Menu
Cost discipline for recurring /loop runs: cache-aware pacing, event-driven-over-polling, no-change short-circuit, compact state handle, stop condition.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | honey-loop |
| description | Cost discipline for recurring /loop runs: cache-aware pacing, event-driven-over-polling, no-change short-circuit, compact state handle, stop condition. |
| homepage | https://github.com/Green-PT/honey-for-devs |
| license | MIT |
A loop is the one place token waste compounds: cost = per_tick × ticks. The
honey levers still apply inside each tick (less code, less prose), but a loop adds
two leaks the single-shot levers don't cover — re-paying for context every
wake-up, and re-doing work that didn't change. Cut those.
Pace to the cache, skip the dead zone. The prompt cache TTL is 5 min.
< 270s → context stays warm; right for actively polling external state.≥ 1200s → one cache miss amortized over a long idle wait; the default for
idle ticks (use 1200–1800s).~300s is the worst choice — pays the cache miss without amortizing it.
Never round to "5 minutes." Drop to 270s or commit to 1200s+.Event-driven over polling. Harness-tracked work (background Bash, Agent,
Workflow) re-invokes you automatically on completion. Don't schedule a short
tick to poll it — that re-reads full context for nothing. Set a long fallback
heartbeat (1200s+) and let the completion notification drive you. Poll only
external state the harness can't see (CI, a deploy, a remote queue), at a
cadence matched to how fast that state actually changes.
No-change short-circuit. Re-generating output is the bulk of per-tick cost.
Check cheaply first (a hash, a timestamp, a git rev-parse, a status field). If
nothing changed: emit one line and reschedule. Don't redo the task, don't
re-read the world, don't write a report.
Carry a compact state handle, don't re-derive. Keep what the loop has done
so far in a small file (.honey-loop/state.json or a <<honey:HASH>> stash via
eso stash), and read that each tick — not the full history re-discovered from
scratch. Lever 3b: the cheapest input token is the one that never re-enters context.
Define done; then stop. An unbounded loop is unbounded cost. State the exit condition up front (target reached, N empty ticks, deadline). When met, end the loop — omit the wake-up reschedule. "Loop forever" is a bug, not a feature.
A loop tick is an agent-to-self / agent-to-log message, not a deliverable. Default
to ultra: status line only, near-zero prose. One line is enough —
tick 4/∞ · no change · next 1800s. Step up to a full report only on the tick
where something actually needs the user (state changed, a decision, an error).
Compress-Cache-Retrieve huge repetitive array tool output before it hits context: keep a sample, cache the rest, leave a hash.
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.
Write less code and say less about it: YAGNI, stdlib-first, terse prose. Cuts agent token cost on every coding and writing task.
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.
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.
Read huge dense read-only text as PNG pages via pxpipe export: ~60-75% fewer input tokens. Lossy on exact strings; never for files you will edit.