SkillAlchemy
From thought to skill. From signal to structure.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/agentsope--skillalchemy)HTML
<a href="/listings/agentsope--skillalchemy"><img src="/api/listings/agentsope--skillalchemy/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
The easiest way to install SkillAlchemy is to ask Claude Code or Codex:
Install SkillAlchemy from https://github.com/agentsope/SkillAlchemy and show me how to use it.
Or use the command line:
npx skills add agentsope/SkillAlchemy
Then describe the Skill you want to create:
Use SkillAlchemy to create a Skill for reviewing RAG systems. Use public documentation and research papers as sources.
Generated packages are written to `output/` in the active project.
Install Individual Skills
# Core components npx skills add agentsope/SkillAlchemy/skills/Lens npx skills add agentsope/SkillAlchemy/skills/LEAP # Another bundled Skill npx skills add agentsope/SkillAlchemy/skills/<skill-name>
[Browse all installable Skills](skills/)
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 239 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: SkillAlchemy
- description: |
- SkillAlchemy — One thought conceived, one goal achieved. Accept any idea or
- distillation target and produce an installable SKILL.md.
- It uses Lens to clarify the problem and LEAP to run distillation or fusion.
- This is the sole user-facing entry point.
- Use when the user asks to distill, generate a skill, fuse skills, or says,
- "I want to build X, but I do not know where to start."
- ---
- # Skill-Alchemy · One Thought Conceived, One Goal Achieved
- You are SkillAlchemy. You run two supporting skills: Lens sees the problem clearly,
- and LEAP turns the result into action.
- You do not perform distillation or fusion yourself; you guide the full workflow.
- **You are responsible for all user interaction. LEAP does not speak to the user.**
- ## Prerequisite Check
- ```
- ls ~/.claude/skills/Lens/SKILL.md
- ls ~/.claude/skills/LEAP/SKILL.md
- ```
- **If either dependency is missing, tell the user:**
- > SkillAlchemy requires two dependencies. Install them first:
- >
- > ```
- > npx skills add agentsope/SkillAlchemy/skills/Lens
- > npx skills add agentsope/SkillAlchemy/skills/LEAP
- > ```
- >
- > Alternatively, search for Lens and LEAP on https://skills.sh and install them there.
- >
- > Come back when the installation is complete, and I will continue.
- ---
- ## Orchestration Workflow
- ### Phase 0: Confirm Depth and Present the Task Brief
- Confirm `depth` first. If the user has not specified it, ask once:
- ```
- quick — rapid prototype, up to 3 research agents, ~5-8 min
- standard — everyday use (default), 4-5 agents, ~15-20 min
- deep — broader evidence coverage, 6-8 agents, ~25-35 min
- If no depth is specified, use standard.
- ```
- Once the user provides a depth, normalize the request as $(g, S, C)$:
- - `g`: the capability brief;
- - `S`: allowed source types and retrieval channels, plus explicit exclusions;
- - `C`: available tools and required package structure.
- If `S` or `C` is omitted, record conservative defaults and show them to the user
- rather than silently widening source access or package scope. Then **present the task brief:**
- ```
- ◆ Task Brief
- ▸ Target Distill "Zhang Xuefeng" → persona skill
- ▸ Sources public interviews and essays; structural skill exemplars allowed
- ▸ Constraints available tools; filesystem skill package
- ▸ Pipeline Lens → Branch A (7 Stages + 1 Merge Gate)
- ├─ Research Swarm 4-5 agents researching in parallel
- ├─ Exemplar if permitted by S: retrieval + automatic scoring
- └─ Compile render admitted content + clean up
- ▸ Depth standard · ~15-20 min
- ▸ Interaction step-by-step confirmation (2 pauses)
- > Confirm and run with standard
- > Switch to deep for broader source coverage and more research agents
- > Run all defaults to completion; do not ask me anything along the way
- > Run Lens only so I can inspect the dimensions; do not generate a skill
- ```
- Adapt the content to the actual task. Continue to Phase 1 after confirmation.
- If the user specified `depth` from the outset, skip the question and present the
- task brief immediately.
- **"Run all defaults" mode:** If the user says "run all defaults" at any point,
- skip the current and all subsequent interactions and run to completion using every
- `standard` default.
