The Skill Library: 20 Composable Capabilities Every AI-Native Team Should Ship First

If skills are the reusable unit and agents are the orchestrator, the next operator question is brutally practical: which skills do I write first? Anthropic's Agent Skills framing—filesystem-addressable capabilities loaded on demand—gives you the container.^[1] It does not tell you what to put inside it. Below is the library I would build first, ordered by category and ruthless about repeatability. Twenty capabilities, each described with the contract it should hold, the work it earns its keep on, and how it composes with the rest of the stack. Build these, version them, regression-test them, and you have an IC-grade operating layer that compounds the way The Individual Contributor Age argues serious operators now must.

What counts as a skill (and what does not)

A skill is narrow, named, and reusable. It takes a defined input, applies a single capability, and returns a structured output. "Write copy" is not a skill. "SCQA brief writer" is. "Analyze the data" is not a skill. "Deep research synthesizer with source validation" is. The right granularity is "a thing one person could trust enough to delegate by name." If you cannot regression-test the output, the skill is too broad.

Writing & content skills

Writing skills are where most operators feel the gap first—long-form output is constant, judgment about structure is rare, and the difference between a good draft and noise is enormous. Five worth writing in the first month.

• 1. SCQA Writing Framework. Converts an unstructured idea into Situation, Complication, Question, Answer—the same backbone used in this essay and every consulting deliverable worth reading. Use as the default frame for any thread, briefing, or article. Output: a structured SCQA block plus short paragraphs and bullet highlights.
• 2. Content Repurposing Engine. Takes a long-form artifact and produces social threads, short-form video scripts, executive summaries, and newsletter formats—preserving the core argument and adjusting tone per channel. Pairs naturally with the SCQA frame above.
• 3. Tone & Style Enforcer. Keeps every output aligned with a defined brand or personal voice—an editorial gate the agent layer can run before publication. This is where machine-readable brand intersects with operator output; see DESIGN.md, Stitch, and the bet on machine-readable brand for the deeper bet, and our Brand Lockup case study for an enterprise-scale version of the same idea.
• 4. Long-Form to Summary Compressor. Reduces dense source material—papers, transcripts, meeting notes—into digestible summaries without dropping the load-bearing claims. Critical input to the knowledge layer below.
• 5. Structured Copywriting Skill. Produces hook, body, and CTA in a defined structure for marketing, social, or product copy. Composes with the tone enforcer for governance. The discipline underneath this category is the bet in More Variants Are Cheap. Creative Learning Is the Bottleneck.

Visual & diagram skills

Most operators undervalue visual skills because the model lets them muddle through with a bullet list. That is a mistake. A diagram beats a thousand-token paragraph for decision-making in hot windows, and these skills are unusually high-leverage relative to the engineering cost of writing them.

• 6. Excalidraw Diagram Generator. Converts a workflow or concept description into the node/connector structure Excalidraw or similar tools render. Good for explaining systems to the team fast.
• 7. Infographic Builder. Turns process or step content into structured visual summaries—steps, headings, optional iconography. Earns its keep when briefing executives who do not read past page one.
• 8. Flowchart & Decision Builder. Generates decision trees from a process description with explicit conditional branches. Most operating playbooks become testable for the first time when they hit this skill.
• 9. UI/UX Layout Advisor. Given an interface goal, suggests structured layout with spacing, hierarchy, and accessibility hints. Compounds when paired with a working design system; see Design Systems as Code.

Research & analysis skills

This category is where the asymmetric gains hide. Research at scale used to require a team; now it requires a stack. Five skills that cover the surface area.

• 10. Deep Research Synthesizer. Pulls signal from large source sets, removes low-information content, and produces structured insights with supporting detail. This is the operator counterpart to the Karpathy autoresearch pattern—the skill at the bottom of the overnight loop.
• 11. Onchain Transaction Analyzer. Traces wallets, contracts, and token movements and returns plain-language explanations. Disproportionately useful for compliance, intelligence, and investigative work—similar in spirit to our ThreatBase agentic OSINT approach.
• 12. Source Validation Skill. Scores references for credibility, freshness, relevance, and bias. Runs as a gate before any research output ships externally. The cheapest possible insurance against confidently wrong assistants.
• 13. Competitive Intelligence Skill. Compares products, protocols, or tools across features, strengths, weaknesses, and opportunities. Output is a structured comparison plus actionable takeaways, not a vibes summary.
• 14. Knowledge Structuring Skill. Takes messy capture—notes, transcripts, screenshots, web clippings—and returns a clean, categorized knowledge structure. This is the engine room behind a personal LLM knowledge base; the case for where PageStash fits in the capture-to-structure path is in the linked piece.

Video & multimedia skills

Video is the format where most operators still treat AI like a toy. The honest read is that a small, composed video stack now produces studio-grade short-form output in an afternoon. Four skills that close the gap.

• 15. Video Script Generator. Produces a structured script with hooks, sections, pacing notes, and explicit calls-to-action. Best paired with the hook generator below for the cold open.
• 16. Video Editing Planner. Returns a scene-by-scene plan with cuts, transitions, and pacing guidance—saves hours of editor staring at a timeline trying to find the cold start.
• 17. Hook Generator. Produces attention-grabbing openings, optimized for retention in short-form formats. Composes with the SCQA writing skill—the hook earns the audience the right to spend three minutes on the situation.
• 18. Caption & Subtitle Formatter. Formats captions for readability and accessibility, enforces line-length and timing rules, and reduces the post-production tax substantially.

Engineering & automation skills

The final category is where the team-level leverage compounds the most. Two skills that every operator should have on their personal stack, even if they do not consider themselves an engineer.

• 19. Code Review Skill. Reviews changes for bugs, inefficiencies, and adherence to a house style. The version worth shipping is opinionated and project-specific; a generic linter will not give you the same compounding return. Particularly valuable when the operator owns the loop end-to-end, as argued in the IC piece.
• 20. Workflow Automation Skill. Takes a goal description and returns a stepwise workflow—mapped to specific tools, with explicit handoffs and observability hooks. This is the skill that turns "we should automate this" into "here is the n8n graph by Friday." The principled execution layer underneath sits in n8n for cautious enterprise automation and LangGraph on the Ground.

The meta-skill: how to keep writing more

Build the library and you will start noticing the gaps. The right twenty-first skill is therefore a meta: a skill creator that takes a goal description and returns a properly structured skill file—name, description, instructions, constraints—ready to drop into the library. Use it sparingly. A library of fifty mediocre skills you do not trust is worse than a library of twenty you do.

How to put this into practice this month

The operator's first move is to pick three skills that are painful in your week—the kind of work you do every Tuesday with mild resentment. Write those three first. Use them for two weeks. Tighten them until you trust the output without re-reading it. Then add another three. Inside a quarter you have an opinionated, tested, named library that genuinely changes the shape of your week.

The team's first move is to publish the library internally the same way you publish a design system. Names, contracts, tests, owners, change log. The skill is a product. Treat it like one, and the rest of the agentic stack stops feeling like an experiment and starts feeling like infrastructure.

Tools are commodity. Models are commodity. The unit that is yours is the skill—named, versioned, trusted, composed. Twenty of them, written deliberately, beat a thousand agents shipped on a slide. Start with three. Compound from there.

Related: Build Skills, Not Agent Armies, Personal LLM Knowledge Bases, and The Individual Contributor Age.