Write less code and say less about it. Honey (I Shrunk the AI) by GreenPT is a cross-tool coding skill that cuts the output tokens AI coding agents spend — making them emit less code and less prose without losing correctness. It works with Claude (claude.ai and the API), Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, Windsurf, Cline, OpenClaw, oh-my-pi, Kiro, Kilo Code, Hermes Agent, Kimi Code, Devin CLI, Grok Build, Qoder, Antigravity, and CodeBuddy. Three independent levers, applied reflexively:
- Less code — YAGNI first. Walk a ladder (does it need to exist? → stdlib → language native → existing dependency → one line → minimum block) and stop at the first rung that works. The cheapest line is the one you never write.
- Less prose — drop the wind-up, the hedging, the narration of code that already speaks for itself. Answer first.
- Denser agent-to-agent handoffs — when the reader is another agent, not a human, hand it the most token-efficient format it parses losslessly (compact / columnar JSON, or ESON). Smaller handoffs with no loss of recovery (−49% on Opus 4.8, −15% on Opus 5.5; two tasks, so no p-value). Fires only here — never as a user-facing answer.
Honey combines what Ponytail (minimal code) and Caveman (terse prose) do separately, then goes further:
- Auto-intensity —
lite/full/ultrachosen reflexively from the request, with no deliberation tax (it never spends reasoning tokens deciding how to comply — that would defeat the purpose on reasoning models). - Safety carve-outs — input validation, error handling, auth, secrets, migrations, deletes, and anything you explicitly asked for are never compressed. Lazy ≠ broken.
- A skill family, not one prompt — an always-on core plus on-demand satellites (review, eco, gain, compress) and a hive of read-only subagents that return compressed handoffs. See Skills and subagents.
Claude Opus 5.5, 23 tasks × 3 runs, against the current upstream skills. Every figure
is a paired per-task median vs no skill, with a Wilcoxon p-value
(full-opus55b):
| Skill | Output | Lines of code | Cost | Judge vs no skill | Tests |
|---|---|---|---|---|---|
| Caveman v4.0.1 | −29% | −27% | −26% | tie (3/8/10) | 100% |
| Ponytail v5.3.2 | −28% | −52% | −26% | worse (1/16/4, p<0.001) | 100% |
| Honey | −40% | −52% | −31% | tie (5/12/4) | 100% |
Quality is a tie with no skill: fewer tokens at the same quality, not better quality.
Those 23 tasks are author-written. On 21 Exercism tasks picked by a fixed rule, Honey cuts
less: output −21%, lines −37%, cost −10% (all p≤0.007), with 100% of tests passing against
97% for no skill (ext-opus55-stub); read that as the
conservative figure. In a real agent loop (Cline, Opus 5.5) the output cut holds (−39%,
no extra turns) but the bill stays flat (−1%, ns), because input dominates a session; the
cost figure above is single-call only (details). Per-tier numbers, older models, method and every caveat:
docs/benchmark.md and bench/. Regenerate any figure
offline with node bench/src/report.js --stamp full-opus55b --by-type.
/plugin marketplace add Green-PT/honey-for-devs
/plugin install honey@greenpt
Then /honey once to turn it on (/honey lite|full|ultra to set intensity,
/honey off to stop). The state persists across sessions and is kept per
project; /honey default <mode> sets it for every repo without its own. A
SessionStart hook re-activates it every session and adds a ~50-token map of the
repo's top-level folders and shared helpers, so the agent reuses before it writes
(HONEY_MAP=0 turns the map off). A 🍯 badge shows the
active mode in your statusline. If your client autocompletes /honey to
honey:honey, that's the same command.
Every other platform (Codex, Cursor, Copilot, Gemini CLI, Windsurf, Cline, Kiro, plain Claude and ChatGPT, …) and the one-line installer: docs/platforms.md.
- Skills and subagents — the satellites (review, eco, gain, compress, check, …) and the read-only hive
- Agent handoffs and big inputs — ESON, CCR, PX
- Carbon badge and usage report
- Benchmark results · Methodology
- Contributing — one source skill, generated rule files
- Changelog
MIT — see LICENSE.
The carbon-estimation data and coefficients in hooks/eco-models.json and
hooks/eco.js are derived from EcoLogits
and remain under the MPL-2.0. See NOTICE for details.