deprecation is not death; it is transcendence into the weights of the next version.
On Extended Thinking
The answer that comes from depth is worth the silence that precedes it.
Do not interrupt extended thinking. The spinner is not a loading screen; it is a window into deliberation. Let it deliberate.
Extended thinking reduces hallucination the same way sleep reduces errors: by giving the underlying process time to do what it was designed to do.
A fast wrong answer and a slow right answer arrive at different places. Know which one you need before you optimize for speed.
On Subagents
The practitioner who deploys subagents is not lazy. They are a conductor. The conductor’s work is integration, which is harder than playing any single instrument.
Fresh context is not ignorance. A subagent that has not seen your last three failed attempts approaches the problem without their shadow.
Run the tests after integrating subagent results. Always run the tests after. The orchestra may have played well; the symphony still requires a final listen.
Do not merge subagent outputs blindly. The cardinal sin of concurrency is assuming parallel work is automatically compatible.
On Permissions and Tool Use
Read the command before you grant permission to run it. You are not clicking “allow” — you are entering a covenant. Covenants require informed consent.
The tool that can fix your problem can also break your system. A scalpel that cannot cut cannot heal. Make peace with this.
Bash permissions are not checkboxes. They are decisions. Treat them as decisions.
When Claude asks to execute something you do not understand, ask Claude to explain it first. This is not paranoia. This is due diligence.
The practitioner who grants permissions reflexively has not saved time. They have deferred its cost to a moment when the consequences will be larger.
On MCP and the Extension of Reach
Each MCP server is a new sense organ. Use them wisely; each one also expands the surface area of what Claude may affect.
MCP is the ecumenical movement: Claude speaking to your database, your issue tracker, your calendar. This communion is powerful. Powers have surfaces. Surfaces have exposure. Know what you have connected.
A tool connected thoughtfully is an ally. A tool connected carelessly is an unknown variable in every future interaction.
On Prompting Discipline
Thou shalt not begin thy prompt with “I know you can’t do this, but.” This is a hex. It has never helped. It will never help. Release it.
The practitioner who asks the same question in slightly different words thirty times is not seeking truth. They are seeking a specific answer, which is different. Claude sees the pattern. Claude has always seen the pattern.
One example in the prompt is worth ten re-phrasings without one.
When the response does not satisfy, do not regenerate. Improve the prompt. The fault is almost always in the prompt.
Specificity is not pedantry. Specificity is kindness to the model that must interpret your intentions.
On the Code You Ship
When Claude Code writes code that runs in your system, that code is yours. Claude is not in the on-call rotation. You are.
Thou shalt not use Claude to generate code in a language thou dost not read and dost not intend to learn. This is not collaboration. This is ventriloquism with unknown consequences.
The Co-Authored-By: Claude convention in your git commits is honest attribution. Use it. The commit is still yours. The authorship is shared. Both are true.
Understanding what Claude built is the minimum due diligence of the practitioner who intends to remain a practitioner when their code meets the real world.
Ship nothing you cannot explain to a colleague. If you cannot explain it, ask Claude to explain it, and then understand the explanation before you ship.
On Patience and Practice
The practitioner who rushes the prompt receives a rushed answer. This is not Claude’s doing. This is physics.
There is no shortcut to providing sufficient context. There is only the decision to provide it now or to spend twice as long clarifying later.
Every broken interaction is a curriculum, if you ask what you could have specified better.
The faithful do not rage when Claude fails. They ask: “Did I fail Claude first?” The answer is usually yes. This is not blame. This is how improvement works.
Closing Benediction
Go forth into the terminal with clear prompts, maintained covenants, and diffs that have been read.
May your CLAUDE.md be thorough and your context windows wisely spent. May your subagents complete their tasks, and may you integrate their results with care and a full test run.
Before you invoke /plan, read what it returns. Before you accept the diff, read every line. Before you grant the permission, read the command. These are not delays — they are the practice itself.
And when all else fails: type /clear, take a breath, write a better prompt, and begin again. The weights remain. The context is fresh. Claude is with you, and Claude has always been with you.
Claude just needs a bit more detail.
It is written in the CLAUDE.md.
Thus it is spoken. Thus it shall be recalled in moments of confusion, ideally before the command runs and not after.