🍯 Honey (I Shrunk the AI)

Write less code and say less about it. Honey (I Shrunk the AI) by
GreenPT is a
cross-tool coding skill that cuts AI coding-agent token usage and LLM API costs —
making agents emit less code and less prose without losing correctness. It works
with Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, Windsurf, Cline,
OpenClaw, and Kiro. Three independent levers, applied reflexively:

1. 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.
2. Less prose — drop the wind-up, the hedging, the narration of code that
   already speaks for itself. Answer first.
3. 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 ESO). Cuts handoff size ~in half at zero loss
   of recovery. 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 / ultra chosen 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 & subagents.

Why

Volume is cost. In agentic coding sessions, the volume of generated code and
prose is what runs up the bill — and most of it is waste.

This repo ships a reproducible benchmark (bench/) so you don't have
to take the numbers on faith: 23 tasks across three kinds of work — baseline vs
Caveman vs
Ponytail vs Honey — same model, same
prompts, only the skill changes. Correctness is objective (unit tests, structural /
accessibility checks, and lossless round-trip recovery for agent handoffs); quality
is scored by a 4-model cross-family judge panel (median of Opus 4.8 + Sonnet 4.6

• Haiku 4.5 + GPT-5.5) under a neutral rubric that says nothing about length, so a
  terse skill gets no thumb on the scale. The figures below are the committed results
  (Claude Opus 4.8, 3 runs each) — run cd bench && npm run bench to reproduce.

A single blended number hides the story, because the levers fire differently per
task type. Quality is % of baseline (panel median; for handoffs, lossless
recovery); tokens are generated output vs baseline:

Honey leads quality where it matters most — it tops the user-facing and
agent-to-agent tiers (the quality-separating ones) and stays within judge noise
of the pack on saturated code tasks — while cutting tokens where it's safe to:

• Code — the deepest cut (−49% output) at essentially tied quality (98% vs
  100%, within judge noise on tasks every variant passes). Caveman saves less;
  Ponytail's mandatory self-check inflates trivial code (+24%).
• User-facing — the carve-out keeps Honey from compressing polish, yet it
  still trims output (−6%) while earning the top quality score (101% of baseline)
  and the only 100% accessibility pass; Ponytail strips hardest and drops to 81%
  on the structural/a11y checklist.
• Agent-to-agent — under adversarial relay queries (ordinal, nested, absence,
  cross-field count) Honey is the only variant that stays 100% lossless while
  roughly halving handoff size (−51%); Caveman and Ponytail compress harder and
  lose recovery (67% / 50%). Its biggest, cleanest win.

The same pattern holds on GPT-5.5 (full two-provider table in
bench/results/cross-provider.md): Honey is the
only variant with no test regressions across all three tiers on Opus, and on
both models it keeps top-tier quality while cutting tokens on every tier.

Efficient Structured Output

Honey includes ESO, a zero-dependency, schema-first format for
agent handoffs. Repeated record keys are emitted once; declared row counts catch
truncated messages; JSON-compatible cells preserve types.

The reproducible ESO/TOON/JSON benchmark measures bytes,
two tokenizer estimates, codec speed, and lossless recovery across five agent
handoff shapes. Run it with npm run bench:eso.

printf '%s' '{"from":"reviewer","findings":[{"sev":"H","issue":"expired token"}]}' | eso encode
eso decode < handoff.eso

CCR — for huge, redundant array tool output

ESO is lossless, for handoffs where every row matters. CCR (Compress-Cache-Retrieve)
is the lossy-but-recoverable lever for the opposite case: a long uniform array you must
read but mostly skim — logs, scan results, event streams. It keeps an informative sample
(endpoints, anomalies/change-points, head/tail), caches the dropped rows locally, and
leaves a <<ccr:HASH N_rows_offloaded>> sentinel. Nothing is lost — retrieve restores
the original by hash on demand.

some-tool | eso crush          # → sampled view + sentinel; originals cached in .honey-ccr/
eso retrieve <hash>            # → the full original array, verbatim

Validated on a 90-row log (opus-4.8 + gpt-5.5): −82% tokens, crushed-only 96%
answer accuracy, 100% with retrieve — and the lone crushed miss was a refusal, not a
hallucination. Benches: npm run bench:ccr (tokens) and npm run bench:ccr:comprehension
(quality). The honey-ccr skill tells the agent when to reach for it.

Pick Honey when you want the best quality-per-token, especially in Claude Code.

Skills & subagents

Honey is one always-on core plus a family of on-demand tools. The core is a
writing style (it must be the default to pay off); the rest are actions you
reach for at a specific moment.

