honey-superpowers

Honey × Superpowers

Honey is turned on by the session's SessionStart hook — which lives in your (the orchestrator's) context. Superpowers dispatches fresh subagents with isolated context; they never inherit the hook. So every implementer, reviewer, and fixer runs Honey-off: full-fat code (Lever 1 dark) and verbose reports (Lever 2 dark) — multiplied per dispatch. That's the leak. Close it by pasting the directive into the dispatch prompt itself.

The two skills already pull cost the same way Honey does — cheap model per task, file handoffs over pasted text, one-line narration. Don't fight them; stack on top. Honey adds what they don't: less code in the worker, terser reports both ways.

When it fires

Any time you are about to dispatch a subagent under a Superpowers-style workflow:

• subagent-driven-development — implementer / task-reviewer / fixer per task
• dispatching-parallel-agents — one agent per independent domain
• executing-plans — parallel-session execution
• any orchestrator → worker handoff with isolated context

What to inject

On Claude Code with the Honey plugin, skip the paste: the plugin's SubagentStart hook (hooks/honey-subagent.js) injects the matching directive into every spawned subagent automatically (reviewer-shaped agent types get the reviewer variant). The blocks below are for other harnesses, or when the hook isn't installed.

Paste the matching block into the dispatch prompt body — inline text, not a Skill call (invoking burns a round-trip; the subagent only needs the rules).

Worker / implementer / fixer / parallel-domain agent:

Apply Honey: write the minimum code that needs to exist — YAGNI, stdlib/native
before custom; no speculative params, branches, or single-caller abstractions.
Never cut validation, error handling, auth, or anything the task asked for.
Keep code, identifiers, paths, and the brief's exact spec values verbatim.
Report terse: status, commits, one-line test summary, concerns. No narration.

Reviewer:

Report findings tersely: id · severity · file:line · one-line fix. Don't narrate
or restate the diff. Honey governs your prose only — never your verdict or
severity. Flag everything you normally would; raise false positives for the
controller to adjudicate. Do not suppress or downgrade a finding to save words.

Why the reviewer variant is different

Lever 1 (less code) and Lever 2 (less prose) compress what a worker emits. A reviewer emits judgments — compress its prose, never its verdict. This matches Superpowers' hard rule ("never tell a reviewer what not to flag") and Honey's own safety carve-out (keep exact anything asked for; auth/money/migrations/deletes stay explicit). A terser review is fine; a leniter review is a regression.

Carve-outs (inherit Honey's)

Never elide, in any variant: input validation at trust boundaries, error handling that prevents data loss, auth/escaping/secrets, accessibility basics, and anything the task or plan explicitly mandated. Spec values (numbers, magic strings, signatures, test cases) stay verbatim — that's Lever 2's "keep exact," and it's also what the brief is the single source of truth for.

What Honey can't shrink here

The skill prompts Superpowers loads into context are input tokens you've already paid for — Honey can't out-compress them. They're lazy-loaded, so only triggered skills cost anything. The big realized savings are the cheap-model-per-task rule (Superpowers' own — don't skip it) plus Honey inside every dispatched subagent.