GPT with Taste prompt
System instructions and targeted reminders for GPT in Codex: fewer ceremonies, more decisions and careful craft
- 499 ★ on the repository
- Last commit
Extended access
Why this level
- Hooks run local Node.js code with your privileges and see prompt and tool metadata
- Installation changes the model's instructions in CODEX_HOME and the agent's behavior across tasks
Install
- Works in
Manual install
git clone https://github.com/NB-Corp/gpt-with-taste.git
cd gpt-with-taste
node scripts/install.mjs --hooksNeeds Node.js 22 or newer. Without --hooks only the prompts are installed. After installing, open a new Codex session and trust the hooks in settings.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Hooks run local Node.js code with your privileges and see prompt and tool metadata
- Installation changes the model's instructions in CODEX_HOME and the agent's behavior across tasks
This is third-party code. Review the repository files before installing.
What it does
GPT with Taste compiles a single behavioral recipe into native instructions for each supported host. The recipe teaches the model to make ordinary decisions itself and absorb complexity. It keeps to supported claims, synthesizes what collaborators return and finishes the work with craft, visual craft included. In Codex an install script writes two managed prompt files and points model_instructions_file at the main one. The second part is the Steer hooks. They add short reminders when the user submits a request, when a child agent starts and when a wait result arrives. For Claude Code the build puts the main instructions file on the system prompt append and loads a plugin with three reminders and an optional taste-worker agent. Variants exist without reminders and for an independent session. Replies follow the language you write in, while the instructions themselves are in English.
Who it is for. For people who run GPT reasoning models in Codex, Claude Code and other harnesses and want the same model behavior in every host.
Good fit when
- Codex answers with a long audit log, and the first useful step gets lost at the end of the reply
- Multi-turn work in Codex Desktop where reminders have to reach child agents
- The model asks confirmation for ordinary decisions it should reasonably take itself
- You run GPT in Pi, Oh My Pi or DeepSeek Harness and want the same recipe in the host's native format
- You want to compare results with and without the recipe: the repository has a case archive with inputs
Not a fit when
- You need project tooling: the recipe tunes the behavior of the model itself
- You do not trust hooks that run local Node.js with your privileges
- You need a strict protocol with mandatory confirmation at every step
- The host is outside the list: you would port the recipe using the porting.md guide
Example request
Use notes.md to write a getting-started guide for a colleague who has just finished installing. Save it as START_HERE.md.Limitations
Installation needs Node.js 22 or newer; no npm packages are required. The recipe targets GPT reasoning models. The Steer hooks run local Node.js code with the user's privileges and receive session, prompt and tool metadata; after installing, review and trust them in Codex Desktop under Settings, Hooks, or with /hooks in the terminal. Coverage of later inputs to child agents depends on the client entry point.
Claude Code needs version 2.1.296 with the --append-system-prompt-file and --plugin-dir flags and exec-form hook arguments. The dist directory is built by build.mjs. Load only a checkout you trust. Do not load the Steer and prompt-only variants together or append the main file through another entry in the same session.
How to disable. For Codex return to the pre-install state with docs/recovery.md, which restores the managed files. Hooks switch off in Codex Desktop under Settings, Hooks. For Claude Code start a session without the --append-system-prompt-file and --plugin-dir flags.
FAQ
How do I install GPT with Taste in Codex?
Clone the repository and run node scripts/install.mjs --hooks. The no-hooks variant is node scripts/install.mjs. The script writes two prompt files into CODEX_HOME (~/.codex by default) and points model_instructions_file at the main one. For another home add --home with an absolute path, then open a new session.
Does it replace my settings and AGENTS.md?
No. Model, provider, effort, authentication, permissions, MCP servers, roles and AGENTS.md stay unchanged. Only the managed prompt files and the model_instructions_file pointer change. You can restore the state with docs/recovery.md.
How do I connect it to Claude Code?
Build the files with node scripts/build.mjs --target claude. Then start claude --append-system-prompt-file pointing at dist/claude/main.md and --plugin-dir pointing at dist/claude/plugin. For the no-reminders variant use the prompt-only folder, and for an independent session use independent.md in place of main.md.
What do the Steer hooks do?
In Codex they send a short reminder when the user submits a request and when a child agent starts, and on wait results where the client supports it. The Claude Code plugin has three targeted reminders. The adapter reads only its own bundled text files and returns extra context. It writes no session state and sends no network requests.
How do I check for updates?
node scripts/check-updates.mjs compares the installed components with official metadata. Installation runs the same check; add --skip-update-check to work offline. Update and conflict steps are in docs/install.md.
Discussion
No questions or reviews yet.
Related
PromptsA collection of system prompts of ChatGPT, Claude, Gemini, Cursor, Devin and other AI systems extracted by the community. File authenticity is not verified
A teaching terminal coding agent in Python, built as a "nanoGPT for agents": read it, take it apart and fork it
A structured prompt system: templates with role, profile, skills, rules and workflow for repeatable model requests
Claude Code Prompts (RepoWise)
claude-code-prompts
Independently written prompt templates for AI coding agents: system prompts, tool prompts, delegation, memory and multi-agent coordination
The author's personal Claude Code toolbox: meta-prompt and todo commands, skills to create skills, hooks and MCP servers, auditors and a Ralph loop
Orchestration layer for Claude Code, Cursor and Codex: a board of 495 expert personas gives a verdict with named dissent, plus hundreds of skills
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| GPT with Taste (this entry) | no | 499 | |
| codex-keysmith | no | 4,785 | deploys instructions into Codex reversibly with a plan preview, while GPT with Taste ships its own behavioral recipe |
| tweakcc | no | 2,574 | patches Claude Code itself with custom system prompts, themes and toolsets |
README in short
The main example compares default model behavior with the recipe on one task: writing a getting-started guide for a teammate. The recipe compiles from one English source in the recipe folder, worked cases sit in an archive and the principles live in a separate Research document.