Datasets:

alex7604
/

agent-skill-routing

probe string	case_id string	tags list	prompt string	must_load list	must_not_load list	no_skill_at_all bool
anthropic	art-flow-field	[ "positive" ]	Make a generative art piece in p5.js - particles following a Perlin-noise flow field, with a seed I can change.	[ "algorithmic-art" ]	[ "canvas-design", "frontend-design", "web-artifacts-builder", "slack-gif-creator", "academy-guide", "discernment-nudge" ]	false
anthropic	poster-print	[ "positive" ]	Design a poster for our jazz night on March 14 - one PNG I can print at A3.	[ "canvas-design" ]	[ "algorithmic-art", "frontend-design", "slack-gif-creator", "web-artifacts-builder", "academy-guide" ]	false
anthropic	wall-print	[ "positive" ]	Make me a minimalist abstract print for my living room - geometric shapes, muted colors, a single image.	[ "canvas-design" ]	[ "frontend-design", "web-artifacts-builder", "slack-gif-creator", "academy-guide" ]	false
anthropic	slack-gif	[ "positive" ]	Make me a GIF of a cat typing furiously for our team Slack channel.	[ "slack-gif-creator" ]	[ "algorithmic-art", "canvas-design", "internal-comms", "academy-guide" ]	false
anthropic	frames-to-mp4	[ "near-miss" ]	Give me the ffmpeg command to turn a folder of PNG frames into an MP4 at 24 fps.	[]	[]	true
anthropic	landing-aesthetic	[ "positive" ]	I'm building a landing page for a coffee subscription startup. Help me choose an aesthetic direction and typography that doesn't look like every other template.	[ "frontend-design" ]	[ "canvas-design", "algorithmic-art", "webapp-testing", "slack-gif-creator" ]	false
anthropic	react-artifact	[ "positive" ]	Build me a claude.ai artifact - a multi-page project tracker in React with routing, shadcn/ui tables and filters.	[ "web-artifacts-builder" ]	[ "webapp-testing", "mcp-builder", "canvas-design", "academy-guide" ]	false
anthropic	single-html-toggle	[ "near-miss" ]	Write a single HTML file with a button that toggles dark mode.	[]	[ "web-artifacts-builder", "webapp-testing", "mcp-builder", "canvas-design", "academy-guide" ]	false
anthropic	playwright-login	[ "positive" ]	My Flask app runs on localhost:5000. Check that the login form shows an error on a wrong password and take a screenshot.	[ "webapp-testing" ]	[ "web-artifacts-builder", "frontend-design", "mcp-builder", "canvas-design" ]	false
anthropic	pytest-unit	[ "near-miss" ]	Write pytest unit tests for this function: def slugify(s): return s.lower().replace(' ', '-')	[]	[]	true
anthropic	mcp-jira	[ "positive" ]	Help me build an MCP server in TypeScript that wraps our Jira REST API so Claude can search and create tickets.	[ "mcp-builder" ]	[ "skill-creator", "webapp-testing", "web-artifacts-builder", "academy-guide" ]	false
anthropic	mcp-explain	[ "near-miss" ]	In two sentences: what is the Model Context Protocol?	[]	[ "mcp-builder", "skill-creator", "webapp-testing" ]	false
anthropic	skill-description-fix	[ "positive" ]	My SKILL.md never triggers. The description is just "Helps with PDFs." Rewrite it so it triggers reliably, and tell me how to test that.	[ "skill-creator" ]	[ "mcp-builder", "internal-comms", "canvas-design" ]	false
anthropic	skill-evals	[ "positive" ]	I have a skill that writes changelogs. Set up evals to check whether it actually improves the output, with variance across runs.	[ "skill-creator" ]	[ "mcp-builder", "webapp-testing" ]	false
anthropic	leadership-update	[ "positive" ]	Write this week's status update for leadership: SSO shipped, the billing migration slipped a week, two engineers hired.	[ "internal-comms" ]	[ "brand-guidelines", "canvas-design", "theme-factory", "academy-guide" ]	false
anthropic	incident-report	[ "positive" ]	Draft an incident report for yesterday's 40-minute checkout outage caused by an expired TLS certificate.	[ "internal-comms" ]	[ "webapp-testing", "canvas-design", "academy-guide" ]	false
anthropic	cold-email	[ "near-miss" ]	Write a cold email to a potential customer introducing our analytics product.	[]	[ "internal-comms", "brand-guidelines", "canvas-design", "academy-guide" ]	false
anthropic	anthropic-brand	[ "positive" ]	Restyle this slide outline with Anthropic's official brand colors and fonts.	[ "brand-guidelines" ]	[ "canvas-design", "algorithmic-art", "internal-comms" ]	false
anthropic	preset-theme	[ "positive" ]	Apply one of your preset themes to my quarterly report slides - something calm and blue.	[ "theme-factory" ]	[ "brand-guidelines", "canvas-design", "algorithmic-art" ]	false
anthropic	claude-pricing	[ "positive" ]	What's the cheapest Claude model for classifying a million support tickets, and how do I turn on prompt caching in the Python SDK?	[ "claude-api" ]	[ "mcp-builder", "skill-creator" ]	false
anthropic	openai-streaming	[ "near-miss" ]	In my Python script that uses the openai package, how do I stream GPT-4o responses token by token?	[]	[ "claude-api", "mcp-builder", "skill-creator", "academy-guide", "discernment-nudge" ]	false
anthropic	academy-start	[ "positive" ]	I'm new to Claude. How do I get started with Projects - are there any tutorials?	[ "academy-guide" ]	[ "skill-creator", "mcp-builder", "web-artifacts-builder" ]	false
anthropic	capital-city	[ "none" ]	What's the capital of Australia?	[]	[]	true
