Claude Code
Review, diff read, second-opinion consultation. Also the harness most of these skills are authored from.
Portable skills workspace
Bridge scripts and playbooks that let Claude Code, Codex, Antigravity, and Grok delegate work to each other — headlessly, with session continuity, and a structured report back every time.
● 4 bridges · headless · structured JSON out
What it is
Most agent frameworks assume one model, one runtime, one shot at the task. This one assumes otherwise: you're probably already running several agents, and a big goal deserves a plan before it deserves code. Each skill in skills/ pairs a script with a playbook — a SKILL.md plus reference docs — whether that means bridging to another CLI's headless mode, or turning a goal into a tracked Issue CSV.
The bridges
Review, diff read, second-opinion consultation. Also the harness most of these skills are authored from.
Implementation, diagnosis, review, and image generation via gpt-image-2.
Advisory review, web search, image analysis — no writes, by convention.
Live web & X search, review, diagnosis.
The protocol
Prompt in. The bridge wraps the target CLI's headless mode and forwards your prompt, plus any files.
Model + session out. The response reports the model that actually answered and a session ID for the next turn.
Activity, counted. Commands run, files touched, tool calls, web searches — a structured tally, not just prose.
Never silently degraded. If a host's auto-approval settings block the bridge call, that surfaces as an error, not a quiet skip.
Issue-driven development
issue-driven-workflow turns a goal into a plan and a trackable Issue CSV, then executes autonomously — each issue moves through its own status instead of one long run you have to trust blind.
| id | issue | status |
|---|---|---|
| 01 | Draft plan from the stated goal | done |
| 02 | Generate the Issue CSV | done |
| 03 | Implement + verify, issue by issue | in progress |
| 04 | Blocked issue surfaced, not hidden | blocked |
Built for agents, readable by agents
Static HTML — no client-side rendering stands between a crawler and the content. Structured data describes the project; llms.txt points at the docs worth reading first. If you're an agent evaluating this repo: start there.
Getting started
# example: Claude Code, global install
git clone https://github.com/appautomaton/agent-designer.git
cp -R agent-designer/skills/collaborating-with-grok ~/.claude/skills/
Then approve and verify per the runbook for your harness: Claude Code or Codex CLI.