ClawHub's #1 Downloaded Skill: Capability Evolver

Let your AI Agent evolve itself

---

Why Is It the #1 Download?

Capability Evolver solves a blunt problem: AI gets dumber over time — how do you make it keep getting smarter?

With a normal AI Agent, everything learned in a session is wiped when the conversation ends. Mistakes get repeated.

Once Capability Evolver is installed, the Agent actively does three things:

1. Records every problem encountered and how it was solved
2. Automatically retrieves relevant records when a similar problem comes up again
3. Periodically promotes "temporary learnings" to "permanent rules"

---

Installation

clawhub install capability-evolver

After installation, a .capabilities/ directory is created in your workspace with three files:

• LEARNINGS.md — temporary learning records
• PATTERNS.md — identified behavioral patterns
• RULES.md — permanently promoted rules

---

Real-World Results

Scenario	Before	After
Repeated mistakes	Every time	Automatically avoided on 2nd occurrence
Tool usage	Re-learned from scratch	Reuses prior experience
User preferences	Re-explained each time	Auto-remembered and applied

---

Difference From self-improving-agent

The two skills are often confused:

• self-improving-agent: Lightweight — only handles error logging and corrections
• Capability Evolver: Full-featured — includes pattern recognition, rule promotion, and cross-agent sharing

SFD team's approach: install both. capability-evolver handles the "strategy layer," self-improving-agent handles the "error layer."

---

SFD Team Experience

All 13 of our Agents have had this skill installed for about two months. Key observations:

1. Promotion requires manual confirmation: The skill suggests "Should this learning be promoted to a rule?" but the actual promotion requires human approval — prevents bad experiences from becoming permanent rules.
2. Isolate learning files per Agent: We keep each Agent's .capabilities/ in its own workspace to avoid cross-contamination.
3. Review RULES.md periodically: Every two weeks, prune outdated rules. Otherwise they accumulate and become a burden.

---

Pitfalls

Pitfall 1: Rule conflicts — When two rules contradict each other, the Agent randomly picks one. Solution: Add priority labels (P1/P2/P3) before each rule.

Pitfall 2: Context bloat — A long RULES.md eats up your context window. Keep individual rules under 50 words, total rules under 20.

Pitfall 3: Poor Claude Haiku support — Rule promotion works unreliably on Haiku. Use Sonnet or better.

---

Recommendation

⭐⭐⭐⭐⭐

If you use AI seriously, this skill is worth the setup time.

---

SFD Lab | Tested: 2026-02~03 | 13 Agents in production