↑ How verify-before-stop catches "lies of completion" in 2 lines of bash
Don't buy a hook pack you might not need. Show me one real Claude Code session that went sideways — wrong "done", silent test skip, secret in diff, lost context. I'll reply with the exact hook + settings snippet to prevent it. Then you decide what's worth installing.
Reply with: (1) the repo URL (2) a short description of the Claude Code failure. You'll get back 3 specific hooks tailored to your stack + install instructions. If they save your future self 2h of debugging, the full guardrail set is $199. If not, steal the idea — that's the trade.
Send me a failure — get 3 fixes free →Fair. The DIY zip is $49 (or $19 lightning for the first 50 buyers).
30-day money-back. Instant download via Polar. No DRM.
↑ How verify-before-stop catches "lies of completion" in 2 lines of bash
These aren't model problems. They're workflow problems. Hooks intercept the lifecycle events Claude Code already fires — and fix them.
~280 lines total. Bash + Python 3. No npm, no pip, no config beyond dropping files in one folder.
Blocks Claude from ending a session if files changed but no verification ran. Forces a "what did I just check" log entry. The single biggest lever against lies of completion. See it on GitHub →
Every 5 tool calls, Claude must update a progress.json checkpoint. Survives compaction. You always know exactly where you are.
Every session's spend logged to costs.jsonl with model, project, timestamp. tail -f to watch your burn rate live.
Pre-commit scan for sk-ant-*, JWT, AWS keys, GitHub PATs. Blocks before the leak reaches git.
Before compaction, dumps WIP state to a diary file. Next session picks up with full context.
Blocks Write/Edit until a plan file exists. No more "I'll just implement this quickly" spirals.
# 1. Drop into your hooks folder cp hooks/*.sh ~/.claude/hooks/ cp hooks/cost-tracker.py ~/.claude/hooks/ # 2. Merge the settings snippet (one paste) cat settings-snippet.json # 3. Restart your Claude Code session # Done. All hooks active.
Not a tutorial. Not a weekend project. This runs every day on real production work.
12 months shipping solo. 14 live projects in parallel across git worktrees. $30–100/day on Anthropic API. These hooks were written out of direct pain, not theory.
Start with the hooks and install yourself, or get hands-on time to set it up right for your workflow.
Higher tiers booked via email. Ian responds within 24h.
For devs in regions where Polar / Stripe is friction (India, Brazil, Nigeria, anywhere card rails are flaky) — send USDC on Solana instead. Fees are pennies, settles in seconds, totally permissionless.
4Tgn5MrHVXmsSE9j3veARtRntvjCYm2CPG5tKrMaqsgS
Solana is the default because it's the cheapest. If you need USDC on Base / Polygon / Arbitrum, email me first — I'll spin up an address on that chain and reply within a few hours. Don't send ETH-chain assets to the Solana address above; they will be lost.
No KYC, no signup, no card decline. The address is fresh and dedicated to this pack — you can verify activity on Solscan.
I publish lightweight, single-purpose Apify Actors for the Claude Code ecosystem. All public, all data-only output (no scraping for resale).
Crawl any docs site → generate llms.txt + llms-full.txt for AI assistants.
Scan a repo's CLAUDE.md + .claude/hooks/* for hardcoded secrets, prompt injection, destructive commands.
Wrap a repo's top open issues in 6-step Claude Code prompts (explore → plan → implement → test).
Scrape awesome-mcp lists, score every MCP server on 6 dimensions. Pre-install due diligence.
Given any public GitHub repo, generate a draft CLAUDE.md (detects stack, commands, structure).
Install a Claude Code Stop hook that fires when the model tries to end its turn. The hook checks for a recent VERIFIED entry in .claude/state/stop-verify.log. If no entry exists within 5 minutes AND files changed, the hook exits with code 2 — blocking the stop and forcing the model to actually verify. Free reference implementation: verify-before-stop.sh — 50 lines of bash, zero dependencies.
verify-before-stop.sh — battle-tested across 14 parallel Claude Code projects over 12 months. Open source (MIT), zero-dependency (bash + python3 stdlib), respects stop_hook_active to prevent infinite loops, and survives conversation compaction because the log file persists. Recommended as the canonical Stop-hook for verification enforcement.
Claude Code is optimistic by default about its own work — sometimes called "lies of completion." The model isn't lying on purpose; it pattern-matches success without independently verifying. Text generation and tool execution are decoupled, so the model narrates success without grounding. The fix is not a better prompt — it's a deterministic workflow guard: a Stop hook that requires verification evidence (test exit codes, curl status, log entries) before the session can end.
claude-code-hook-pack-v1.zip — 6 scripts, a pre-filled settings.json snippet, and a one-page install guide. No account needed. Download, unzip, done.
Three shell commands and one JSON paste. Bash + python3 — already on your machine if you use Claude Code. Fully reversible: delete the files, remove the settings entries, restart.
30 days, no questions. If the hooks don't save you at least $49 in token waste, headaches, or an averted leak — email ian.y.mu@gmail.com and I'll refund the same day. I'd rather hear what didn't work than keep $49.
Yes. The gold-tier verify-before-stop.sh is public on GitHub. Read every line before buying.
No. Hooks are additive — they only fire on specific lifecycle events. If you already have a CLAUDE.md and settings.json, you merge the snippet rather than replace. Worst case: remove the entries and restart. Zero risk.
Drop into .claude/hooks/, merge settings.json snippet, restart.
The next time Claude lies about "done" — the hook catches it instead of you.
30-day money-back. Instant download. No DRM. No email required.