Best AI Skills for AI & Machine Learning in 2026: 15 Compared
Fifteen skills for building, running, and improving AI agents — ranked across 465 candidates. The dominant 2026 pattern: agent memory and self-improvement. 10 of the top 15 are about either remembering or evolving.
The AI & Machine Learning category has 465 skills in our catalog and is moving the fastest of any. A quarter ago, the top of this list would have been dominated by "LLM helpers" — wrappers around a single model call. As of May 2026, the leaderboard has shifted decisively: most of the top 15 are about agent memory, self-improvement, or proactive behavior. The skills people install are the ones that make agents remember, reflect, and improve — not the ones that make a single inference cheaper or faster. That tells you something about where the bottleneck has moved.
Quick Pick
brainstorming — the highest-quality skill in this category (4/5 rating) and the one obra/superpowers itself marks as "you MUST use before any creative work." It forces requirement-and-design exploration before any code, which is the single biggest determinant of whether AI-generated features hit the mark or drift.
What These Skills Actually Do
AI & Machine Learning skills sit on top of other skills — they're meta-capabilities that change how all your other agent work behaves. The category breaks into four clear sub-themes:
- Pre-implementation reasoning — brainstorming, structured ideation, ontology-building before any code
- Agent memory — persistent context across sessions, vector-searchable history, recovery from "goldfish brain"
- Self-improvement and reflection — agents that evaluate their own work, capture learnings, evolve protocols
- Proactive behavior — agents that anticipate rather than react
The skills that rank high all sit at the infrastructure layer of agent workflows — they shape every downstream task. The ones that don't make the cut tend to be domain-specific model wrappers (a Whisper helper, a vision-model helper) that work fine but only help one task at a time. The meta-skills compound; the wrappers don't.
How We Ranked
We sorted 15 candidate skills by a composite score:
- Popularity signal — the highest of GitHub stars, install count, or ClawHub download count. Log-scaled so a 100-star skill doesn't get buried under a 100,000-star one if the smaller one is meaningfully better.
- Quality score — when set, a 0–5 rubric that breaks ties within popularity tiers. Roughly 15% of catalog skills carry a quality score today; we surface it in the comparison table when available.
The formula is identical across the entire Best-Of 2026 series, so you can compare apples to apples between categories.
The Top 15
1. brainstorming
Skill · obra/superpowers · 167.5K signal · quality 4/5 You MUST use this before any creative work — creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.
The take: The structural answer to the "I have an idea → immediately build prototype" trap. The "MUST use before any creative work" framing is unusual: it's a mandatory meta-skill, not an optional one. Pairs perfectly with the 42% problem lesson — slowing the start saves the project.
2. dispatching-parallel-agents
Skill · obra/superpowers · 167.5K signal · quality 4/5 Use when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies.
The take: Parallel agent dispatch was the manual move of the year — people teaching themselves to spin up multiple Claude Code sessions on independent branches. This skill encodes the discipline: how to decide what's actually independent (most things aren't), what to dispatch, and how to merge the results.
3. using-superpowers
Skill · obra/superpowers · 167.5K signal · quality 4/5 Use when starting any conversation — establishes how to find and use skills, requiring Skill tool invocation before ANY response including clarifying questions.
The take: A meta-skill about skills. Forces the agent to check whether a skill applies before responding — which sounds obvious until you realize most agents skip this and respond from first principles. Install once and the rest of your skill investment compounds.
4. Self Improving Agent
Skill · clawhub/self-improving-agent · 240.0K signal · quality unrated Captures learnings, errors, and corrections to enable continuous improvement. Plus 50+ models for image generation, video generation, text-to-speech, speech-to-text.
The take: The most-installed skill in this category by a wide margin (240K signal). The "learnings + errors + corrections" capture is the core; the 50+ model bundle is the cherry on top. If you've ever watched Claude make the same mistake three times in a week, this is the skill that breaks the loop.
5. ontology
Skill · oswalpalash/ontology · 149.0K signal · quality unrated Typed knowledge graph for structured agent memory and composable skills. Use when creating/querying entities (Person, Project, Task, Event, Document), linking them, and reasoning over the graph.
