what 700+ AI skills teach us about automation

Table of content

by Ray Svitla

travis maintains awesome-claude-skills — a curated library of Claude skills that people have built and shared.

as of february 2026: 700+ skills.

content writing, code review, language learning, financial analysis, meal planning, resume optimization, legal document drafting, research synthesis, email management, SEO optimization, and 690 other things.

I spent a week reading through all of them.

not using them. not testing them. just reading the instructions, understanding the patterns, seeing what people are actually automating.

here’s what I learned.

the distribution is weird

50% of skills: writing variants

blog posts, tweets, emails, documentation, marketing copy, technical specs, product descriptions, press releases, social media content.

all the same pattern:

30% of skills: analysis/synthesis

reading papers, summarizing threads, extracting insights, comparing options, doing competitive analysis, financial modeling, data interpretation.

pattern:

15% of skills: workflow automation

code review, test generation, documentation updates, git commit messages, project management updates, meeting notes.

pattern:

5% of skills: weird/niche

recipe generation, DnD campaign planning, language translation with cultural context, meditation guide customization, music theory tutoring.

these are the interesting ones. highly specific. often deeply personal.

what people actually automate

looking at usage stats (github stars, forks, mentions in discord):

top 10 most-used skills:

  1. technical documentation writer
  2. code review assistant
  3. twitter/X thread generator
  4. research paper summarizer
  5. email drafter (professional voice)
  6. SEO content optimizer
  7. blog post writer (markdown output)
  8. meeting notes synthesizer
  9. git commit message generator
  10. product requirements doc writer

notice what’s missing: creative work.

people aren’t automating “write a novel” or “design a logo” or “compose music”.

they’re automating the scaffolding around creative work. the docs, the summaries, the administrative layer.

the three types of skills

type 1: template fillers

“here’s a format, here’s some data, fill in the blanks”

example: resume optimizer. input: your experience. output: resume formatted for ATS systems.

these are the most common. they’re also the least interesting. pre-AI, these were mail merge templates.

type 2: synthesis engines

“here’s a bunch of information, extract the signal, format it”

example: research paper summarizer. input: PDF. output: key findings, methodology, implications.

these are more valuable. pre-AI, this required actually reading and understanding.

type 3: judgment assistants

“here’s a situation, here’s context, help me decide”

example: code review assistant that doesn’t just flag style issues but asks “is this the right architecture for scale?”

these are rare. they require domain expertise baked into the instructions.

what makes a skill good

I looked at the most-forked, most-adapted skills. common traits:

1. specific output format

vague: “write a blog post”
good: “write 1200-1500 word blog post, markdown format, lowercase headers, three sections, end with questions”

specificity prevents drift.

2. clear voice/style

vague: “professional tone”
good: “write like Paul Graham essays: short sentences, concrete examples, conversational but precise, avoid jargon”

voice is the difference between generic and useful.

3. examples included

skills without examples: AI guesses.
skills with examples: AI matches the pattern.

even one good example 10x’s the output quality.

4. error handling

bad: “if data is unclear, make your best guess”
good: “if data is unclear, output ‘[NEEDS INPUT: specific thing]’ and stop”

you want failures to be obvious, not hidden.

5. iteration hooks

bad: skill generates output, done.
good: skill generates output, asks “what should I adjust?”, refines.

the best skills are conversational, not one-shot.

the context pattern

almost every good skill has this structure:

IDENTITY: who/what the AI should act as
CONTEXT: what information it has access to
CONSTRAINTS: what it must/must not do
OUTPUT: format, length, style
EXAMPLES: 1-3 samples of good output
ITERATION: how to refine

bad skills skip steps. usually examples and iteration.

what people get wrong

mistake 1: too much personality

“you are an enthusiastic AI assistant who loves helping with X!”

nobody wants enthusiastic AI. they want competent AI.

best skills treat AI like a tool, not a character.

mistake 2: not enough constraints

“write a summary”

okay. how long? what format? what tone? what details to include/exclude?

constraints aren’t limiting. they’re clarifying.

mistake 3: one-size-fits-all

“universal content writer”

a skill that tries to do everything does nothing well.

best skills are narrow. “twitter thread for technical audience” beats “social media post”.

mistake 4: no quality bar

skill generates output. user accepts it. ships it. it’s mediocre.

best skills include: “quality checklist” at the end. “does this output meet criteria X, Y, Z?”

mistake 5: ignoring context engineering

skill has great instructions. but assumes AI has context it doesn’t.

