agent-workflows

Twenty-one experimental multi-agent workflow skills for Claude Code: parallel research, brainstorming, prototyping, debate, stress-testing, planning, execution, and review, with Beads for cross-session task tracking.

The premise: one agent produces one answer; several agents with uncorrelated context produce a map of the answer space. Each skill spawns independent agents, forces variety (different lenses, different decompositions, different failure narratives), and treats disagreement between them as signal about where the problem is ambiguous. When debate stops paying, the skills switch to building parallel prototypes in isolated git worktrees and let code argue.

Skills

Category	Skills
Ideation and research	brainstorm, diverge, fracture
Decision and debate	converge, constraint-inversion, distill
Risk and validation	premortem, stress-test, diffuse
Prototyping	diverge-prototype, crossbreed, entangle, migrate
Specification	contract, replicate
Debugging	bisect
Execution	focus, scaffold, prd-build
Pipelines and meta	research-project, compose

Per-skill reference pages live in docs/, one per skill plus workflow guides. The skills/ tree also carries supporting utilities (simplify, code-review, prd-queue, prd-build-status/-resume/-cancel, mcp-eval).

Picking an entry point

The core pipeline runs research through implementation:

/brainstorm > /diverge > /converge > /premortem > /diverge-prototype
    > /stress-test > pick winner > /scaffold > /focus or /prd-build

Not every project needs every step. Match the entry point to what you already know:

You have	Chain
A single task, ready to execute	`/focus "build auth with OAuth2 and RBAC"`
Clear requirements, multiple components	`/scaffold design.md`, then `/focus` through the seeded beads
Two or more viable designs	`/converge`, then `/premortem`, then `/scaffold`
A PRD, want to spike implementations	`/diverge-prototype`, then `/stress-test`, pick the winner
A vague idea	`/brainstorm`, then `/diverge`, `/converge`, `/premortem`
A topic, want code at the end	`/research-project "topic"`, then `/prd-build prd_topic.md`
A risky change to assess	`/diffuse`, then `/stress-test`
A regression to isolate	`/bisect`
No idea which skills to use	`/compose "your goal"`

Worked examples: full pipeline, solo ideation, fast debate on known options, research-project walkthrough.

Installation

Install the repo as a Claude Code plugin. One step covers all 21 skills, four agents (code-reviewer, code-simplifier, research-project, security-pipeline), and the hooks:

git clone https://github.com/sjarmak/agent-workflows.git
claude --plugin-dir ./agent-workflows

Skills are then invoked as /agent-workflows:<skill>:

/agent-workflows:diverge "How should we design the auth system?"
/agent-workflows:premortem path/to/design.md
/agent-workflows:stress-test src/auth/

Pipeline agents appear under /agents. To load the plugin for every session, add its path to plugins in your project's .claude/settings.json.

To adopt a single skill instead of the plugin, copy its directory into your Claude Code configuration:

cp -r skills/diverge ~/.claude/skills/diverge   # global; use .claude/skills/ for one project

Prerequisites for specific skills

Beads (/focus, and the bead-seeding phase of /scaffold): install bd and Dolt per the Beads README, then run bd init in your project. Without Beads, /scaffold skips seeding and /focus prompts for bd init.

Brainstorm ships a Python backend (skills/brainstorm/scripts/) that tracks ideas and enforces shape uniqueness:

pip install -r skills/brainstorm/requirements.txt

numpy is required; sentence-transformers is optional and enables the semantic-similarity layer. Without it, brainstorm falls back to lexical similarity.

Converge requires Claude Code agent teams: set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in the env block of your settings.

Hooks

hooks/hooks.json ships five hooks: UserPromptSubmit injects the changed-file list from git diff so skills scope to what you actually touched; PostToolUse runs prettier, black, or gofmt on edited files; Stop verifies that expected output files (prd_*.md, premortem_*.md, and friends) exist and are non-empty; WorktreeCreate installs dependencies in fresh worktrees so parallel prototype agents start runnable; Notification raises a desktop notification, useful because these skills spawn multiple agents and take minutes.

Status

Experimental and in active use. Skill interfaces and internals change without deprecation cycles; BEST_PRACTICES.md records what has held up, including the working defaults: start implementation sessions with /focus, one bead at a time, plan before executing, and let /scaffold seed the task graph.

Related repositories

coding-agent-workflows: a portable, pre-rendered instantiation of these standards for Claude Code, Codex, Amp, and any AGENTS.md-reading agent; now superseded, with development continuing here.
beads: the dependency-aware issue tracker (bd) that /focus and /scaffold use for cross-session task state.

License

MIT