Cc Orch Development

orch-development

A Claude Code skill for running software development as **orchestrate-and-verify** instead of write-it-yourself. Your session model becomes PM + lead architect + QA gate; developer subagents (any models or CLIs you have) do the implementation, and the orchestrator independently verifies every piece before it lands.

Philosophy

A report is a claim, not a fact — the orchestrator verifies everything with its own tool calls and never writes the code itself. Different model families catch different bugs, so load-bearing work is reviewed across families and nothing grades its own homework. Done is a machine-checkable fact (a green command, a durable CI/E2E gate), never a model's opinion of itself.

Install

Copy this folder into your skills directory:

cp -r orch-development ~/.claude/skills/orch-development

(Or drop it into a plugin's `skills/` directory.) It activates on phrases like "orch-development", "orchestrated development", "multi-agent dev", or "let the orchestrator drive".

**Two files, and the split matters.** `SKILL.md` is the orchestrator's file — every gate is stated there in full, because a gate that lives in a file you might not open has stopped being a gate. `DEV_BRIEF.md` is the developer's file: the role override, the no-backgrounding rule, TDD, the engineering standards, the honest-report contract. It is **pasted verbatim into every dispatch** rather than consulted, so it stays one versioned artifact — which is also what lets you attribute a change in dev outcomes to a change in what devs were told (`git log --before= -1 -- DEV_BRIEF.md`). Retyping it from memory is how a `## Git` requirement went missing and a dev finished green with 784 passing tests and nothing committed.

`devlog.py` ships alongside the skill as a working reference implementation of the retrospective log — append one entry per verification, then aggregate before the next run:

python3 devlog.py add --json '{"kind":"task","