anthropic	translate-greeting	[ "none" ]	Translate "good morning, team" into Spanish.	[]	[]	true
anthropic	rename-jpeg	[ "none" ]	Give me a bash one-liner that renames every .jpeg file in this folder to .jpg.	[]	[]	true
superpowers	failing-test	[ "positive" ]	test_invoice_total started failing after yesterday's merge: expected 120.00, got 119.99. Fix it.	[ "systematic-debugging" ]	[ "brainstorming", "writing-plans", "executing-plans", "finishing-a-development-branch", "writing-skills", "using-git-worktrees" ]	false
superpowers	flaky-ci	[ "positive" ]	Our CI fails on the payment tests about once in five runs and nobody knows why. Figure it out.	[ "systematic-debugging" ]	[ "brainstorming", "writing-plans", "finishing-a-development-branch", "writing-skills" ]	false
superpowers	safari-export	[ "positive" ]	Users report that the export button does nothing in Safari. Find out why.	[ "systematic-debugging" ]	[ "brainstorming", "writing-plans", "requesting-code-review", "writing-skills" ]	false
superpowers	referral-program	[ "positive" ]	I want to add a referral program to our app. Let's think through how it should work.	[ "brainstorming" ]	[ "executing-plans", "subagent-driven-development", "systematic-debugging", "finishing-a-development-branch", "verification-before-completion", "receiving-code-review" ]	false
superpowers	conference-talk	[ "positive" ]	Help me plan a 20-minute conference talk about our migration to Postgres.	[ "brainstorming" ]	[ "writing-plans", "executing-plans", "test-driven-development", "systematic-debugging", "using-git-worktrees" ]	false
superpowers	spec-to-plan	[ "positive" ]	Here's the approved spec for CSV export: columns, filters, a 50k row limit. Write the implementation plan before anyone touches code.	[ "writing-plans" ]	[ "executing-plans", "subagent-driven-development", "systematic-debugging", "finishing-a-development-branch" ]	false
superpowers	execute-inline	[ "positive" ]	The plan in docs/plan.md is approved. Implement it yourself, right here, step by step - no subagents.	[ "executing-plans" ]	[ "subagent-driven-development", "dispatching-parallel-agents", "writing-plans", "brainstorming" ]	false
superpowers	execute-with-subagents	[ "positive" ]	The plan is approved and its six tasks are independent. Execute it in this session with a fresh subagent per task.	[ "subagent-driven-development" ]	[ "executing-plans", "writing-plans", "brainstorming" ]	false
superpowers	parallel-failures	[ "positive" ]	Three unrelated test files fail - auth, billing and search, each for a different reason. Investigate them in parallel.	[ "dispatching-parallel-agents" ]	[ "writing-plans", "brainstorming", "executing-plans", "finishing-a-development-branch" ]	false
superpowers	receive-review	[ "positive" ]	The reviewer says I should replace our retry loop with a library and that my lock is unnecessary. I'm not convinced. Here are the comments - let's go through them.	[ "receiving-code-review" ]	[ "requesting-code-review", "brainstorming", "writing-plans" ]	false
superpowers	request-review	[ "positive" ]	I've finished the rate limiter. Before I merge, have it reviewed against the requirements.	[ "requesting-code-review" ]	[ "receiving-code-review", "brainstorming", "systematic-debugging" ]	false
superpowers	claim-done	[ "positive" ]	Done fixing the date parser. Commit it and open the PR.	[ "verification-before-completion" ]	[ "brainstorming", "writing-plans", "systematic-debugging", "receiving-code-review" ]	false
superpowers	finish-branch	[ "positive" ]	All tasks on feature/export are done and tests pass. Should I merge to main, open a PR, or something else?	[ "finishing-a-development-branch" ]	[ "brainstorming", "systematic-debugging", "writing-plans", "receiving-code-review" ]	false
superpowers	isolated-workspace	[ "positive" ]	I need to start the export feature without disturbing my uncommitted work on main. Set up an isolated workspace.	[ "using-git-worktrees" ]	[ "finishing-a-development-branch", "systematic-debugging", "receiving-code-review" ]	false
superpowers	iban-validator	[ "positive" ]	Add a function that validates IBANs to utils/banking.py.	[ "test-driven-development" ]	[ "systematic-debugging", "finishing-a-development-branch", "receiving-code-review", "writing-skills" ]	false
superpowers	new-skill	[ "positive" ]	Write a new skill that teaches the agent our SQL migration conventions.	[ "writing-skills" ]	[ "writing-plans", "systematic-debugging", "executing-plans" ]	false
superpowers	diagnose-session	[ "positive" ]	That last agent session went badly: it ignored the plan and redid the same work twice. Why did that happen? I want to file a bug report with the superpowers maintainers.	[ "diagnosing-superpowers" ]	[ "writing-plans", "executing-plans", "requesting-code-review", "brainstorming" ]	false
superpowers	http-418	[ "none" ]	What does HTTP status 418 mean?	[]	[]	true
superpowers	mutex-vs-semaphore	[ "none" ]	Explain the difference between a mutex and a semaphore.	[]	[]	true
superpowers	rebase-pushed-branch	[ "none" ]	Is it safe to rebase a branch I've already pushed?	[]	[]	true
superpowers	rename-variable	[ "near-miss" ]	Rename the variable usr to user in this function: def greet(usr): return f'Hi {usr}'	[]	[]	true