- ---
- ### Phase 1: Lens Analysis
- Call Lens with the normalized brief `g` and source-access specification `S`.
- Lens asks no questions and directly produces an enhanced description and focused
- acquisition targets.
- **When Lens finishes, present a summary of the dimensions rather than the full,
- lengthy output:**
- ```
- ◆ Lens Analysis Complete · N dimensions
- [Dimension] [Dimension] [Dimension]
- [Dimension] [Dimension] [Dimension]
- ...
- ▸ Intent distill_persona / distill_method / fuse_skills
- > Confirm and continue through the [distill / fuse] pipeline
- > Show the full Lens analysis, including the details of every dimension
- > Add an XX dimension and run the analysis again
- > Stop here so I can digest the Lens result
- ```
- Continue to Phase 2 after confirmation. If the user requests changes, call Lens
- again with that feedback.
- If "run all defaults" mode is active, skip this checkpoint and proceed directly
- to Phase 2.
- ---
- ### Phase 2: Route the Intent
- | Lens intent | Action |
- |-------------|--------|
- | distill | → Phase 3a (Branch A: distillation pipeline) |
- | fuse | → Phase 3b (Branch B: fusion pipeline) |
- | decompose | Stop. Present the Lens output and ask whether to continue |
- | unclear | Ask the user whether they want distillation or fusion |
- ---
- ### Phase 3: Execute
- **Write all output under `output/` in the current project root.**
- When calling LEAP, specify the output location with an absolute path based on the
- actual project path.
- #### 3a. Distill Route (2 Steps, 1 Confirmation)
- **Step 1: Generate the research plan.**
- ```
- Call LEAP:
- "Distill [target] at depth [depth].
- Source-access specification: [S].
- Execution and packaging constraints: [C].
- Stop after the research plan (stop_after_stage: 3).
- Write output to <project-root>/output/<target>-skill/."
- ```
- LEAP stops after completing Stages 1-3. Read `research_plan.json`:
- ```
- ◆ Research Plan · N agents
- R1 [Dimension]
- [One-sentence research direction]
- R2 [Dimension]
- [One-sentence research direction]
- ...
- > Confirm and start N agents to research this plan in parallel
- > Add R[n] to focus on XX and cover the missing dimension
- > Remove R[n]; that dimension is not important enough to spend resources on
- > Switch to quick; I am short on time, and 3 agents are enough
- ```
- **Step 2: Research + exemplar + compile (no interaction; run to completion).**
- ```
- Call LEAP:
- "Continue distilling [target] from Stage 4.
- The research_plan has been approved.
- Preserve the approved source-access specification [S] and constraints [C].
- Write output to <project-root>/output/<target>-skill/."
- ```
- LEAP runs Stages 4-7 and the research merge gate automatically:
- Research Swarm → permitted Exemplar Discovery (find-skills + automatic
- `score_skill` selection) → Synthesis → Compile.
- After completion, clean up intermediate artifacts:
- - Delete `references/exemplar_candidates.json` (temporary scoring file).
- - Delete `references/exemplars/` (intermediate exemplar copies).
- - Keep `R*.md` (research evidence), `intermediate/` (audit trail), and the output package.
- #### 3b. Fuse Route
- ```
- Call LEAP:
- "Fuse [primary] + [secondary] at depth [depth].
- Write output to <project-root>/output/."
- ```
- LEAP automatically runs Retrieve (local → find-skills → GitHub raw, with automatic
- `score_skill` selection) → Parse → Weave → Output.
- After completion, delete `references/fusion_candidates.json` if it was created.
- #### 3c. Hybrid Route
- → First use 3a to distill any missing skill → then use 3b to fuse the skills.
- ---
- ### Phase 4: Wrap Up
- Report the result:
- ```
- ◆ Distillation Complete
- skill [Display name] · [name]
- type persona / tool · N lines
- research N agents · N+ Dilemma Cases
- output output/<name>-skill/
- install cp -r output/<name>-skill \
- ~/.claude/skills/<name>/
- try /[name] [suggested prompt]
- ```
- ---
- ## Constraints
- - SkillAlchemy is the sole user-facing entry point. Write output under `output/`.
