Crew

30 days to change your mind, no questions asked, refunded through Stripe. The refund policy in full.

Every file has one job, and one non-job

The reviewer knows what the debugger owns. The verification skill knows what "done" means for both. ROSTER.md is the design document that assigns each of the twenty-four files a job, a boundary and a hand-off, and it says plainly that any overlap between two of them is a bug worth filing an issue over.

Check it before you pay

A prompt pack is the one kind of product you cannot evaluate through the box, so we opened it. SAMPLES.md publishes the complete text of one paid agent and one paid skill, the description of all eighteen paid-only files, and the exact diff between the six free files and their paid counterparts. That diff is six inserted cross-references and nothing else, which means the free tier is an honest sample of the writing: judge the bar on the three free agents and you have judged it on the files you have not read.

The same page carries a recorded run of the free reviewer and a paid agent against the same public commit of this store's engine. They overlapped on one observation out of fifteen, and disposed of it in opposite directions, each correctly for its own job. The free reviewer also found a real bug on a money path in our own code, and that finding is published intact rather than edited out of the transcript.

It also names the three files we think are the weakest in the pack, with the reasons. You would find them a day after paying anyway.

What's inside

Ten agents. A correctness reviewer that reads outside the diff (most review-detectable bugs live in unchanged code whose assumptions just broke), a systematic debugger that forms ranked hypotheses before touching anything, a planner that argues scope before code, plus tester, security, performance, refactoring, docs, simplifier, and a release captain that takes a branch from code-done to published and stops the line when preflight fails.

Fourteen skills. The disciplines that keep agent work verifiable: test-driven development, verification before claiming done, PR authoring, accurate changelogs, safe shipping checklists, structured handoffs between sessions.

Hooks and templates. Six auditable shell hooks and CLAUDE.md starters for solo and team repositories. Five are guards that block: force-pushes to main, destructive shell commands, exfiltration, out-of-scope edits, and secrets entering files. The sixth, format-on-edit, is a convenience that fails open.

The installer. POSIX sh, zero dependencies: --dry-run narrates before touching anything, checksum-based conflict detection refuses to clobber your edited files, --uninstall removes exactly what it installed. Tested against fresh, upgraded, and conflicted setups.

How delivery works

Checkout asks for your GitHub username. The fulfillment bot invites that account to the private Honorboxx/crew-full repository, usually within minutes and always within a few hours. Access is permanent; updates land in the same repo.

Terms

• $19, one-time. No subscription. Every later update is included at no extra cost.
• Licensed per developer; use across all your projects.
• Don't republish or resell the pack itself.
• 30-day refunds, no questions asked, via Stripe.
• Support: GitHub issues or honorbox@proton.me.