NB-Corp/gpt-with-taste

gpt‑with‑taste

What it is – A lightweight, “drop‑in” recipe that adds a layer of judgment and craft to the responses of large‑language‑model coding assistants (OpenAI Codex, Anthropic Claude, Kiki, Pi, DeepSeek, etc.). The project ships a single English‑language system prompt and a set of event‑driven “reminders” that are compiled into native instructions for each supported host, so the model behaves with more sensible defaults, better collaboration handling, and a focus on engineering‑level output.

Key ideas

• Single source authoring – The behavioral recipe lives in recipe/. A build step generates the appropriate prompt files and hook definitions for every harness.
• Harnesses – Ready‑made integrations for:
  • OpenAI Codex (native model‑instruction overlay and optional hooks)
  • Claude Code (instruction append + plugin)
  • Kiki, Pi, Oh My Pi, DeepSeek (each with a native “recipe” or request‑local steering)
• Hooks & events – When installed with --hooks, Codex Desktop receives three managed prompt files that can be enabled per‑event (e.g., user submission, child‑process start, wait‑result). The system prompt establishes standing priorities (judgment, execution, collaboration, craft) and the hooks refresh those priorities during a session.
• Language‑aware – The system prompt is English, but replies follow the language the user writes in, as demonstrated with a Chinese‑language guide example.

Installation (Codex example)

# Prerequisite: Node.js 22+
git clone https://github.com/NB-Corp/gpt-with-taste.git
cd gpt-with-taste
node scripts/install.mjs --hooks   # installs hooks into CODEX_HOME (default ~/.codex)

• --hooks adds three prompt definitions that appear under Settings → Hooks in Codex Desktop.
• Omit --hooks for a prompt‑only install.
• Use --home "<path>" to target a different Codex configuration directory.
• Update checks are performed automatically; node scripts/check-updates.mjs can be run manually.

How it works

1. System prompt – Sets “standing priorities” (judgment, execution, collaboration, craft) that guide every model reply.
2. Steering events – Local hooks or harness‑specific extensions fire at defined moments, refreshing the priorities and routing context to the main prompt or delegated agents.
3. Compiled per‑harness – The single recipe is transformed into the native instruction format required by each host, so developers only maintain one source of truth.

Typical use‑case

• A developer asks Codex to generate a START_HERE.md onboarding guide. With the default Codex prompt the model may include unnecessary self‑justification and installation warnings. After installing gpt‑with‑taste, the same request yields a concise, reader‑first guide, demonstrating the “better judgment” and “less ceremony” goals.

Documentation & resources

• Harness details – docs/harnesses/*.md
• Authoring guide – docs/authoring.md
• Installation guide – docs/install.md
• Case archive – Real input/output comparisons in docs/cases.md
• Research – Design rationale and prompt‑architecture discussion in docs/research.md

License – MIT (see LICENSE).

Who might like this – Teams that already use LLM‑powered coding assistants and want a consistent, higher‑quality behavioral layer without rewriting prompts for each tool.