English · 简体中文

A drop-in behavioral recipe across coding harnesses: make ordinary decisions, digest complexity, and deliver work with engineering and visual craft.
Built for the GPT reasoning family, especially GPT-6.1 Sol and GPT-6 Astra. Replies follow the language you write in.

Harnesses · Install Codex · What changes · Comparison · How it works · Research

Harnesses

GPT with Taste compiles from a single behavioral recipe into native instructions and event steering for each supported host:

Running in another host? See docs/harnesses/porting.md. Maintainers edit the single English authoring source in recipe/ and compile across targets: see docs/authoring.md.

Install · Codex

With Node.js 22 or newer installed:

git clone https://github.com/NB-Corp/gpt-with-taste.git
cd gpt-with-taste
node scripts/install.mjs --hooks

We recommend Hooks for multi-turn work in Codex Desktop: user submissions, child startup and supported wait results receive scoped reminders. Later child-input coverage depends on the client entry point. For a prompt-only installation, run node scripts/install.mjs instead. Neither route needs npm packages.

Both routes install into the active CODEX_HOME (otherwise ~/.codex), write the two managed prompt files, and point model_instructions_file at the main one. Your model, provider, effort, authentication, permissions, MCP servers, roles, and AGENTS.md remain unchanged. To target another home, add --home with a quoted absolute path, for example --home "C:/Users/you/.codex".

With --hooks, review and trust the three installed definitions in Codex Desktop under Settings → Hooks, or in the interactive terminal with /hooks. Enable only the events you want. Open a new Codex session after installation.

Check updates with node scripts/check-updates.mjs. Installation also checks official metadata before writing; add --skip-update-check for offline installation. Update instructions cover the installed components and customized checkouts.

Prefer to delegate? Give your coding assistant this repository and docs/install-for-ai.md. To recover the managed state, use docs/recovery.md. The complete installation contract is in docs/install.md.

What changes

GPT with Taste puts judgment into the outcome: choose sensible defaults, preserve supported claims, synthesize useful collaborator results, and finish work with care.

The recipe combines judgment, continuous execution, collaboration, and engineering and visual craft. Explore the principles in Research, or inspect complete runs, task inputs, and side-by-side outputs in the Case Archive and Web Comparison.

Comparison · From installation logs to a first task

Task request: Use notes.md to write a Chinese getting-started guide for a colleague who has just finished installing. Save it as START_HERE.md.

The visible improvement is straightforward: get the reader started first, while installation audits and process disclaimers step out of the way.

Inspect full inputs, generated documents, and run conditions: Task & run conditions · Web comparison page · Case index.

How it works

1. The system prompt sets standing priorities. It covers judgment, execution, collaboration, and craft across hosts. The instructions are English; replies follow your latest language.
2. Steer refreshes priorities at supported events. Using local hooks, extensions, or native recipe events, reminders route user context to main, scoped reminders to delegated agents, and input-consumption guidance to main upon receiving collaborator results.

Standalone prompt definitions provide the complete behavior for each agent role; supported harnesses connect the matching reminder paths.

Research

docs/research.md details the findings, prompt architecture, and iteration loop behind the recipe: how defensive writing affects the reader's experience, why models need room for ordinary judgment, and how collaborator findings are turned into direct action.

MIT License · Credits: SpadeLushen/pi-gpt-with-taste · Community friend link: Linux DO