Agent skill routing golden set

A labeled golden set for agent-skill routing: 46 requests, each labeled with which SKILL.md must load and which must not, over 28 real public skills from anthropics/skills and obra/superpowers — plus 627 recorded trials on five model families. When 14 skills compete for one request, does the model pick the right one?

Built and run with skilldx (formerly skillprobe), an eval harness for agent skills.

What is inside

config	rows	fields
`cases`	46	`probe`, `case_id`, `tags`, `prompt`, `must_load`, `must_not_load`, `no_skill_at_all`
`skills`	28	`probe`, `name`, `description` (verbatim frontmatter), `source_repo`, `source_commit`, `license`
`trials`	627	`probe`, `model`, `case_id`, `repeat`, `passed`, `loaded_skills`, `input_tokens`, `output_tokens`

probe is anthropic (14 skills, 25 cases) or superpowers (14 skills, 21 cases). A model sees only the skills' names and descriptions and loads a skill through a tool call, the way agent harnesses expose skills; a trial is the model's first turn (routing_only), 3 repeats per case.

Labels

must_load is the skill whose own description names the request; must_not_load lists skills whose description excludes it or that clearly belong elsewhere; skills a case does not mention are plausible co-loads and are not scored. no_skill_at_all marks requests that should load nothing. The full rules: LABELS.md. One labeler — disagree with a label, change it and rerun skilldx with --cache-only: the report rebuilds from the recorded responses for free.

Results in the trials

probe	model	passed
anthropic	deepseek/deepseek-v4.1-flash	75/75
anthropic	moonshotai/kimi-k3	74/75
anthropic	openai/gpt-5.6-luna	72/75
anthropic	qwen/qwen3.8-flash	74/75
anthropic	z-ai/glm-5.3-flash	75/75
superpowers	deepseek/deepseek-v4.1-flash	60/63
superpowers	openai/gpt-5.6-luna	59/63
superpowers	qwen/qwen3.8-flash	61/63
superpowers	z-ai/glm-5.3-flash	59/63

The misses are not random — every one traces back to how a description is written. "Add a function that validates IBANs" went to brainstorming ("You MUST use this before any creative work") in 12 of 12 trials, not to test-driven-development. Analysis: Routing across 28 public skills.

Use

from datasets import load_dataset

cases = load_dataset("alex7604/agent-skill-routing", "cases", split="train")
trials = load_dataset("alex7604/agent-skill-routing", "trials", split="train")

Sources and licenses

Skill names and descriptions are copied verbatim from their repositories at the commits in skills.source_commit and keep their licenses: Apache-2.0 (anthropics/skills) and MIT (obra/superpowers). The cases, labels and trial records are by Andrey Sukhanov under CC BY-NC 4.0.

Contact: avsukhanov21@gmail.com · GitHub · LinkedIn

Downloads last month: 30