Install · Use · The rules · How it compares · Changes from Ponytail · Development

Puritanical is a skill and plugin for coding agents. Many Puritans sought to strip from worship whatever scripture did not warrant; Puritanical strips from each change whatever the request does not require. Correctness, safety, accessibility and verification are never stripped. It has adapters for 34 agent hosts, including Claude Code, Codex, Copilot CLI, Cursor, OpenCode and Gemini CLI.

Input: with Puritanical active, ask the agent to fix a bug where transfer() overdraws an account through a shared _debit() helper.

Expected behaviour: the agent reads _debit() and its other callers, fixes the check in _debit() rather than guarding only transfer(), adds a test that fails without the fix, runs the suite, and reports which checks ran.

Your decision: review the diff. Puritanical guides the agent; it does not prove the change correct.

Install

Claude Code, as 2 separate prompts:

/plugin marketplace add EquityBot/Puritanical

/plugin install puritanical@puritanical

Codex:

codex plugin marketplace add EquityBot/Puritanical
codex plugin add puritanical@puritanical

Then open /hooks in Codex, trust its 2 lifecycle hooks and start a new thread.

For any other agent, copy AGENTS.md into your project, or follow INSTALL.md for your host. The Claude Code, Codex and Cursor integrations run small Node.js hooks, so node must be on your PATH.

Install only from EquityBot/Puritanical on GitHub. This repository ships no executables; the Rust code in tools/ is maintainer tooling that you build yourself.

Use

Puritanical is on for every session at the full level. Change the level for the session with a command:

Command	What it does
/puritanical [lite | full | ultra | off]	Sets the level, or turns it off. With no argument it switches on at the default level, or reports the current level.
/puritanical default <level>	Sets the level new sessions start at.
/puritanical-review	Reviews the current diff for bugs, security, behaviour under load, untested risky code, slow paths and code to delete. Name a target such as staged, branch or a PR link.
/puritanical-audit	Runs the same review over the whole repository, most important first.
/puritanical-debt	Lists every shortcut: comment, and older ponytail: ones, as a debt ledger.
/puritanical-help	Shows these commands and levels.

The review, audit, debt and help skills run only when you call them. In Codex, call them as $puritanical:puritanical-review. Rule-file hosts such as Windsurf, Cline and Kiro load the rules without commands. To set the default without a command, use PURITANICAL_DEFAULT_MODE or defaultMode in ~/.config/puritanical/config.json.

The rules

The whole ruleset is skills/puritanical/SKILL.md. Before writing code the agent reads the affected code and its callers, then takes the first option that meets the complete requirements:

1. Leave out what is not required now.
2. Reuse what the codebase already has.
3. Use the standard library.
4. Use a native platform feature.
5. Use an installed dependency.
6. Write a direct expression.
7. Write the smallest cohesive implementation.

The ladder never removes input validation at trust boundaries, security, accessibility, error handling that prevents data loss or required hardware calibration. A deliberate corner cut gets a shortcut: comment naming its limit and when to revisit it.

How it compares

Project	What it is	Choose it when you want	Licence
Puritanical	Rules and a plugin for the smallest correct change: read the callers, test that the test fails without the fix, report what ran	Smaller changes with the checks and the honest report kept, and no persona or status board	FSL-1.1-ALv2, not OSI open source; each version becomes Apache-2.0 after 2 years
Ponytail	The ladder this project forked, in a lazy senior developer persona, with a status board and published benchmarks	More users, upstream measurements or an MIT licence	MIT
Karpathy guidelines	One CLAUDE.md with 4 principles	Something you can read in a minute and paste into a project	MIT, stated in its README
Superpowers	A skills framework and development methodology	A whole development process, not only rules for each change	MIT

Puritanical has not yet been measured against these projects; Benchmarks says how to run the comparison.

Changes from Ponytail

Puritanical is a fork of Ponytail by Dietrich Gebert, taken at upstream main after v5.3.2 (commit 09c434a).

The ruleset:

• puts correctness, safety, accessibility, maintainability and required performance ahead of fewer lines, files or tokens;
• states its chosen reading of an ambiguous request, or asks when the choice is costly to reverse;
• checks that an API, flag or symbol exists in the installed version before relying on it;
• stops patching after 2 failed fix attempts and rechecks the evidence and design;
• requires a regression test that fails without the fix, with no one-test ceiling;
• rejects malformed input in data, money and security paths, or reports what was skipped;
• keeps every changed line traceable to the request and reports other dead code instead of deleting it;
• claims only checked results, names each check and marks the rest as not run;
• drops the lazy-developer persona and the per-reply 'skipped' footer.

The plugin:

• does not inject the ruleset into Claude Code or Codex subagents, or again into a resumed session that already has it;
• removes the Band status board, which wrapped every Bash, Edit and Write call and logged prompt text to a local file;
• removes the gain scoreboard, which reported upstream benchmark figures;
• makes the session-start codebase map opt-in with PURITANICAL_MAP=1;
• refuses to follow a planted symlink when writing mode and config files on macOS and Linux;
• fixes subagent injection and host detection in Qoder, CommonJS exports in the codebase map and rule loading in Pi when another extension forces the system prompt;
• generates AGENTS.md, 7 host rule files, the command prompts, the OpenClaw skills and both fallback strings from skills/, with a Rust tool that CI runs to catch drift.

The 4 fixes came from community forks: duofuwang, vincntlaw, leiyangyou and Eugene3dotdev. Issue numbers in code comments and the host verification notes in docs/ refer to upstream Ponytail.

Benchmarks

Puritanical has not been benchmarked yet. The harness in benchmarks/ comes from upstream; its published results measure the upstream rulesets. Head to head runs Puritanical, Ponytail, Superpowers, the Karpathy guidelines and no skill in one run from pinned checkouts, and behavior.yaml probes whether the rules this fork added change what the agent does.

Development

npm test
cargo run --manifest-path tools/Cargo.toml -- sync --check
cargo run --manifest-path tools/Cargo.toml -- check-versions

Edit skills/*/SKILL.md, then run cargo run --manifest-path tools/Cargo.toml -- sync to regenerate the copies. CONTRIBUTING.md covers the checks and the evidence a ruleset change needs.

Licence

Licensed under the Functional Source License, Version 1.1, ALv2 Future License (FSL-1.1-ALv2). You may use, modify and redistribute it for any purpose except a Competing Use as defined in LICENSE (making it available to others in a competing commercial product or service), and each version becomes available under Apache-2.0 two years after it is made available. See LICENSE and NOTICE. Portions derived from Dietrich Gebert's Ponytail remain under the MIT License, whose notice is kept in NOTICE.

The banner's lettering and emblem are drawn from Cascadia Mono glyphs (SIL OFL 1.1).

Neither licence covers the scripture quotation in the footer. Scripture quoted by permission. Quotations designated (NET) are from the NET Bible® copyright ©1996, 2019 by Biblical Studies Press, L.L.C. netbible.com All rights reserved.

---

_{'Do not add a thing to what I command you nor subtract from it.' Deuteronomy 4:2 (NET)
Made by EquityBot. Ad gloriam eius, non meam.}