Prompts Are Dead. Skills Are the New Moat.
The format that unified Anthropic, OpenAI, Cursor, Microsoft, Vercel, and Lovable in six months, and what it means for every AI company building right now.
Lovable Just Shipped Something Worth Paying Attention To
One thing worth flagging before the full breakdown, because it connects directly.
Lovable shipped Skills this week. Most people are treating it as a product update. It is not.
Lovable crossed $400M ARR in February 2026, adding $100M in a single month with 146 employees at $2.77M revenue per head. Anton Osika told Bloomberg the company is “pacing five months ahead of projections.” 40 million projects built on the platform. 200,000 new ones every single day.
Skills is the feature that takes Lovable from “the tool that vibe-codes your app” to “the tool that encodes how your entire company operates.”
Here is what shipped.
Skills are workspace-scoped playbooks written as SKILL.md files. A name, a description starting with “Use when...”, and a markdown body of instructions. You invoke them by typing / or just asking naturally. Lovable figures out which one applies.
Three ways to create one:
Type “save that as a skill” inside any chat
Import a GitHub repo with a SKILL.md at the root
Upload a ZIP
No extra cost. No revenue share. Standard credits cover execution.
Lovable shipped Skills to 8 million non-technical founders, operators, and marketers. People who have never written a line of code and never will. That is a distribution bet nobody else in this space has made at this scale.
If you have been watching Lovable from a distance, this is the week to try it.
Now, the bigger story. Because Lovable is one piece of a much larger shift across the entire AI stack in the past six months.
What a Skill Actually Is
Six months ago, only developers knew what a Skill was. Today, every major coding agent has adopted the same format. Anthropic, OpenAI, Cursor, Microsoft, Vercel, Windsurf, Cognition, and Lovable all converged on the same primitive in under a year.
A Skill is a folder with a SKILL.md file inside. It has a name, a description that starts with “Use when...”, a markdown body of instructions, and optional scripts and references.
The architecture solves a real problem. Loading everything into context at once degrades agent performance. Skills route around that by loading procedural knowledge only when the agent needs it.
At session startup, only the name and description load. Thirty to 100 tokens per Skill. When a request matches the description, the full body loads. Forty Skills installed costs roughly 1,500 tokens at startup. Stacking is free until invoked.
Anthropic published the spec as an open standard at agentskills.io in December 2025. The anthropics/skills repo crossed 117,000 GitHub stars.
The Convergence
▫️ October 2025 — Anthropic launches Agent Skills
▫️ December 2025 — published as open standard at agentskills.io
▫️ January 2026 — Cursor ships Skills; Vercel ships skills.sh marketplace with 34,000+ skills
▫️ February to April 2026 — OpenAI ships Skills, codenamed “Hazelnut,” SKILL.md compatible
▫️ April to May 2026 — Microsoft ships Agent Skills in Visual Studio 2026 and Copilot Cowork
▫️ Q1 2026 — Windsurf, Cognition, Factory, Goose, and Cline all adopt SKILL.md
▫️ May 2026 — Lovable ships Skills to 8 million non-technical builders
A Skill written for Claude Code runs unmodified in Cursor, Copilot, OpenAI Codex, Windsurf, and Lovable. MCP got cross-vendor conformance with two years of effort. Skills got there in six months because the format is markdown plus YAML.
Why Skills Beat Everything Else
Skills sit in a gap that every other primitive leaves open.
▫️ Prompts live in someone’s notes app and break when the task changes
▫️ RAG retrieves information but carries no procedure for what to do with it
▫️ Tools and MCP expose capability without telling the agent when or how to use it
▫️ Fine-tuning costs five figures, takes weeks, and locks you to one model
▫️ Long context dumps everything in at once and degrades performance as the window fills
Skills fill the gap between knowing what to do and knowing exactly how your team wants it done. MCP is the plumbing for tool access. Skills are the procedural memory for how to use those tools well.
One more thing. Skills are model-portable. A SKILL.md file works in Codex CLI, Gemini CLI, and any agent that can read a folder. No protocol implementation required. That is the entire reason they spread in six months.
What Gets Commoditized
Prompt Management Platforms
Anthropic acquired Humanloop in July 2025 and shut it down on September 8. One foundation lab absorbed an entire category by acquihiring its leader. PromptLayer and Vellum are pivoting to enterprise registries. The “prompt is the product” thesis is done.
Thin Wrappers
Generic AI writing tools, single-prompt apps, and basic copilots face commoditization from open-source Skills templates anyone can clone in hours. If your only moat is a clever prompt, you have a head start measured in weeks.
Pure Prompt-Orchestration Frameworks
LangChain explicitly repositioned as an “agent engineering platform.” Lightweight orchestration is being absorbed into IDEs and foundation labs directly.
Who Wins
Foundation labs — Anthropic defined two AI primitives in one year. Claude Code exceeded $500M run-rate by September 2025.
Distribution-layer companies — Cursor at $29.3B. Lovable at $6.6B. Replit at $9B. Vercel at $9.3B. They own the user surface. Skills make their products more powerful without changing who owns distribution.
Evals and observability — Braintrust closed $80M at $800M. When Skills are commodity markdown, the moat shifts to proving whether your Skills work in production. Evals are the new gross margin.
Vertical AI with proprietary data — Harvey ($8B), Abridge, Decagon ($1.5B), Glean ($7.2B at $100M ARR). The switching cost lives in the workflow integration and the data corpus. The prompts are irrelevant.
The Four-Question Skills Test for Any AI Startup
Run this on every pitch deck you see this year.
1. Do you have proprietary data feeding your Skills? Generic Skills are markdown files anyone can copy. Skills trained on your unique data are not.
2. Do you own the user surface? Cursor, Lovable, and Replit do. Most AI startups do not.
3. Do you have evals proving your Skills measurably improve outcomes? “We have great Skills” is not a moat. “Our Skills improve task success by 34% across 12,000 production hours” is.
4. Is your workflow integration deep enough to survive an MCP-compliant competitor? REST APIs anyone can wrap is arbitrage, not infrastructure.
Pass on four nos. Position on three yeses. Back hard on four yeses.
The premium section below is the full operational playbook: 30+ ready-to-use Skills, copy-paste SKILL.md templates for every role, the complete library of prompts that build them, and the exact guidelines for making Skills that actually work in production.
The Complete Skills Playbook
Part 1: The SKILL.md Master Template
Every Skill you build should follow this structure. Copy it, fill in the blanks, and install it in your Project:
Keep reading with a 7-day free trial
Subscribe to The VC Corner to keep reading this post and get 7 days of free access to the full post archives.