Teaching Claude to stop talking in commit hashes

What it does

Open Steps is a set of prompt-engineered skills and shell hooks that reframe an AI coding agent’s output into plain-language reports with clear verdicts. It tells you whether a session is done, what debt was created, and what you need to do next—without requiring you to parse commit hashes or middleware logs. It targets Claude Code first, with adapters for Codex, Cursor, and Gemini CLI.

The interesting bit

The author is a self-described “market-led builder” with no engineering background who started “vibecoding” and found the agent reporting in jargon he could not decode. He built eight interlocking skills—os-done-or-not, os-what-could-go-wrong, os-big-picture, and others—that form a closed loop from picking work to verifying it. It is product-management guardrails expressed as prompt engineering, and the author freely admits where the tooling is thin.

Key highlights

• Eight skills that fire at different moments: session wrap-up, human hand-off, risk assessment, backlog grooming, and plain-language rewriting.
• os-big-picture maintains a living BIG-PICTURE.md file tracking feature status and queue, refreshed by a scripts/census.sh script that measures code age and liveness.
• os-check-work explicitly distrusts other sessions’ reports and verifies claims against actual repository state before signing off.
• Built and measured on Claude Code; other agents receive the same skills and routing block, though the README admits reliability there “is not checked.”
• The os-big-picture skill refuses to invent tasks or delete backlog items, and honestly flags when its own liveness metrics are too young to trust.

Caveats

• Full validation exists only for Claude Code; Codex, Cursor, and Gemini CLI support is present but the author notes “how reliably the skills fire there is not checked.”
• Some configuration stays manual by design—no installer silently overwrites your existing CLAUDE.md writing style.
• On repositories younger than six months, the os-big-picture liveness column explicitly admits it “cannot mean anything yet.”

Verdict

Grab this if you are a product person or founder vibe-coding with agents but need to understand what actually happened without learning git internals. Skip it if you are already comfortable reading raw agent output or do not use Claude Code, Codex, Cursor, or Gemini CLI.