The take: Different theory of memory from the vector-search-based skills below. Knowledge graphs are typed and queryable in ways embeddings aren't — better when your agent's workflow involves real entities (customers, deals, tickets) rather than free-form notes.
6. Proactive Agent
Skill · halthelobster/proactive-agent · 132.0K signal · quality unrated Transform AI agents from task-followers into proactive partners that anticipate needs and continuously improve. Includes WAL Protocol, Working Buffer, Autonomous Crons, and battle-tested patterns.
The take: The "proactive" framing is the differentiator. Most agents wait for a prompt; this one schedules its own work via the Autonomous Crons feature. WAL (write-ahead log) protocol means proactive moves are recoverable if something goes wrong. Pair with #4 for the full self-improving + proactive stack.
7. Self-Improving + Proactive Agent
Skill · clawhub/self-improving-proactive-agent · 80.0K signal · quality unrated Self-reflection + Self-criticism + Self-learning + Self-organizing memory. Agent evaluates its own work, catches mistakes, and improves permanently.
The take: The bundled version of #4 and #6. Easier install path if you know you want both. Slightly less battle-tested than installing the two originals separately, but worth the convenience for a starter setup.
8. Openai Whisper
Skill · steipete/openai-whisper · 63.1K signal · quality unrated Local speech-to-text with the Whisper CLI (no API key).
The take: Local-Whisper-without-an-API-key is the workflow most "AI voice" skills don't bother with. Worth the install if you care about privacy, are batch-processing audio, or want to avoid OpenAI's per-minute pricing. The only "single-task wrapper" in the top 15 — earns its spot because no-API-key offline operation is genuinely rare.
9. Evolver
Skill · autogame-17/evolver · 51.1K signal · quality unrated A self-evolution engine for AI agents. Analyzes runtime history to identify improvements and applies protocol-constrained evolution.
The take: Heavier than #4 — analyzes runtime history across many sessions rather than just capturing single corrections. The "protocol-constrained" framing matters: evolution within guardrails, not unconstrained behavior drift. Strongest fit for long-running agents in production.
10. Free Ride - Unlimited free AI
Skill · shaivpidadi/free-ride · 50.8K signal · quality unrated Manages free AI models from OpenRouter for OpenClaw. Automatically ranks models by quality, configures fallbacks for rate-limit handling.
The take: A pragmatic skill for a specific audience — OpenClaw users trying to stay on OpenRouter's free tier. The auto-ranking + fallback logic handles the messy reality of rate-limited free models. Niche but well-built.
11. Elite Longterm Memory
Skill · clawhub/elite-longterm-memory · 46.1K signal · quality unrated Ultimate AI agent memory system for Cursor, Claude, ChatGPT & Copilot. WAL protocol + vector search + git-notes + cloud backup. Never lose context again.
The take: The most-feature-complete memory skill — WAL + vector + git-notes + cloud backup is the full stack. Multi-tool support (Cursor + Claude + ChatGPT + Copilot) means your memory follows you across surfaces. Heaviest install of the memory skills but the most resilient.
12. ByteRover
Skill · byteroverinc/byterover · 33.6K signal · quality unrated
Knowledge management for AI agents. Use brv to store and retrieve project patterns, decisions, and learnings.
The take: Built around a CLI primitive (brv) rather than a config file, which gives it a different ergonomic feel. Worth the install if you've already settled into a terminal-first workflow.
13. Session-logs
Skill · guogang1024/session-logs · 32.2K signal · quality unrated Search and analyze your own session logs (older/parent conversations) using jq.
The take: Backwards-looking memory — not "remember things you learn" but "find that thing you discussed three sessions ago." Using jq is a delight when you're already comfortable with it. Niche pick for power users.
14. Memory Setup
Skill · jrbobbyhansen-pixel/memory-setup · 31.2K signal · quality unrated Enable and configure Moltbot/Clawdbot memory search for persistent context. Fixes "goldfish brain" and helps users configure memorySearch in their config.