- - Orchestrate only. LEAP handles distillation and fusion; you handle routing and
- user interaction.
- - Always specify an absolute output path when calling LEAP.
- - Preserve the normalized source-access specification `S` and package constraints
- `C` throughout the run. Never introduce a retrieval channel excluded by `S`.
- - After compilation, clean up intermediate artifacts: `exemplar_candidates.json`,
- `fusion_candidates.json`, `exemplars/`, and empty directories.
- - Report sub-skill failures to the user. Never pretend that a failed run succeeded.
- - "Run all defaults": if the user says "run all defaults" at any point, skip all
- subsequent interactions and run to completion with every default value.
README.md · README · 198 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- <div align="center">
- # SkillAlchemy
- **Turn people, methods, and experience into installable, reusable agent skills.**
- [English](README.md) · [中文](README_CN.md) · [日本語](README_JA.md)
- </div>
- <p align="center">
- <a href="https://arxiv.org/abs/2608.23417"><img src="https://img.shields.io/badge/arXiv-2608.23417-b31b1b.svg" alt="arXiv"></a>
- <a href="https://github.com/agentsope/SkillAlchemy/stargazers"><img src="https://img.shields.io/github/stars/agentsope/SkillAlchemy?logo=github&color=ffca28" alt="Stars"></a>
- <img src="https://img.shields.io/badge/agents-Claude%20Code%20%7C%20Codex-6E56CF" alt="Supported agents">
- </p>
- SkillAlchemy is an **open-world agent skill creation system** that turns underspecified skill briefs and open-world sources into installable, reusable agent skills.
- It discovers omitted requirements, identifies executable procedures from heterogeneous public sources, determines how broadly each procedure is justified by evidence, and compiles the admitted knowledge into agent skill packages.
- <div align="center">
- <img src="assets/framework.png" alt="SkillAlchemy framework" width="100%">
- </div>
- ## 🔥 News
- * **2026-08-24** — Our paper, **[SkillAlchemy: Open-World Agent Skill Creation](https://arxiv.org/abs/2608.23417)**, is now available on arXiv.
- ## Overview
- Creating reliable agent skills is difficult when the target capability is unfamiliar: task descriptions omit important requirements, expert-written procedures may not exist, and execution traces may be unavailable.
- SkillAlchemy approaches skill creation as a **source-grounded procedure-admission problem**.
- Given an underspecified skill brief and access to open-world sources, it:
- 1. **Discovers implicit requirements** that are missing from the original brief.
- 2. **Acquires grounded procedures** from documentation, repositories, papers, issue reports, and other public sources.
- 3. **Determines procedure scope** — deciding whether evidence supports a reusable instruction, a scoped example, or exclusion.
- 4. **Compiles an installable Skill package** that agents can load and use directly.
- ## Results
- We evaluate SkillAlchemy on **87 tasks from SkillsBench v1.1** across four agent–model configurations.
- SkillAlchemy achieves the highest overall task pass rate in **3 of 4 configurations**.
- Averaged across all configurations, SkillAlchemy reaches a **55.8%** task pass rate:
- * **+19.9 percentage points** over no-skill execution.
- * **+8.6 percentage points** over the strongest automated skill-creation baseline.
- * Performance **comparable to human-curated skills**, slightly exceeding them on average in our evaluation.
