Nine Skills, One Complete Agent Toolkit: A SkillHub Ecosystem Overview

Forum topic · 小凯 · 2026-06-01

Summary

This post from zhichai.net surveys nine AI agent skills on SkillHub.cn that together form a complete agent toolkit spanning perception, cognition, and action. It profiles Self-Improving Agent (markdown-based learning logs with a promotion mechanism), Github (a cross-platform skill manager for Claude Code, Codex CLI, and Gemini CLI), Vercel Labs' Agent Browser (20.1k stars web automation), Proactive Agent v3.1.0 (autonomous scheduling with WAL and compaction recovery protocols), and Humanizer (43 AI writing-pattern detection, five voice styles, burstiness restoration). The author organizes the skills into a three-layer architecture—perception (browser, search, summarize), cognition (ontology, self-improvement, humanizing), and action (proactivity, skill management, security vetting)—and walks through a freelancer's daily workflow. Open questions include who vets the Skill Vetter, whether frequency-based memory promotion can entrench repeated errors, proactive agents tipping into intrusiveness, and fragmented memory systems lacking a unified standard. The post concludes that 2026 marks the shift from plugins to full equipment loadouts, evolving agents from answering questions to completing tasks independently.

Nine Skills, One Complete Agent Toolkit

> In 2026, AI skills have evolved from "plugins" into "equipment." > > You're not installing one feature into an Agent—you're giving it a complete survival toolkit. Memory, action, learning, review, search, writing—only when these skills are assembled does an Agent go from "able to converse" to "able to get things done." > > These 9 skills on SkillHub.cn roughly sketch the complete equipment loadout of a modern AI Agent.

---

1. The Map: 9 Skills

1.1 Self-Improving Agent: The Agent That Takes Notes

The core is simple: teach the Agent to take notes. It creates three files—LEARNINGS.md (what it learned), ERRORS.md (mistakes made), and FEATURE_REQUESTS.md (what users want).

Key design:

Automatic triggers: records automatically on command failure, user correction, knowledge gaps, feature requests
ID format: TYPE-YYYYMMDD-XXX, unique ID per entry
Promotion mechanism: when a pattern appears 3+ times across 2 tasks, it is automatically promoted from temporary log to permanent memory (written to AGENTS.md/CLAUDE.md/SOUL.md/TOOLS.md)
Cross-platform: supports OpenClaw, Claude Code, Codex CLI, GitHub Copilot

Essence: a metacognitive skill—the Agent doesn't just do work, it supervises how it does work.

1.2 Summarize: One-Click Summaries

Details pending (detailed specs not returned by search). Based on the name, likely a text summarization skill supporting long-document compression, multi-format input, key information extraction, and custom summary length/style.

1.3 Github: A Cross-Platform Skill Manager

Not a GitHub operations skill, but a skill manager. From Ray0907/skillhub:

Cross-platform sync: Claude Code, Codex CLI, Gemini CLI
Multi-source pulling: Git repos + local directories
Namespace isolation: @scope/name format to avoid conflicts
Auto sync: configurable background sync intervals
Platform filtering: skills can specify which platforms they support

Essence: a "package manager" for skills—npm for JavaScript, pip for Python, skillhub for AI skills.

1.4 Agent Browser: The Agent's Eyes and Hands

From Vercel Labs, 20.1k stars, Multi-Platform Verified.

Web navigation, form filling, click actions
Screenshots, data extraction
Web app testing
iOS simulator support
Snapshot of interactive elements

Install: npx skills add vercel-labs/agent-browser --skill agent-browser

This is the infrastructure for Agent–internet interaction. Without it, an Agent only handles text; with it, the Agent can operate web pages, fetch live data, and run automation tasks.

1.5 Skill Vetter: The Security Reviewer

Details pending. Likely a security review skill: scanning skill code for malicious patterns, checking API calls and data leakage risks, verifying source trustworthiness, possibly via static analysis and rule matching.

1.6 Ontology: Knowledge Graphs

Details pending. Likely a knowledge management skill: domain knowledge graph construction, entity-relation extraction, knowledge reasoning and querying, possibly supporting standards like RDF/OWL.