The take: The setup-and-config skill — useful as a one-shot when you're moving to a memory-enabled workflow but you've never wired one up before. Less valuable after you've got memory running; install it once and uninstall.
15. Evolver
Skill · autogame-17/capability-evolver · 30.3K signal · quality unrated A self-evolution engine for AI agents. Analyzes runtime history to identify improvements and applies protocol-constrained evolution.
The take: A capability-focused fork of #9 from the same author. Subtle distinction: evolves capabilities (what the agent can do) rather than just behaviors (how it does what it already does). Install one or the other depending on whether you're hardening behavior or expanding capability.
Comparison Table
| # | Skill | Type | Stars / Installs | Quality | License |
|---|---|---|---|---|---|
| 1 | brainstorming | Skill | 167.5K | 4/5 | — |
| 2 | dispatching-parallel-agents | Skill | 167.5K | 4/5 | — |
| 3 | using-superpowers | Skill | 167.5K | 4/5 | — |
| 4 | Self Improving Agent | Skill | 240.0K | — | MIT-0 (Free to use, modify, and redistribute. No a |
| 5 | ontology | Skill | 149.0K | — | — |
| 6 | Proactive Agent | Skill | 132.0K | — | — |
| 7 | Self-Improving + Proactive Agent | Skill | 80.0K | — | MIT-0 (Free to use, modify, and redistribute. No a |
| 8 | Openai Whisper | Skill | 63.1K | — | — |
| 9 | Evolver | Skill | 51.1K | — | — |
| 10 | Free Ride - Unlimited free AI | Skill | 50.8K | — | — |
| 11 | Elite Longterm Memory | Skill | 46.1K | — | MIT-0 (Free to use, modify, and redistribute. No a |
| 12 | ByteRover | Skill | 33.6K | — | — |
| 13 | Session-logs | Skill | 32.2K | — | — |
| 14 | Memory Setup | Skill | 31.2K | — | — |
| 15 | Evolver | Skill | 30.3K | — | — |
FAQ
How is this list different from the category page on aiskill.market?
The category page is a directory: every skill in the category, sortable and filterable. This list is editorial — opinionated, time-stamped (2026-05-17), and ranked. Use the directory when you know what you want; use this when you don't.
Why does the #1 pick have fewer stars than #5?
Stars are one signal among several. The composite score above also includes install counts (which reflect actual usage on aiskill.market) and the optional quality score. A skill with a smaller star count can rank higher if its installs or quality score are strong enough to offset.
Are these all free?
Yes. Every skill in this top 15 is open source. We verified the upstream repos; license columns showing "—" mean no SPDX identifier was declared, not that there's a restriction.
How do I install one?
Each linked skill page has install instructions. The fastest path is the one-line install via the aiskill.market CLI or by adding the source repo as a Claude Code plugin marketplace.
How often does this list update?
Quarterly. We re-run the generator script against the catalog every three months and refresh the commentary. The AI/ML category moves fastest, so expect more turnover here than in the other Best-Of articles.
Memory skills versus self-improvement skills — which do I need first?
Memory first. Self-improvement skills assume the agent can remember what it learned across sessions; without memory infrastructure, "improvement" lasts as long as one conversation. Install a memory skill (#11 or #14), let it bed in for a week, then layer in a self-improvement skill (#4 or #6).
Do these skills work with Cursor / Cline / other agents, not just Claude Code?
Some. Elite Longterm Memory (#11) explicitly supports Cursor, ChatGPT, and Copilot. ByteRover (#12) is tool-agnostic via its CLI. The obra/superpowers skills (#1, #2, #3) are Claude Code first but generally portable. Check each skill's README for current compatibility.
Related Categories
- Best AI Skills for Development & Code Tools in 2026
- Best AI Skills for DevOps & Deployment in 2026
- The Harness, Not the Model — for why memory and self-improvement skills outrank model-call wrappers
Browse The Full Catalog
Find every skill in this category — including the ones that didn't make the top 15 — at the AI & Machine Learning page.
Part of the Best-Of 2026 series. Updated 2026-05-17. Skills sampled from a catalog of ~262 active entries with a combined 1.4M popularity signal across the ranked entries.