“write in my usual style” — what style? include a style guide. attach examples.

context is half the battle.

the vertical skill hypothesis

general skills are commodities. vertical skills are valuable.

commodity: “write a blog post”
100 versions of this exist. they’re all fine. none are great.

valuable: “write a SaaS landing page in the style of Basecamp’s marketing (concrete benefits, no jargon, skeptical customer POV)”

that’s specific. that requires taste. that’s worth sharing.

the next wave: highly vertical skill libraries.

“the lawyer’s AI toolkit” — 50 skills for contract review, legal research, document drafting, client communication.

“the designer’s AI toolkit” — 50 skills for design critique, style guide creation, responsive layout suggestions, accessibility audits.

someone will build these. probably charge $500-2000 per vertical. people will pay.

the automation ceiling

skills are great for tasks that are:

skills struggle with tasks that are:

example: “write my weekly newsletter” — great skill candidate.

example: “write the apology email to the client we screwed over” — bad skill candidate. too much context, too high-stakes.

knowing which is which is the skill.

the CHOP connection

steve yegge’s Chat-Oriented Programming is basically “skills for code”.

instead of writing code, you describe what you want. AI writes it. you review, iterate.

skills are the same idea extended to all knowledge work.

the 700+ skills in awesome-claude-skills are proof that almost everything white-collar workers do can be described, templated, and automated.

not should be. not must be. but can be.

what this means for personal AI

if skills replace apps, and 700+ skills already exist, you don’t need most SaaS anymore.

need a CRM? there’s a skill.
need email management? there’s a skill.
need content calendar? there’s a skill.

you just need:

  1. AI access (Claude API, OpenAI, whatever)
  2. skills library
  3. personal infrastructure to run them

everything else is optional.

the skill marketplace that should exist

right now awesome-claude-skills is a github repo. markdown files. copy-paste to use.

that’s fine for early adopters. not fine for mainstream.

what’s needed:

whoever builds that — the npm for AI skills — wins.

what the data reveals about work

700+ skills is a dataset about what people actually do.

and what it shows: most “knowledge work” is pattern matching.

those are all templateable. which means automatable.

the things that aren’t templateable:

those show up as skills way less. because they’re hard to template.

the ADHD insight

I have ADHD. skills are a cognitive prosthetic.

pre-AI: “I should write that blog post” → context switch cost → never happens.

post-AI with skills: “write blog post about X” → skill runs → draft appears → I edit.

the activation energy dropped from “summon executive function” to “type one sentence”.

for ADHD brains , that’s transformative.

and looking at the skill library, a lot of them are ADHD accommodations:

skills as accessibility tools. that’s an underrated use case.

what skills don’t replace

after reading 700+ skills, I notice what’s not there:

taste
skills can execute. they can’t decide if something is good.

relationships
skills can draft emails. they can’t read between the lines in a conversation.

strategy
skills can analyze options. they can’t make the bet.

integration
skills can write code for component A and B. they can’t figure out how A and B should work together.

those are still human domains. for now.

the skill-building skill

meta observation: the most valuable skill might be “skill builder”.

being good at:

that’s a learnable skill. and it’s becoming a valuable one.

building in public with skill creation is content. sharing your skill library is marketing. teaching people to build skills is a business.

the timeline

we’re in the middle of that curve.

what I built after this analysis

I went through the 700 skills and forked ~40 that I thought I’d use.

then I customized them:

now I have a personal skill library. 40 skills that know my voice, my workflows, my preferences.

that’s way more valuable than 700 generic skills.

and it took ~3 hours to build.

the forward question

if 700 skills exist now, how many in a year? 5000? 10,000?

at some point the library is too big to navigate. you need:

that’s the next layer.

700+ skills is a snapshot of what people are automating in early 2026.

it’s also a preview of what work looks like in 2030.

most of the templateable stuff: automated.

what’s left: judgment, taste, strategy, creativity.

and figuring out which is which.

are you using skills? building them? what’s the one task you wish was a skill but isn’t yet?

Ray Svitla
stay evolving 🐌