A portable skills workspace for AI coding agents, the kind that gives Claude Code, Codex, and Gemini clear workflows, safe defaults, and the ability to delegate to one another with session continuity. Clone it as a starting point for any project where you want agents to work deliberately rather than freestyle.
Models are good at the next token. They are less good at the next hour: keeping the work coherent across long sessions, multiple files, and the moments where one agent hands off to another.
agent-designer is built around that gap. Rather than another tool, it is a small workspace of disciplines you bring with you into a project: a folder of skills in plain markdown, a few bridge scripts, and a way to break large work down so an agent can drive it on its own.
Each skill follows the same shape, with frontmatter the agent always sees, a workflow body it loads on invocation, and references it pulls only when the path narrows. That shape is portable: the same skills work under Codex, under Claude Code, under Gemini, with minimal adaptation. Most changes here are text, not code. The repository is workflow-first.
Out of the box you get two complementary disciplines. Cross-agent collaboration gives you three bridge scripts that let one agent consult another with full session continuity, returning the same JSON envelope on every turn. The issue-driven workflow lets you plan once, then hand the agent a CSV it can execute end to end, every row a granular, testable, markable unit of work.
When Codex hits a problem it wants a second opinion on, it can ask Claude and stay in the same conversation through five turns. When Claude needs the web, it can hand the question to Gemini and read what comes back. Each bridge wraps a CLI and returns the same shape: a JSON envelope with a SESSION_ID you pass back to resume.
Delegate review, diff, and design consultation. Multi-turn back-and-forth while you implement locally, with second opinions, edge-case checks, and test gaps.
Delegate implementation, diagnosis, or adversarial review. Sandboxed by default, read-only unless you explicitly grant workspace writes.
Delegate research with built-in web grounding, or hand off image and screenshot analysis. Returns a unified diff when you ask for code changes.
Front-load the thinking and the agent has a full plate of useful work per run. The plan and Issue CSV act as a work amplifier, every row a granular, testable, markable unit. The agent reads, sets Dev_Status = DOING, completes the row, and immediately moves to the next.
| ID | Task | Dependencies | Dev_Status | Review1_Status |
|---|---|---|---|---|
| 001 | Audit current scoring weights in ranker.py | — | Done | Done |
| 002 | Add BM25 baseline test fixtures | 001 | Done | Done |
| 003 | Implement reciprocal-rank fusion in ranker.py | 001, 002 | Doing | Todo |
| 004 | Wire the fusion path through search/handler.py | 003 | Todo | Todo |
| 005 | Add latency benchmark, assert p95 < 80ms | 004 | Todo | Todo |
| 006 | Document the new ranking path in README | 004 | Todo | Todo |
| 007 | Regression sweep across saved query corpus | 003, 004, 005 | Todo | Todo |
The CSV is the agent's state. It reads to know where it is, updates as it works, and keeps driving forward. It stops only for genuinely blocking unknowns and decides anything routine itself. Loop: plan → issues → implement → test → review.
Skills are markdown. Most changes here are prose, not refactors. The discipline lives in how you write the workflow, not in the surrounding scaffolding.
The same skills run under Codex, Claude Code, and Gemini. The same envelope returns from every bridge. Move between agents without re-learning the surface.
The Issue CSV turns vague intent into a queue the agent can complete autonomously. Stops only for genuinely blocking unknowns.