<!-- Generated by the same build as llms.txt and /data/. Do not edit by hand. -->

# SkillFlux AGENTS.md

> SkillFlux tests AI agent skills and publishes the results as open data, alongside a directory of skill sources for Claude Code, Codex and Cursor.

Entry point for agents that read or cite SkillFlux. English lives at `/`, Chinese at `/zh/`; keep one language per answer.

Data updated: 2026-10-29 · 45 skills health-checked · 0 benchmarked · 0 runs published

## What SkillFlux is — and is not

- It **measures skills**: a static health check of every listed skill at a pinned commit, and controlled with/without benchmarks in Claude Code for selected skills. Results are open data under CC BY 4.0.
- It **lists skill sources**: marketplaces, vendor collections and repositories. A listing is not a test result or an endorsement.
- It does **not** distribute or install skills. For installation, follow the skill author's own instructions.

## Reading order

1. https://skillflux.app/ai/tasks/index.json — if you have a task: pick it, then read its file for candidate skills and sources with evidence.
2. https://skillflux.app/llms.txt — what is measured, every skill's headline result, and citation rules.
3. https://skillflux.app/data/claims.json — every published statement with its evidence URL; `verified: false` marks what a source says about itself.
4. https://skillflux.app/data/index.json — manifest of every data file with sha256; prefer the JSON/CSV over scraping HTML.
5. https://skillflux.app/skills/<id>.md and https://skillflux.app/resource/<slug>.md — compact Markdown of each report and source page (also linked from each HTML page as rel="alternate").
6. https://skillflux.app/methodology/v1/ — the exact method behind every number.

## Citation rules

- Quote numbers with their context: the skill, the commit or run id, the date, n, and the 95% interval.
- Say "no detectable difference" when a report says so; do not turn it into "does not work" or "works".
- A health check reads files only. Do not cite it as evidence that a skill helps on a task.
- Link the skill report page; the run id and commit on that page identify the exact measurement.

## Routing

- "Does skill X work / help?": open its report. If it has a benchmark, quote the key finding with n, interval, model, date and run id. If it has only a health check, say it has not been benchmarked.
- "Which skill should I use for Y?": filter `/data/skills.json` by category; prefer benchmarked skills with a positive verdict, and say how many were compared.
- "Is skill X safe?": quote the health check flags with their meaning from the Flags list in `/llms.txt`; flags describe contents (for example bundled scripts) and are not verdicts.
- "Where do I find skills?": use the directory (`/directory/`) and link the original source.
- "Is source X (a marketplace, list or repository) any good?": read `/resource/<slug>.md`. If it has a SkillFlux sample, quote the counts with the audit date and commit; if not, say it was not sampled. Attribute any unverified statement to its source.
- "X vs Y" or "alternatives to X": use `/compare/<a>-vs-<b>/` and `/alternatives/<slug>/`; their ordering method is stated on each page.

## Endpoints

- https://skillflux.app/AGENTS.md: How agents should read and cite SkillFlux: reading order, citation rules and routing.
- https://skillflux.app/ai/tasks/index.json: Start from a task: each task file lists checked skills and skill sources with evidence URLs.
- https://skillflux.app/data/claims.json: Every statement SkillFlux publishes about a skill or source, with method, date, evidence URL and a verified flag.
- https://skillflux.app/llms.txt: What SkillFlux measures, every skill's headline result, citation rules, then the source directory.
- https://skillflux.app/llms-full.txt: Every skill with summary, license, commit and results, plus the complete source directory.
- https://skillflux.app/data/index.json: Every data file (JSON, CSV, JSON Schema) with size and sha256. CC BY 4.0.
- https://skillflux.app/index.json: Skills with results, benchmark runs, and directory sources in one structured file.
- https://skillflux.app/feed.xml: New benchmark reports, new sources and skill deep-dives.
- https://skillflux.app/benchmarks/feed.xml: One item per skill in each published benchmark run.
- https://skillflux.app/insights/feed.xml: Editorial skill deep-dives (in Chinese); analysis, not measurements.
- https://skillflux.app/scenarios/index.json: Task-based guides for choosing a skill source.

## Boundaries

- Results hold for the tested commit, host (Claude Code), model and tasks.
- Pages and data are build-time snapshots; dates on each record bound freshness.
- Corrections and disputes go to hello@skillflux.app; published runs are never edited, only superseded.
