/next-move
The flagship skill. Tell it what you're working on (or don't — it reads your git state), and it produces a predicted DAG of the highest-impact agents to run next.
/next-move # auto-detect from git + conversation
/next-move fix the flaky auth tests # tell it what you care about
What you get:
- Problem classification — well-structured, ill-structured, or wicked
- 3-8 subtasks decomposed into parallel waves
- Each subtask matched to the best skill from 594 via BM25 retrieval
- Risk analysis with mitigations (PreMortem agent)
- Time and cost estimates per node
- Accept / Modify / Reject feedback loop
How it works: A 5-agent meta-DAG runs inside your Claude Code session:
Wave 0: Sensemaker — classify the problem, halt gate if uncertain
Wave 1: Decomposer — break into subtasks with dependencies
Wave 2: Skill Selector — BM25 search → match skills to subtasks (parallel)
Wave 2: PreMortem — scan for failure patterns (parallel)
Wave 3: Synthesizer — assemble the predicted DAG + present to user
No extra API keys. No external services. Runs entirely in your Claude session.
Skill Search
BM25 ranked retrieval across the full catalog. Porter stemming, real stopword removal, document-length normalization. Zero API keys — runs locally.
Queries like "drone computer vision" find computer-vision-pipeline and drone-cv-expert. "Wedding photo 3d gaussian" finds wedding-immortalist at 35.9. Stemming handles morphological variants — "deploy" matches "deployment" and "deployed".
DAG Execution
Skills don't just get recommended — they get executed in parallel waves. The WinDAGs engine spawns real agents (claude -p) with skill-injected prompts, running independent nodes concurrently.
Claude Code (plugin):
claude plugin add /path/to/windags-skills
Manual (any agent):
git clone https://github.com/curiositech/windags-skills.git
# Use /next-move:
cp -r windags-skills/skills/next-move ~/.claude/commands/
# Or symlink the whole catalog:
ln -s $(pwd)/windags-skills/skills ~/.claude/skills
Single skill:
cp -r windags-skills/skills/api-architect ~/.claude/commands/
| Category | Count | Highlights |
|---|
| Agent & Orchestration | 87 | dag-orchestrator, next-move, skill-architect, task-decomposer |
| Research & Academic | 76 | raft-consensus, bdi-agents, chain-of-thought, tree-of-thoughts |
| Backend & Infrastructure | 62 | api-architect, microservices, caching, websocket |
| Design & Creative | 50 | design-system-creator, pixel-art, typography, vibe-matcher |
| DevOps & Infrastructure | 41 | github-actions, kubernetes, terraform, docker |
| Cognitive Science & Decision Making | 37 | naturalistic-decision-making, sensemaking, expertise-elicitation |
| Frontend & UI | 37 | nextjs, react-performance, animation, framer-motion |
| AI & Machine Learning | 34 | ai-engineer, RAG, computer-vision, embeddings |
| Code Quality & Testing | 23 | vitest, playwright, code-review, refactoring |
| Productivity & Meta | 20 | prompt-engineer, documentation, skill-creator |
| Data & Analytics | 18 | data-pipeline, dbt, data-viz, dimensional-modeling |
| Mobile Development | 17 | ios, react-native, flutter, swiftui |
| Lifestyle & Personal | 14 | ADHD, grief, finance, interior-design |
| Content & Marketing | 10 | SEO, copywriting, product-launches |
| Career & Interview | 9 | interview-prep, resume, hiring-manager |
| Recovery & Wellness | 8 | sobriety, crisis-intervention, speech-pathology |
Academic Formatting
- APA formatting: APA 7 student and professional papers, author–date citations, and References.
- MLA formatting: MLA 9 papers, author–page citations, and Works Cited.
Both skills include source guidance, worked examples, and document scaffolds. They flag missing bibliographic details and distinguish text checks from rendered document checks.
Design Tools
- Product palette designer: bespoke OKLCH ramps, light/dark semantic tokens, Tailwind/shadcn/native exports, and coordinated brand assets.
How Skills Work
Each skill is a markdown file (SKILL.md) with YAML frontmatter:
---
name: api-architect
description: "Expert API designer for REST, GraphQL, gRPC architectures"
category: Backend & Infrastructure
tags: [api, rest, graphql, grpc]
---
The body contains activation triggers, core capabilities, anti-patterns, and reference files. When an agent loads a skill, it gains bespoke expertise — not just instructions, but deep domain knowledge with working examples.
Skills follow the Agent Skills open standard and work with Claude Code, OpenAI Codex CLI, Gemini CLI, Cursor, VS Code Copilot, and 30+ other agents.
Install
Claude Code (one command)
claude plugin marketplace add curiositech/windags-skills
claude plugin install windags-skills
This installs all 594 skills + the 5 /next-move subagents (sensemaker, decomposer, skill-selector, premortem, synthesizer).
Cross-tool install (Codex, Gemini CLI, Cursor, Aider, …)
Clone the plugin and run the installer. It detects which tools you have under $HOME and links skills + subagents into the right place. No hardcoded paths — uses $HOME and the plugin checkout dir.
git clone https://github.com/curiositech/windags-skills.git ~/coding/windags-skills
~/coding/windags-skills/scripts/install.sh
What it does (idempotent — safe to re-run after git pull):
| Tool | What gets installed |
|---|
| Claude Code | ~/.claude/skills/{next-move, windags-*} + ~/.claude/agents/{5 subagents}.md |
| Codex | ~/.codex/skills/{next-move, windags-*} + ~/.codex/AGENTS.md |
| Cursor / Cline / Aider | ~/AGENTS.md (auto-discovered by AGENTS.md-aware tools) |
| Gemini CLI | Marked block appended to ~/.gemini/GEMINI.md |
Override the install location with WINDAGS_HOME=/path ./install.sh. Preview without changes with --dry-run.
The MCP server lives at <plugin>/mcp-server/index.js — wire it into any MCP-aware client. It exposes 10 tools and 1 prompt:
| Tool | Purpose |
|---|
windags_skill_search | Cascade-ranked skill search (BM25 + MiniLM + RRF + cross-encoder reranking) |
windags_skill_graft | Full SKILL.md bodies for top primaries + adjacencies + asset manifest |
windags_skill_reference | Load one reference file from a skill's references/ |
windags_history | Recent /next-move predictions for a project |
windags_skill_search_batch | N searches in one round-trip (capped at 20) |
windags_skill_graft_batch | N grafts in one round-trip (capped at 20) |
windags_node_requirements | Per-skill allowed-tools, pairs-with, suggested model_tier, and provider-native model IDs |
windags_validate_dag | Schema-check a candidate DAG before saving |
windags_estimate_cost | Per-node + total cost estimate during planning |
windags_run_pipeline | Run the full 5-stage /next-move pipeline server-side. Provider-agnostic — auto-detects from env (Anthropic / OpenAI / Google / Groq / OpenRouter / Together / DeepSeek / Fireworks / Cerebras / xAI). Pin one with WINDAGS_PROVIDER=<id>. ~$0.02–$0.15 per run. Returns a halt response or a validated PredictedDAG. |
| Prompt | Purpose |
|---|
next_move | Templated /next-move runner. Drives the 5-stage pipeline (sensemaker → decomposer → skill-selector + premortem → synthesizer) for non-Claude clients that don't have Claude Code's slash commands. Args: task (focus hint), fresh ("true" to ignore conversation history). |
License
BUSL-1.1 (Business Source License). Free for non-commercial and personal use. Converts to Apache 2.0 on 2030-03-03.
Commercial licensing: [email protected]
