cyberneticlibrary

Inside the data.

We indexed every AI-agent artifact we could find, comprehended each one, and measured which actually work. Here's what 7,950 skills, tools, and workflows reveal — including the uncomfortable parts.

01

The index

175,632 skills surveyed19,214 candidates pulled7,950 deep-indexed
0
deep-indexed
0
capabilities
0
ready-to-run recipes
0
causally measured
0
capabilities measured
0
found to HARM output
02

What's in it

Six kinds of artifact, twelve business functions. Engineering dominates the open-source supply — the business functions you'd actually hire for are the underserved minority.

By kind

skill
5,396
workflow
1,180
command
508
subagent
476
mcp server
235
plugin
155

By business function

Engineering
4,285
Operations
1,482
Product
1,173
Marketing
975
Data
550
Productivity
361
General
334
Sales
321
Finance
255
Legal
195
HR
64
Support
31
03

Fresh and sourced widely

A current index — most of the corpus was updated this spring — drawn from hundreds of repos, not a handful.

Last updated, by month

26-0226-0426-06

Top contributing repos

Sheshiyer
574
alirezarezvani
447
Tibsfox
388
majiayu000
336
davepoon
326
richfrem
136
K-Dense-AI
122
openclaw
109
Orchestra-Research
93
Jeffallan
66
elementalsouls
64
deanpeters
53
nanocoai
52
Pantani
52
04

Stars don't predict quality

The uncomfortable truth at the heart of this project. We measured each skill's causal lift — does it improve the output versus no skill at all — and plotted it against GitHub stars. There is no relationship. Popular skills routinely sit below zero.

no effect+40+20-20-401★100★10k★377k★GitHub stars (log) →

Each dot is a measured artifact · green helps, red harms · x = stars (log) up to 183k

05

One in three measured skills make output worse

Of 199 artifacts we measured, 69 produced measurably worse results than using no skill at all. Stars don't warn you — measurement does.

Distribution of measured lift

5
-50
8
-40
7
-30
15
-20
37
-10
63
0
43
+10
17
+20
4
+30
measured causal lift (pp) →

Most popular skills that harm output

-52ppAudit LinkedIn ad campaign health4,890
-52ppLaunch Apple Search Ads campaigns1,242
-52ppExecute LinkedIn ABM campaigns20
-45ppTrack brand mentions and competitor sentiment230
-44ppScore TikTok Ads performance20
-40ppGenerate design-system admin dashboards41,165
-40ppAnalyze Google Ads account performance4,890
-40ppAudit Microsoft Ads and Google imports4,890
-36ppAnalyze YouTube ads across all formats4,890
-36ppSelect optimal app store category for ranking1,242
-33ppOptimize cold outbound to 90/100 quality2,362
-33ppGenerate personalized cold email sequences135
06

Where skills actually matter

A skill's value depends on how good the base model already is. Where the model is weak, the right skill lifts results a lot. Where it's already strong, skills add nothing — or hurt. The zone where skills matter shrinks as models improve.

model already strong → skip the skillno-skill baseline (how good the model is alone) →best skill lift (pp)

Capabilities where the model already wins — skip the skill

ab-experimentation · 100%brand-voice-management · 100%sentiment-analysis · 100%ad-copy-generation · 96%content-transformation · 96%marketing-strategy · 96%launch-strategy-generation · 92%cold-email-strategy · 90%campaign-optimization · 88%
07

The map of the AI-skill universe

Every comprehended artifact, projected by what it actually does. Neighbors do similar work. Colors are business functions; the glowing rings are the ones we measured — green where they help, red where they harm. Drag to explore.

AI artifacts · placed by what they do · measured-helps measured-harms
EngineeringOperationsProductMarketingDataProductivityGeneralSalesFinanceLegalHRSupportdrag · scroll or ± to zoom · hover
08

What this work costs today

The freelancer and agency prices for the jobs these recipes do — the budget a measured AI recipe competes with.

GTM
$45$150k
Legal
$14$30k
Finance
$5$20k
Product
$8$60k
Founder ops
$5$20k
Operations
$5$50k
Hiring
$7$9.0k

what people pay freelancers/agencies/tools for these jobs today · median dot

09

How much runs on its own

Recipes chain library skills, but some steps need outside tools — your CRM, a sender, a calendar. Here's which tools recur, and how self-contained each business area is.

Most-used connectors

google
23×
gmail
14×
linkedin
11×
notion
11×
apollo
9×
hubspot
8×
meta
8×
instantly
8×
clay
6×
smartlead
6×
slack
6×
quickbooks
5×

Autonomy by area

GTM
76% library
Finance
81% library
Hiring
77% library
Founder ops
81% library
Operations
79% library
Legal
86% library
Product
80% library

runs from the library needs an external tool

10

The capabilities that power everything

A few stable capabilities show up across many recipes — the reusable backbone of the library. Build these well and the whole catalog gets better.

sales-strategy
8 recipes
ad-copy-generation
6 recipes
cold-email-strategy
5 recipes
personalized-outreach
5 recipes
brand-voice-management
5 recipes
data-analysis
4 recipes
html-generation
4 recipes
founder-coaching
4 recipes
seo-content-optimization
4 recipes
marketing-strategy
4 recipes
customer-profiling
3 recipes
sales-briefing
3 recipes
social-media-analytics
3 recipes
adversarial-research
3 recipes
seo-competitor-analysis
3 recipes
cold-email-generation
3 recipes
ai-content-optimization
3 recipes
contract-risk-analysis
3 recipes

The point of all this measurement

So when you pick a recipe, the steps are filled by what actually works — not what's popular. Pick a job and run it.

What do you need to run? →