Claude Code Engineer, Migrate a Working AI Governance System from Cowork Prose Rules to Real Hooks

JOB DESCRIPTION I run an AI-assisted consulting practice on a platform called Claude Cowork. Over the past several months I built a real enforcement system for it: skills, templates, and about 47 Python gate scripts that check things like "did you actually verify this before calling it done," "did this file get saved to the right folder," "does this document still sound like me and not like AI filler." The problem is that all of that enforcement is prose — the AI is instructed to run these checks and follow these rules, but it's reading and remembering, not being forced. Under a long session it skips things. I've had real, repeated incidents of gates being skipped even though the rule was written down clearly. Claude Code (a different product from the same company) has an actual fix for this: hooks. A hook is a real shell command the platform runs automatically at a defined point (before a tool runs, after a tool runs, when the AI tries to finish its turn) and it returns an exit code. Some hook events can genuinely block — the AI cannot proceed until the check passes. That's a different category of enforcement than a prose instruction, because it's outside the model's discretion. I already have a working proof of concept: a sealed test environment with 6 hooks built and 72 of 72 tests passing (PreToolUse guards blocking destructive file operations and wrong save locations, a Stop-hook gate checking that required deliverables actually exist, a post-write logger). I also already have a written spec mapping my real production gate scripts to the hook events they should run on. What I need is someone who can take that foundation and my real scripts and actually do the migration: wire the real checks into real hooks, close the gaps the prototype already surfaced, and get the whole thing running end-to-end on my Windows machine, not just passing tests in isolation. This is not a "build me an AI agent from scratch" job. The architecture and the plan already exist. I need someone who has actually shipped Claude Code hooks in a real project before and can execute a defined migration, not learn hooks on my dollar. WHAT'S ALREADY BUILT (you'd be extending this, not starting from zero) - A sealed Claude Code hook environment, 72/72 automated tests passing, already run successfully live on Windows via Git Bash. - A written spec mapping specific existing Python gate scripts to specific hook events (PreToolUse, PostToolUse, Stop, SessionStart), including which ones should be blocking and which are feedback-only. - Real production scripts to migrate in, for example: a pre-delivery verification gate, a document voice/style checker, a spreadsheet data-integrity checker, a stale-content checker, a test-suite runner. KNOWN GAPS TO CLOSE (found by testing the prototype, not guessed) 1. A Windows-only LibreOffice headless-conversion hang (a profile/lock flag issue) that times out file-conversion checks at 60 seconds. 2. Shell-level commands bypass tool-level guards entirely — anything routed through a raw shell call (mv/rm/del/Remove-Item equivalents) currently evades the Write/Edit hook checks. 3. A "finish" gate that verified a filename matched, not that the file's actual content was correct — needs a real content-level check. 4. Verdict-parsing bugs in a couple of gate scripts: brittle string-matching that misreads passing output as failing (and vice versa) depending on incidental wording. 5. A blocking Stop hook needs a recursion guard, and I need the exact, correct mechanism — guessing wrong here means an infinite loop, not a soft failure. 6. A subagent config declares a Claude-Code-style "Bash" tool that doesn't exist in Cowork — I need someone who understands the difference between Claude Code's tool surface and other Claude deployment surfaces well enough to catch this class of mismatch generally, not just this one instance. WHAT "DONE" LOOKS LIKE - The real production gate scripts (not just the prototype's sample scripts) wired into hooks per the migration spec, with each one verified by deliberately breaking the thing it's supposed to catch and confirming the hook actually fires and blocks. - All 6 known gaps above fixed, with a test proving each fix. - Full test suite still green, plus new tests for what you fixed. - A short handoff doc: what you changed, why, and how I verify it still works if I touch it later. - Everything running end-to-end on my real Claude Code install on Windows, not just in your own environment. PAID TEST TASK (required before hire, not optional) Fixed-price, $50, paid regardless of outcome, timeboxed to about an hour: I'll send one small, isolated bug reproduced from the real hook code. You fix it, write a test that proves the fix, and send back a short screen recording (Loom or similar, 3-5 minutes) walking through how you found and fixed it. This is the actual hire decision — the interview is a filter, the test task is the evidence. WHAT I'M NOT LOOKING FOR Proposals that restate this post back to me in different words, general "I'm an AI/automation expert" pitches with no hook-specific detail, or anyone unwilling to do the paid test task before a full-scope hire.