Hey everyone,
I code with an agent all day and kept hitting the same wall: give it a one-line request and it happily sprints straight into a pile of files, and I only find out it misread the scope after it’s already written half of it. So I built a protocol to slow that first step down and keep a human in the loop at the moments that actually matter.
It’s called ADP (Agent-Driven Protocol), a methodology (a set of skills the agent follows), open source, with the reference implementation for Claude Code. The flow: reverse interview to discover the real scope → PRD → sliced issues → TDD implementation → QA → PR, stopping at 4 human approval gates so nothing big becomes code without a “yes” from you. It’s deliberately the opposite of the fully autonomous direction. I think the interesting problem is designing where the human belongs in the loop, not removing them.
The two pieces people react to most:
-
A living architecture map. Every time a feature lands, the agent re-renders an n8n-style canvas of the codebase, features, dependencies, integrations, each node linking the real files. It’s the first thing I open when I come back to a project cold.
-
A secrets guardrail on the PR. A PreToolUse hook scans the diff before the PR goes out and blocks it if it catches keys/secrets, so the agent can’t quietly commit a
.envline.
One detail this crowd might appreciate: the discovery skill (grill-me) researches the state of the art first (GitHub, HN, Reddit, dev.to, Product Hunt) before it interviews you, so the scoping conversation starts from what already exists instead of a blank page.
Free / open source:
-
Repo (the free “Fase Zero” tier, a discovery interviewer + a PRD generator):
https://github.com/murilomn58/Claude-Spec-Driven-Fase-Zero -
Direct download, no signup: https://claudespecdriven.com.br/starter
Native install (Claude Code):
/plugin marketplace add murilomn58/Claude-Spec-Driven-Fase-Zero
/plugin install adp-fase-zero@adp
Honest disclosure: the site and the plugin’s prompts are written in PT-BR (I’m a Brazilian dev and built it for the BR crowd first). The methodology is language-agnostic. The agent follows the PT-BR skills and answers in whatever language you’re working in. A paid full version exists, but the free tier above is the point of this post and enough to try the approach.
I’d genuinely like this group’s take on the gate design: where would you place human approval in an agentic coding loop, and what would you automate away? Open to issues/PRs.