honey

Write less code and say less about it. Applies YAGNI and stdlib/native-first so the agent writes the minimum code that needs to exist, and responds tersely — stripping filler, hedging, and pleasantries while keeping code, identifiers, and technical terms exact. Use whenever writing, modifying, refactoring, reviewing, or explaining code, or any response where output volume drives token cost — even if the user never says "minimal" or "concise". Especially in agentic coding, where the volume of generated code and prose runs up the bill.

Compress-Cache-Retrieve for huge, repetitive array tool output (logs, scan results, time series, event streams) before it enters context. Keeps an informative sample — endpoints, anomalies/change-points, head/tail — drops the redundant rest to a local cache, and leaves a retrievable hash. Use when a tool returns a long uniform JSON array you must read but mostly skim, and the full set is one command away if needed. Lossy-but-recoverable.

Honey for plain Claude — the terse-prose core with no agent-harness features. Paste into a claude.ai Project's custom instructions, a Style, or an API system prompt. Strips filler, hedging, and pleasantries from every response while keeping facts, names, numbers, steps, and code exact. No tools required.

Rewrite a memory or context file (CLAUDE.md, AGENTS.md, a todo or notes file) into Honey-terse form so it costs fewer input tokens every session, without losing meaning. Backs up the original first. Use when asked to shrink or compress context files, trim CLAUDE.md, or cut per-session input cost. Prose only — never code, config, or data.

Harvest every `honey:` comment in the codebase into a debt ledger, so the deliberate shortcuts Lever 1 leaves behind get tracked instead of rotting into "later means never". Flags markers with no revisit trigger — those are the ones that go permanent silently. Use when the user says "honey debt", "/honey-debt", "what did honey defer", "list the shortcuts", or "what did we mark to do later". One-shot report; changes nothing.

Same pixels, fewer tokens — for user-facing deliverables where visual polish is the spec. Use when building or editing a landing page, marketing site, hero, pricing/feature section, dashboard, or any HTML/CSS UI component. Keeps the full rendered design (layout depth, hierarchy, motion, responsive richness, a11y) and cuts tokens by expressing that design densely — CSS custom properties, shared classes, shorthand, fluid units — instead of by cutting the design. The honey core trims code and prose; this trims how the design is *written*, never how it looks. Reach for it whenever output is user-facing markup, even if the user never says "minimal".

Report this session's output tokens and CO₂ by running the repo's committed EcoLogits port, plus the modelled CO₂/$ saved vs a no-Honey baseline — always labelled with the bench stamp it came from. Use when asked how much Honey saved, the session's carbon/token footprint, or to expand the 🍯 statusline badge into a full breakdown.

Show Honey's benchmark scoreboard — the committed quality and token results per task tier (code, user-facing, agent-to-agent) from bench/. Reports only the reproducible committed figures, never invents per-repo numbers. Use when asked how much Honey saves, how it compares to Caveman / Ponytail / no-skill baseline, or for the headline numbers.

Decide when to delegate to Honey's read-only subagents (hive-scout, hive-reviewer) instead of working inline, so the expensive tokens — large file reads and review passes injected back into the orchestrator's context — come back compressed (Honey Lever 3). Use when a task is search-heavy or review-heavy, spans many files, or you want to keep the orchestrator's context small. Not for trivial one-file work.

Cost discipline for recurring /loop runs. A loop multiplies per-iteration cost by iteration count, so waste compounds. Encodes cache-aware pacing (avoid the 300s dead zone), event-driven-over-polling, a no-change short-circuit, a compact state handle carried between ticks, and a stop condition. Use when a /loop is running or being set up — especially long-lived or frequently-firing ones.

Create and maintain one committed per-project memory file (PROJECT.md) so agents stop re-discovering the same facts every cold session. Stores only the stable, expensive-to-rediscover, not-in-the-code context — architecture + why, build/test/run commands, decisions and dead-ends — as human-readable markdown, versioned with the code so it can't silently rot. Use when asked to set up project memory/context, stop agents re-grepping the codebase, scaffold or refresh a PROJECT.md / CLAUDE.md memory section, or cut cold-start rediscovery cost. Prose context only — never code, config, or data.

Read huge, dense, read-only text as rendered PNG pages instead of raw text — image tokens scale with pixels, not characters, so token-dense bulk (big files, vendored code, diffs, logs) costs ~60–75% less as an image. Use when you must skim or reason over thousands of lines you will NOT edit or byte-copy. Lossy on exact strings: never for files you'll Edit, secrets, hashes, or byte-exact values. Fable-class readers only.

Review a diff for what Honey would cut — over-engineering (speculative generality, hand-rolled stdlib, single-caller abstractions) and over-verbosity (dead code, narration, redundant comments). Outputs a terse delete-list, one line per finding, with the line count each cut saves. Use when asked to review, trim, or de-bloat a change, or before opening a PR. Quality lever only — it finds waste, not correctness bugs.

Stack Honey onto Superpowers-style workflows (subagent-driven-development, dispatching-parallel-agents, executing-plans) and any orchestration that dispatches fresh subagents. Dispatched subagents run in isolated context and do NOT inherit the session's Honey hook, so they emit full-fat code and verbose reports. Use when you are about to dispatch implementer/reviewer/fixer/parallel subagents: it gives the exact Honey directive to paste into each dispatch prompt, with a worker variant and a reviewer variant, so the levers apply inside the subagents too — where the multiplied cost actually is.