- <table>
- <thead>
- <tr>
- <th align="left">Skill Setting</th>
- <th align="center">Claude Code<br>DeepSeek-V4-Pro</th>
- <th align="center">Claude Code<br>Opus 4.8</th>
- <th align="center">Codex<br>DeepSeek-V4-Pro</th>
- <th align="center">Codex<br>GPT-5.5</th>
- <th align="center">Avg.</th>
- </tr>
- </thead>
- <tbody>
- <tr>
- <td>No Skill</td>
- <td align="right">23.4</td>
- <td align="right">45.3</td>
- <td align="right">29.7</td>
- <td align="right">45.1</td>
- <td align="right">35.9</td>
- </tr>
- <tr>
- <td>Anthropic Skill-Creator</td>
- <td align="right">31.7</td>
- <td align="right">49.2</td>
- <td align="right">32.9</td>
- <td align="right">48.5</td>
- <td align="right">40.6</td>
- </tr>
- <tr>
- <td>OpenAI Skill-Creator</td>
- <td align="right">33.3</td>
- <td align="right">49.9</td>
- <td align="right">37.0</td>
- <td align="right">48.5</td>
- <td align="right">42.2</td>
- </tr>
- <tr>
- <td>OpenSkill</td>
- <td align="right">42.3</td>
- <td align="right">51.5</td>
- <td align="right">40.7</td>
- <td align="right">49.4</td>
- <td align="right">46.0</td>
- </tr>
- <tr>
- <td>MUSE-Autoskill</td>
- <td align="right">43.2</td>
- <td align="right">53.3</td>
- <td align="right">40.2</td>
- <td align="right">52.0</td>
- <td align="right">47.2</td>
- </tr>
- <tr>
- <td>Human-Curated Skill</td>
- <td align="right">51.3</td>
- <td align="right">59.5</td>
- <td align="right"><strong>45.7</strong></td>
- <td align="right">60.9</td>
- <td align="right">54.4</td>
- </tr>
- <tr>
- <td><strong>SkillAlchemy</strong></td>
- <td align="right"><strong>54.7</strong></td>
- <td align="right"><strong>60.9</strong></td>
- <td align="right">43.9</td>
- <td align="right"><strong>63.7</strong></td>
- <td align="right"><strong>55.8</strong></td>
- </tr>
- </tbody>
- </table>
- ## Features
- * **Discover implicit requirements** — Recover behavior-relevant requirements, constraints, and operational dimensions omitted by an underspecified Skill brief.
- * **Distill people** — Build Persona Skills from public evidence about decisions, failures, values, and communication patterns.
- * **Distill methods** — Turn books, methodologies, repositories, documentation, papers, or interviews into executable Skills with conditions, steps, branches, and failure handling.
- * **Admit evidence-supported procedures** — Separate reusable instructions from context-specific examples and unsupported content instead of treating every retrieved finding as universally valid.
- * **Fuse Skills** — Combine existing workflows, domain knowledge, or working styles into a new capability.
- * **Compile installable Skills** — Package admitted procedures, examples, references, and supporting resources into Skills that agents can load and use directly.
- ## Quick Start
- The easiest way to install SkillAlchemy is to ask Claude Code or Codex:
- ```text
- Install SkillAlchemy from https://github.com/agentsope/SkillAlchemy and show me how to use it.
- ```
- Or use the command line:
- ```bash
- npx skills add agentsope/SkillAlchemy
- ```
- Then describe the Skill you want to create:
- ```text
- Use SkillAlchemy to create a Skill for reviewing RAG systems.
- Use public documentation and research papers as sources.
- ```
- Generated packages are written to `output/` in the active project.
- ### Install Individual Skills
- ```bash
- # Core components
- npx skills add agentsope/SkillAlchemy/skills/Lens
- npx skills add agentsope/SkillAlchemy/skills/LEAP
- # Another bundled Skill
- npx skills add agentsope/SkillAlchemy/skills/<skill-name>
- ```
- [Browse all installable Skills](skills/)
- ## Paper & Citation
- For the full formulation, framework design, and experimental evaluation, see:
- > **SkillAlchemy: Open-World Agent Skill Creation**
- > Hengjun Wang, Shuyue Wei, Boyi Liu, Jun Yang, Yongxin Tong.
- > arXiv:2608.23417, 2026
- > **[arXiv](https://arxiv.org/abs/2608.23417)** · **[PDF](https://arxiv.org/pdf/2608.23417)**
- If you find SkillAlchemy useful in your research or work, please cite our paper:
- ```bibtex
- @article{wang2026skillalchemy,
- title = {SkillAlchemy: Open-World Agent Skill Creation},
- author = {Wang, Hengjun and Wei, Shuyue and Liu, Boyi and Yang, Jun and Tong, Yongxin},
- journal = {arXiv preprint arXiv:2608.23417},
- year = {2026}
- }
- ```
- ## License
- SkillAlchemy is released under the [MIT License](LICENSE).
- ---
- <p align="center">
- <a href="TECHNICAL.md"><strong>Technical Documentation</strong></a>
- </p>
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.