gpt-with-taste gives Codex and Claude Code one behavioural recipe

A single behavioural recipe that installs into Codex, Claude Code and four other coding harnesses, and makes each one answer the same way, is the pitch of gpt-with-taste. The repository, by NB-Corp, picked up 355 stars in its first seven days under an MIT licence. The authoring source is one English recipe in a recipe/ folder, and the project compiles it to each host's native instruction format.

Six harnesses, one recipe

The compile targets are OpenAI Codex (native model instructions plus a hooks overlay), Claude Code (main instructions append and a Sub plus three Steer plugin), Kiki (Native Recipe bundle with Sol and Astra compatibility), Pi (request-local extension with a payload guard), Oh My Pi (identity-aware extension), and DeepSeek Harness (Cordis plugin with request-local steer). A porting guide covers other hosts. The project targets the GPT reasoning family, especially GPT-6.1 Sol and GPT-6 Astra, and replies follow the language the user wrote in.

Installation into Codex is a git clone plus node scripts/install.mjs --hooks. The install writes two managed prompt files into the active CODEX_HOME (default ~/.codex), points Codex's model_instructions_file at the main one, and leaves the model, provider, effort, authentication, permissions, MCP servers, roles and AGENTS.md untouched. The hooks route scoped reminders at user submissions, child startup and supported wait results. Users review and trust the three installed hook definitions in Codex Desktop under Settings, Hooks, or in the terminal with /hooks. A prompt-only install runs without hooks.

What the recipe changes

The README's featured comparison is a task about writing a Chinese getting-started document from an installation note. The default Codex run spends its first paragraphs on installation audits and process disclaimers before getting to the content the reader asked for. The gpt-with-taste run leads with the content and keeps the audit text out of the reader's path. The project's own description of the change is that it chooses sensible defaults, preserves supported claims, synthesises collaborator results into direct action, and finishes work with care. The system prompt sets standing priorities for judgment, execution, collaboration, and craft, and the steer layer refreshes those priorities at supported events through hooks, extensions or native recipe events.

The authors have published the findings and the iteration loop behind the recipe at docs/research.md, a case archive at docs/cases.md, and a web comparison page with full inputs, generated outputs and run conditions. A node scripts/check-updates.mjs command checks the installed version against the official metadata, and the installation does the same check before writing unless --skip-update-check is passed.

For a team that switches between coding agents

A developer who uses Codex at the desk and Claude Code on a laptop maintains two system prompts today, each in that host's own format, and keeps them in sync by hand. One recipe compiled across hosts means the two agents produce the same shape of answer, with one place to change the behaviour. The project is MIT-licensed, so a team can fork the recipe and keep its own version of what judgment looks like in its codebase. The compile is a plain Node script, with no npm packages installed, which is a smaller surface than most coding-agent tooling asks for.

Source

The repository is on GitHub at NB-Corp/gpt-with-taste. The web comparison page is at nb-corp.github.io/gpt-with-taste.