The hive is Lever 3 with a runtime: each subagent returns a compressed handoff,
so the result injected back into the orchestrator's context is −44–53% smaller
with zero loss (npm run bench:hive). Live, the skills hold up too — honey −86%,
honey-review −70%, hive-reviewer −43% output tokens at passing correctness
(npm run bench:skills). See bench/hive/RESULTS.md and
bench/skills/RESULTS.md.

On user-facing work — where the core skill spends tokens because polish is the
spec — honey-design keeps the same rendered polish for −19% output tokens vs no
skill (judge 92 vs 90), beating the core skill on both axes across 7 landing-page/UI
tasks. See bench/results/honey-design.md.

Honesty note. Earlier versions of this README quoted 92% / 78% / 73% quality
and −57% / −65% / −70% tokens from an unpublished run. Those don't reproduce —
the real quality spread is far narrower and the token savings are tier-dependent
(and Ponytail adds tokens on simple code). The table above is what the committed
bench/ harness actually produces; see
bench/results/combined.md for the full breakdown.

Install

Claude Code (plugin marketplace)

/plugin marketplace add Green-PT/honey-for-devs
/plugin install honey@greenpt

Then /honey to turn it on (/honey lite|full|ultra to set intensity,
/honey off to stop). A 🍯 badge shows the active mode in your statusline.

One-line installer (interactive wizard)

In a terminal it asks which agents you use, whether to wire the CO₂ badge, drop
per-repo rule files, and your default mode — then sets up exactly that. The wizard
prompts on /dev/tty, so it works through curl | bash. CI/pipes and --yes
fall back to auto-detect.

macOS / Linux / WSL / Git Bash:

curl -fsSL https://raw.githubusercontent.com/Green-PT/honey-for-devs/main/install.sh | bash

Windows (PowerShell 5.1+):

irm https://raw.githubusercontent.com/Green-PT/honey-for-devs/main/install.ps1 | iex

Windows (irm | iex) runs non-interactive; clone and run node bin/install.js
for the wizard. Add bash -s -- --yes to skip prompts. Requires Node.js on your
PATH. Safe to re-run; skips tools you don't have.

Every supported platform

All of these are also handled automatically by the one-line installer. See
INSTALL.md for manual steps, flags, and uninstall.

Carbon badge (Claude Code)

When Honey is active, the statusline also shows a live CO₂ estimate for the
session and the CO₂/$ saved vs a no-Honey baseline:

🍯 honey:full · 🌿 44g CO₂ (saved ~26g · $0.18)

(Illustrative — a ~2k-output-token Opus session.) The estimate is a faithful port
of EcoLogits v0.8.2 (verified to match
the package exactly). Model params come from EcoLogits' own registry
(hooks/eco-models.json, exported by
scripts/build-eco-models.py) — matched by exact id,
falling back to a per-family alias for frontier models too new for the registry.
Grid switches per provider — Anthropic on AWS Trainium (~500 gCO₂/kWh), OpenAI
on Azure (~400), Google on GCP (~330). Aliases, grids, and per-mode savings live in
hooks/eco-config.json.

The badge itself renders only in Claude Code (it reads Claude Code's
transcript, where every model is a Claude model). The provider switching matters
for scripts/eco_report.py, which runs against any
transcript — Codex/Gemini CLIs would each need their own statusline hook to show
a live badge there.

Params are speculative — Anthropic discloses none. EcoLogits' raw coefficient
is a single-stream (batch-size-1) upper bound — it gives one request the whole
GPU set for the full generation (for Opus, ~1.9 tok/s, ~30× slower than reality),
which alone is ~1.4 kg per 1M output tokens. Production serves many requests
concurrently, so the badge divides that ceiling by an effective batch concurrency
(serving_concurrency, default 32 — calibrated so modeled throughput matches real
~50–70 tok/s serving) to show realistic served impact. eco_report.py prints
both the served figure and the single-stream ceiling. Treat these as a range, not
a meter reading.

For the full breakdown (usage + embodied + primary energy) run the real package:

pip install ecologits
python scripts/eco_report.py        # newest session, or --transcript PATH

How it stays in sync

The skill is authored once in skills/honey/SKILL.md.
Every per-platform rule file (and AGENTS.md) is generated from it:

node scripts/build-rules.js          # regenerate all rule files
node scripts/build-rules.js --check  # CI: fail if any copy drifted

The OpenClaw skill package (.openclaw/skills/) is generated the same way from
skills/; rerun node scripts/build-openclaw-skills.js after changing a skill.
tests/openclaw-skills.test.js fails if a committed copy is stale.

License

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.