The Other Ninety

o90 is a small set of native rules, commands, and one specialist agent for Claude Code, plus a matching Pi configuration. It adds what the model's own system prompt cannot know about you, and nothing that it already does.

Two rules carry most of the value:

• Finish the request within its scope. Complete necessary supporting work and verification without re-asking for existing authorization. Investigate technical uncertainty yourself. Ask for missing product decisions or materially broader scope; reversibility does not authorize new features or refactors.
• Stakes decide review, not size. Auth, money, data integrity, security, privacy, or hard-to-undo changes get an independent adversarial-reviewer pass even when the diff is one line.

Everything else in the repository exists to install those rules safely, keep them from drifting, and keep private context out of the public copy.

What ships

Surface	Contents	Source
Claude plugin	/brainstorm, /impl, /plan, /trim; the adversarial-reviewer agent; the clean-writing and design-taste skills	claude/plugin/
Claude config	Public-safe global CLAUDE.md, rules/, hooks, the teammate agent protocol, and reusable skills (conductor, i-have-adhd, wizard, summarize, svelte5-best-practices, retro, ponytail)	claude/config/
Pi	Agents, extensions, prompt templates, themes, and pinned packages; the o90 behavior text is opt-in (--with pi-text)	pi/
Private Codex config	Overlay-only personal instructions and a portable named profile; no public Codex adapter	--with codex
Repository tooling	Bootstrap, rollback, drift, leak, and verification checks	bootstrap.sh, install.sh, scripts/

The Claude plugin and Pi configuration work independently. Plugin-only installs provide commands, the reviewer, clean-writing, and design-taste; they do not install global CLAUDE.md, rules, hooks, or global skills. Linked hooks run only when your Claude settings register them. Fresh settings enable session discovery and the push guard; the WebFetch guard is available but not enabled there.

Pi installs extensions and packages by default. Only the additional o90 behavior text is opt-in. See Pi's README for runtime options.

When to move up a rung says which workflow to reach for as work grows from a single prompt to a conductor session. Choices with more than one defensible answer are numbered in docs/decisions.md; D5 to D9 record the 2026-09 strip-back and the evidence behind it. Agent provenance documents the optional attribution line. Scope examples make the completion boundary reviewable across Claude, Codex plugin migration, and Pi.

Choose an install

This repository configures applications; it does not install Claude Code, Pi, or credentials. The scripts require Git and Python 3.9+. Full Pi bootstrap also requires Bun and Pi; Claude bootstrap requires Claude Code.

git clone https://github.com/vrennat/other-ninety.git
cd other-ninety

Choose the smallest setup you need:

Setup	Action
Claude plugin only	In Claude: /plugin marketplace add vrennat/other-ninety, then /plugin install other-ninety@other-ninety
Configuration only	./install.sh --with claude --with pi
Configuration plus dependencies and plugins	./bootstrap.sh --with claude --with pi

Both scripts preview without writes. Review the output, then repeat with --apply. Omit a component you do not use. With no --with flags, the default is Pi; otherwise the flags select the exact set. Add --with pi-text alongside --with pi only to load the additional o90 behavior text.

The install matrix explains file ownership and target overrides. The new-machine checklist covers verification. Package and plugin operations persist outside the config rollback manifest.

Private overlay

Keep personal configuration outside this repository. An overlay mirrors the destination groups and replaces the same public path:

my-private-overlay/
├── claude/
│   ├── CLAUDE.md
│   ├── settings.json
│   └── skills/
│       └── conductor/
└── pi/
    ├── agents/
    └── settings.json

./bootstrap.sh --apply --with claude --with pi --overlay ../my-private-overlay

There is no JSON merge or template engine. Maintain a complete replacement when the overlay owns a path. Overlay groups apply only when their component is selected. Codex is an explicit private-only component:

./install.sh --with codex --overlay ../my-private-overlay

It manages only codex/AGENTS.md and codex/other-ninety.config.toml under CODEX_HOME (or --codex-dir); base configuration, credentials, host trust, and plugin/app state remain local. Review the plan and repeat with --apply.

Installer, bootstrap package selection, and drift checks share the same target plan. Preserved real Claude skill directories are reported as unmanaged; overlay-owned copies are compared exactly.

Rollback, drift, and leaks

A successful apply prints its manifest path. Restore only manifests created by this checkout:

./install.sh --rollback ~/.local/state/other-ninety/backups/<timestamp>/manifest.json

Check drift and private data with the same component and overlay arguments you installed with:

./check-drift.sh --with claude --with pi --overlay ../my-private-overlay
./check-leaks.sh --patterns-file ../my-private-patterns.txt

check-leaks.sh proves each scanner rule is active with a positive control, then checks the working tree and Git history. It cannot prove free-form prose is safe; public releases still require manual review.

Updating

Pull changes into this checkout, review the install plan, then rerun the same setup command and component/overlay arguments. Linked files follow the checkout; preserved settings stay local and overlay settings are copied again on apply. Restart the selected runtimes after updating.

Plugin caches are separate from checkout links. Use /plugin marketplace update other-ninety and /plugin update other-ninety@other-ninety in Claude, or rerun Claude bootstrap to update both. Package and plugin updates are outside configuration rollback.

Repository layout

other-ninety/
├── .claude-plugin/       # marketplace catalog
├── claude/
│   ├── plugin/           # distributable Claude Code plugin
│   └── config/           # shared global Claude configuration
├── pi/                   # Pi defaults, prompts, and extensions
├── shared/               # Pi's compact-writing policy
├── docs/                 # ladder, decisions, install notes, archive
├── scripts/              # installer, checks, and tests
├── bootstrap.sh          # single-command config bootstrap
└── install.sh            # dry-run-first config installer and rollback

Develop

git config core.hooksPath .githooks
scripts/verify.sh

The repository hook runs plugin lint and the leak scanner before pushes. The verification script also runs installer and bootstrap tests, shell and JSON checks, Pi typechecking, and Pi tests.

Verification evidence

Implementation reports carry a compact verification receipt tying checks to the baseline and final work, with commands, exit status, behavior evidence, and an explicit PASS, FAIL, or INCONCLUSIVE verdict. Requested behavior and preserved behavior are checked separately; test changes receive scrutiny. See verification guidance. Claude's plugin command and Pi's implementation prompt and validator carry the same agreement without relying on global Claude rules. Codex has no public adapter here; private overlays require a separate adoption change. These are reporting practices, not mechanically enforced merge gates.

Release policy

Publishing a release and changing live configuration are separate actions. A release must pass repository verification, a shadow apply and rollback rehearsal, and manual privacy review before it is used as a new machine or live configuration baseline.

Name

The name comes from Tom Cargill's observation at Bell Labs:

The first 90 percent of the code accounts for the first 90 percent of the development time. The remaining 10 percent of the code accounts for the other 90 percent of the development time.

Acknowledgments

The Pi timer, effort control, MCP health display, session alias, cache telemetry, and original clean-writing modes were independently adapted from ideas in Michael Lam's mical-pi. Later clean-writing refinements draw on Michael's clean-writing guidance and Lauren Tan's unslop skill.

The verification receipt, requested/preserved behavior distinction, and explicit inconclusive outcome draw on ideas in Assure. The guidance is written for this toolkit; it does not incorporate Assure's runner or code.

License

MIT