NB-Corp/gpt-with-taste
A system prompt that tells GPT to stop hedging and just decide
One behavioral recipe, compiled into native instructions and hooks for six coding harnesses, all aimed at making GPT exercise ordinary judgment.
Collecting fresh signals — velocity needs a few days of history.
collecting data…
star history
What it does
GPT with Taste is a single-authored behavioral recipe (in recipe/) that compiles into native system instructions and event-based steering for OpenAI Codex, Claude Code, and four other harnesses. The goal is behavioral: pick sensible defaults, stop burying the reader in process disclaimers, and finish work with actual craft. It targets the GPT reasoning family — the README name-drops GPT-6.1 Sol and GPT-6 Astra — and replies follow whatever language you write in.
The interesting bit
The two-part design: a standing system prompt sets priorities, then hooks/extensions re-inject scoped reminders at events like user submissions or child-agent startup, so priorities don’t evaporate over a long session. Maintainers edit one English source and compile across all six targets, which is a genuinely tidy way to keep prompt glue from forking.
Key highlights
- Single recipe, six harness integrations (Codex, Claude Code, Kiki, Pi, Oh My Pi, dsh), plus a porting guide for others
- Hooks route different reminders to main vs. delegated agents, and guide input consumption when collaborator results arrive
- Codex install leaves your model, provider, permissions, MCP servers, and
AGENTS.mduntouched; runs offline with--skip-update-check - Ships a case archive and web comparison with full inputs and side-by-side outputs, not just vibes
- Includes a research doc on why defensive writing hurts the reader and how the prompt architecture was iterated
Caveats
- The Codex hooks route depends on the client entry point for later child-input coverage — multi-turn steering isn’t uniform across hosts
- “Better taste” is inherently subjective; the featured comparison is one curated case, so your mileage on real tasks is the real test
- It’s prompt engineering all the way down — no model changes, so ceiling effects of the underlying model still apply
Verdict
Worth a look if you use Codex or Claude Code and are tired of output that reads like an audit log. If you’re happy with your current agent’s demeanor, this is a lot of ceremony for a system prompt.
Frequently asked
- What is NB-Corp/gpt-with-taste?
- One behavioral recipe, compiled into native instructions and hooks for six coding harnesses, all aimed at making GPT exercise ordinary judgment.
- Is gpt-with-taste open source?
- Yes — NB-Corp/gpt-with-taste is an open-source project tracked on heatdrop.
- What language is gpt-with-taste written in?
- NB-Corp/gpt-with-taste is primarily written in JavaScript.
- How popular is gpt-with-taste?
- NB-Corp/gpt-with-taste has 504 stars on GitHub.
- Where can I find gpt-with-taste?
- NB-Corp/gpt-with-taste is on GitHub at https://github.com/NB-Corp/gpt-with-taste.