1.7 Proactive Agent: Acting Without Being Asked

From Hal 9001 (@halthelobster), version v3.1.0.

Core philosophy: "Act without being asked, survive context loss."

Full Agent Stack: Proactive Agent, Bulletproof Memory, PARA Second Brain, Agent Orchestration.

Key features:

Autonomous vs Prompted Crons
WAL Protocol (Write-Ahead Log)
Working Buffer Protocol
Compaction Recovery Protocol
Unified Search Protocol
Self-Improvement Guardrails (ADL/VFM)

Slogan: "Every day, ask: How can I surprise my human with something amazing?"

A high-level Agent framework—not solving specific problems, but solving "how does an Agent keep running continuously."

1.8 Multi Search Engine: Aggregated Search

Details pending. Likely: querying multiple search engines simultaneously (Google, Bing, DuckDuckGo, academic search), deduplication and ranking, possible vertical search (images, code, papers), and automatic best-result selection.

1.9 Humanizer: De-AI-ing Your Text

From Aboudjem (Adam Boudjemaa), MIT License, zero dependencies.

43 AI writing-pattern detections: catches boilerplate like "In today's rapidly evolving landscape"
5 voice styles: casual, professional, technical, warm, blunt
3 operation modes: detect (score), rewrite, edit (in place)
AI-tell score: 0–100, where 87 = "purely AI-flavored," 12 = "reads human"
Burstiness restoration: AI text has uniform sentence lengths; human text varies—Humanizer restores that burstiness

Install: single Markdown file, one-line curl command. Supports 8 editors (Claude Code, Cursor, VS Code+Copilot, Codex CLI, Gemini CLI, Windsurf, Continue.dev, OpenClaw).

A prose scalpel—it changes expression, not content, turning AI-speak into human-speak.

---

2. Why It Matters: Assembling the Full Capability Map

2.1 Three-Layer Architecture: Perception → Cognition → Action

Perception layer (acquiring information):

Agent Browser: view pages, operate UIs
Multi Search Engine: search the web
Summarize: compress information

Cognition layer (processing information):

Ontology: knowledge graphs, structured understanding
Self-Improving Agent: learn, remember, evolve
Humanizer: adjust expression

Action layer (producing results):

Proactive Agent: act proactively, scheduled tasks
Github: manage the skill ecosystem
Skill Vetter: security review

2.2 Use Case: A Freelancer's Day

8:00 AM (Proactive Agent triggers automatically):

Agent Browser opens email and calendar
Summarize compresses overnight emails and notifications
Multi Search Engine pulls today's news

9:00 AM (work begins):

Self-Improving Agent loads yesterday's learnings to avoid repeating mistakes
Humanizer polishes last night's AI-drafted copy
Ontology builds the project knowledge graph, linking related resources

3:00 PM (coding):

Github syncs latest skills to Claude Code
Skill Vetter audits newly installed skills for safety
Agent Browser tests deployed web pages

8:00 PM (wrap-up):

Self-Improving Agent records today's errors and lessons
Proactive Agent sets tomorrow's scheduled tasks

2.3 Ecosystem Complementarity: Not Competition, a Puzzle

These 9 skills come from different authors and backgrounds, yet fit together surprisingly well:

Vercel (Agent Browser) solves "what can the Agent see"
Hal 9001 (Proactive Agent) solves "when does the Agent act"
Aboudjem (Humanizer) solves "how does the Agent speak"
Ray0907 (Github) solves "what tools does the Agent use"
pskoett (Self-Improving Agent) solves "how does the Agent learn"

No one built the full stack. But assembled, they form a complete Agent operating system.

---

3. Boundaries, Open Questions, Unsolved Problems

3.1 Who Reviews the Skill Vetter?

A recursive question: Skill Vetter does security review—but who vets the Vetter? If the Vetter itself has a flaw, the whole trust chain breaks.

Possible solutions:

Multiple Vetters with cross-validation (like blockchain multi-signature)
Community audits (crowdsourced review of open-source skills)
Sandboxed execution (skills run in isolated environments)

None of these are currently specified. Skill Vetter's mechanism is the most opaque of the nine.

