NoBuzz
A Claude Code skill that takes Claude’s last answer and runs it through the Antigravity CLI so Gemini can say it like a person, with the translation printed verbatim and three audience modes.
GitHub avatar of adnanakil, not the project’s own logo — taken from github.com on 2026-10-02.
What it is
NoBuzz is a Claude Code skill invoked as /debuzz. It takes Claude’s last response, hands it to the Antigravity CLI, and prints Gemini’s rewrite verbatim — the skill states it will not let Claude tidy up the translation, because that reintroduces the voice being removed. It ships three modes: colleague keeps every file path and code block and drops the theatrics, manager gives roughly a third of the length without code, and director gives three to five sentences.
How it is put together
The parts · 3A two-model relay with one rule holding it together: the second model’s output is printed verbatim. The skill is a directory of instructions for one agent, plus a call out to a command-line tool from a different vendor, so the whole system is the routing between them — which response is being rewritten, which audience it is being rewritten for, and what is not allowed to happen to the result.
- skills/debuzz/
- The skill itself, installed by copying the directory into
~/.claude/skills/. It carries the invocation, the three modes and the verbatim rule. - The Antigravity CLI (
agy) - The second model in the relay, and the only external requirement besides Claude Code. It runs Gemini and is invoked to translate a response that is already written.
- The three modes
- Audience contracts rather than style settings:
colleaguepreserves code and paths,managerdrops both and cuts to about a third,directorreduces to three to five sentences.
Choices, and what they beat
Hand the answer to a different model instead of prompting this one to be plainer over more system-prompt instructions telling Claude to drop the voice
The README states the conclusion the project started from: no amount of prompting fully cures the register, so the honest move is to pass the response to a model that does not have it.
Print the second model’s translation verbatim over letting Claude review or polish the rewrite
Stated in the README as the reason the skill works at all — any tidying at that step reintroduces exactly the voice being removed.
Read fromREADME of adnanakil/nobuzz, read 2026-10-02.
Build log
4 stages- 01
Built with Claude, and the commit trailers say so
The repository was created on 2026-08-21 and its last push is 2026-08-22 — two days, five commits, no releases. Four of the five carry a
Co-authored-bytrailer naming a Claude model (Fable 5), and all five are linked to a GitHub account. The README opens by saying the skill was built by Claude and the author together, then states the observation it was built on: ask Claude why a test is flaky and the answer arrives with a load-bearing assumption, three numbered findings, and a note that the third is the most instructive yet. - 02
The decision is to stop trying to prompt the voice away
The skill accepts that prompting does not cure the register and does something else instead: it hands the response to a second model whose only job is to write plainly. The choice is stated as a rule rather than a preference — the translation is printed verbatim, because letting the first model “tidy up” the rewrite puts back exactly the voice the skill exists to remove. That single constraint is what makes the skill a pipeline between two models rather than a prompt.
- 03
Three audiences, one command
/debuzz [mode] [text]takes an optional mode and an optional body of text; with neither it works on the previous response. The modes are defined by what survives rather than by tone:colleaguekeeps the content, every file path and every code block;managerreports what happened, why it matters and what is next at about a third of the length with no code;directorreturns three to five sentences covering outcome, impact and ask, and assumes thirty seconds of attention. - 04
The backend changed on the second day
The fifth commit is a merge titled
Merge PR #1: switch backend from Gemini CLI to Antigravity CLI (agy), contributed by a second GitHub account. The README that ships today documents the Antigravity CLI as the requirement, with install commands for macOS, Linux and Windows and a note thatagyhas to be run once to complete the Google sign-in before the skill works. The earlier Gemini CLI path is gone from the installation section.
Adjacent records
All records →No. 125
sepia
A portable de-AI writing skill: four operations over one canonical rules file, narrative architecture repaired before word choice on fiction, a thin rule file matched to the venue on professional prose, and every rule labelled as measured, consulted or the project’s own inference.
No. 114
agy-staff
A plugin that hires Google’s Antigravity CLI as a member of staff: the host agent — Claude Code, Codex or Pi — keeps the decisions and hands the surveys, reviews and scoped edits to a Gemini 3.8 Flash worker that runs in the background and answers through a job id.
No. 083
agent-manager
Ten coding agents, one terminal list: live status read from each tmux pane, a quick prompt that answers them without attaching, a dead session revived on its own conversation, and a diff review whose comments go back to the agent as one numbered round.