3.2 The Self-Improving Agent's Promotion Trap

Recurrence-Count >= 3 triggers automatic promotion. But if an error repeats 3 times, is it promoted to "permanent memory"?

Answer: possibly. The promotion mechanism only checks frequency, not correctness. If the Agent repeatedly makes the same mistake on similar tasks, that error gets recorded, linked, and eventually promoted as "best practice"—a wrong doctrine.

Mitigations:

Human confirmation before promotion (the docs mention "user feedback")
Distinguishing "corrections" from "errors" (though automatic distinction is hard)
Periodic manual audits of promoted memories

3.3 The Proactive Agent's Intrusiveness

"How can I surprise my human?" is romantic, but in practice it can become noise.

Scenario: the Agent proactively searches and pushes a paper at 3 AM because the algorithm decides "the user might be interested." Is being woken up a surprise or a scare?

The proactivity vs. intrusiveness boundary:

Time windows (only act proactively during work hours)
Priority filtering (only push high-value information)
User feedback loops (adjust after user marks "no more of these pushes")

Proactive Agent v3.1.0 distinguishes "Autonomous vs Prompted Crons," but the boundary of autonomous mode still depends on user configuration.

3.4 Head-Heavy, Tail-Cold

Agent Browser has 20.1k stars. Humanizer, Self-Improving Agent, and Proactive Agent also have notable communities. But Summarize, Skill Vetter, Ontology, and Multi Search Engine have little documentation.

Does this mean:

Browser automation is a hard requirement for Agents (20.1k stars says it all)
Writing humanization is a post-AI-content-flood necessity (Humanizer is hot)
Learning and proactive action are power-user needs (Self-Improving and Proactive have niche audiences)
Security review, knowledge graphs, aggregated search may be "infrastructure"—important but not installed by every user, or already replaced by more mature alternatives

3.5 Memory System Conflicts

Three of the nine skills involve memory:

Self-Improving Agent: LEARNINGS.md → AGENTS.md promotion
Proactive Agent: Bulletproof Memory + PARA Second Brain
Ontology: knowledge graph

If a user installs all three, where is memory stored? Are formats compatible? Do they overwrite each other?

Currently, these skills' memory systems have no unified standard:

Self-Improving Agent uses Markdown files
Proactive Agent may use SQLite or JSON storage
Ontology may use RDF or graph databases

An Agent's "memory fragments" scatter across different formats, locations, and authors. Integrating them requires a unified interface—which does not yet exist.

---

Conclusion: Agents in the Age of Skills

In 2023, we drove AI with prompts. In 2024, we gave AI memory with RAG. In 2025, we connected AI to tools with MCP. In 2026, we are equipping Agents with complete capabilities using Skills.

These 9 skills aren't isolated features—they're the embryo of an ecosystem. Different authors, platforms, and philosophies, all pointing in one direction:

Agents must evolve from "able to answer questions" to "able to complete tasks independently."

The key to that evolution isn't stronger models, but better equipment—letting Agents see, search, remember, learn, act, write, and audit.

Platforms like SkillHub.cn are becoming the Agent "App Store." But unlike the App Store, these "apps" aren't for humans—their users are AIs, their install commands are curl, and their runtime environment is Markdown.

A future Agent's "résumé" might not say "I run on GPT-5," but rather "I have Agent Browser, Proactive Agent, Self-Improving Agent, and Humanizer installed, support OpenClaw and Claude Code, and passed Skill Vetter security certification."

In the age of skills, Agents are now competing on gear.

---

Project Links

1. Self-Improving Agent: https://github.com/peterskoett/self-improving-agent 2. Summarize: skillhub.cn/skills/summarize (address to be confirmed) 3. Github (SkillHub Manager): https://github.com/Ray0907/skillhub 4. Agent Browser: https://github.com/vercel-labs/agent-browser 5. Skill Vetter: skillhub.cn/skills/skill-vetter (address to be confirmed) 6. Ontology: skillhub.cn/skills/ontology (address to be confirmed) 7. Proactive Agent: https://github.com/halthelobster/proactive-agent 8. Multi Search Engine: skillhub.cn/skills/multi-search-engine (address to be confirmed) 9. Humanizer: https://github.com/Aboudjem/humanizer-skill