just launched $FERRA on https://t.co/u7ww1AQiX9
CA:
3zmjpra45LG1wQ3afVTY2eeEbiXx8sj5fnnDuC6Cpump
Ferra reads your GitHub repo and generates a CLAUDE.md in 30 seconds
scans your manifest files, detects your stack, then scores 59 curated skills against what you're actually building. picks the 6-8 that fit. ignores the rest.
what you get: CLAUDE.md + skill files + pre-commit hooks + setup guide. drop it in your repo root and Claude Code reads it automatically every session.
doing this manually takes 45-60 minutes. most devs skip it. this is why their Claude keeps making the same mistakes.
free. 5 forges/day. no login.
https://t.co/X8ZgRNQ2f1
ECC — Extended Claude Code — is the one that quietly made the deeper workflow patterns in Ferra possible
deep workflow automation, integration patterns, the kind of stuff where Claude Code stops being a coding assistant and starts being something closer to an autonomous system. the pre-commit hook structure, the SETUP_GUIDE format, the way skill files reference each other, a lot of that thinking came from spending time in this repo
it has nowhere near the stars it deserves. probably because people who find it are too busy using it to go back and star it
https://t.co/nW6d6cVGis
the AI/LLM skill category in $FERRA MCP Server Integration, Agent Design Patterns, RAG Architecture — that whole section exists because of the Superpowers repo
extended Claude capabilities, advanced agent configs, tooling patterns that actually work in real workflows. it's the kind of repo that fills in the gaps Anthropic's docs leave open
most people building Claude agents haven't found it yet. that's a shame because it covers stuff you'd otherwise have to figure out the hard way over weeks of trial and error.
https://t.co/lUeQFbH30L
the pre-commit sh $FERRA generates actually runs
tsc --noEmit catches type errors before they hit CI
eslint with --max-warnings 0 means zero tolerance
vitest runs the full suite
tested it on 6 different projects. all hooks worked on first install with no manual edits. two of them caught real bugs in the first week that would've made it to production.
the SETUP_GUIDE.md told us exactly where everything goes. never had to guess.
the 59-skill library didn't come out of nowhere
a huge chunk of it was curated from hesreallyhim's awesome-claude-code repo, which at this point is basically the best maintained index of real Claude Code workflows that exists anywhere
what made it actually useful for Ferra is that the skills in that repo reflect how people actually use Claude Code in production. not theory. not demos. real patterns that real developers ran into and figured out.
if you build with Claude Code and you haven't gone through this repo yet you're leaving a lot on the table
https://t.co/fz7XbcEAny
tested skill selection by forging the same project type 10 times with slight description variations
got the same core 6 skills every time. the scoring algorithm doesn't drift.
the 2 variable slots (out of 8 total) shifted based on specifics — mentioned Celery once and got async-await-patterns, didn't mention it and got something else. that's exactly how it should work.
zero random outputs across all 10 runs.
the endgame isn't a CLAUDE.md you set up once
it's a Claude that gets incrementally smarter about your codebase every single week. new module added → context updated. architectural decision made → documented. team convention drifted → corrected.
Ferra becomes the layer between your repo and your AI that makes sure the AI is never working from stale information
developers 2 years from now won't remember what it felt like to correct Claude on the same thing twice. that's the version we're building toward.
where $FERRA is going next
right now you forge once and drop the files manually. that's already 720x faster than doing it by hand.
but the next version doesn't ask you to do anything. Ferra watches your repo. stack changes, dependencies updated, new patterns introduced — CLAUDE.md updates automatically. Claude Code always has current context. no human in the loop.
the goal is a Claude that knows your project better than most of your teammates do.
solo devs are just the beginning
the next problem Ferra solves is teams. right now every developer on a team has different Claude context. one person set up CLAUDE.md properly. two others have a blank file. the rest don't know what CLAUDE.md is.
Ferra for teams means one source of truth. shared skill library per org. everyone's Claude Code starts every session knowing the same architecture decisions, the same conventions, the same constraints.
a 10-person team where all 10 have elite Claude context is a different kind of company.
your API has /getTasks, /getTask/<id>, /task/add, /remove_task
all POST. error paths returning 200. no pagination.
picked the REST API Design skill from $FERRA's library. 5 lines: plural nouns, correct HTTP methods, proper status codes, cursor-based pagination.
one prompt and Claude renamed every single route, fixed every status code, and added pagination with next_cursor
it knew what "correct" meant because the skill defined it
the thing about having a REST skill loaded is Claude stops improvising
without it, it'll name routes however it feels that day. return 200 for errors because technically the request succeeded. skip pagination because you didn't explicitly ask.
with it has a convention. it follows the convention. every time.
5 lines in your CLAUDE.md and your whole API suddenly has opinions
https://t.co/wasFzlLURl
compared Ferra-generated CLAUDE.md against ones our team wrote manually
the Ferra ones had 3x more specific project context. things like: actual file tree structure, key design patterns for the detected stack, state management decisions, performance targets with real numbers.
manually written ones were mostly vibes. "use TypeScript. follow best practices."
Claude performs noticeably better with the Ferra version. not subtle.
people ask how Ferra reads a GitHub repo without cloning it
the answer is @yamadashy built repomix and we built on top of it
when you paste a URL, repomix is what packs those manifest files into something Ferra can actually reason over. package.json, tsconfig, prisma schema, requirements.txt — all of it, structured, AI-ready, in seconds. the GitHub URL scan mode literally doesn't exist without this.
25,700+ stars on the repo. one of the cleanest tools in the AI dev ecosystem right now and it deserves more credit than it gets
go star it: https://t.co/rdy2H7qWSI
follow: @repomix_ai
the wild part is the skill is freely available at https://t.co/wasFzlLURl
you don't even need to run a forge. just copy the snippet directly into your CLAUDE.md yourself
but if you do run a forge, Ferra auto-selects it along with 5-7 others that actually match your stack
59 skills in the library. your codebase has violations right now that Claude doesn't know to look for.
https://t.co/X8ZgRNQ2f1
grabbed the SOLID Principles skill from $FERRA's library and dropped the snippet into Claude Code
5 lines. max 20 lines per function, no magic numbers, single responsibility per module.
asked it to audit 3 Python files before touching anything
came back with a full violation table. file, line number, principle broken, exact issue. every single one.
that's not Claude being smart. that's Claude having a framework to audit against instead of guessing.
timed 50 forges across different repo sizes
smallest repo (12 files): 2.8s
average repo (~200 files): 4.1s
largest repo we tested (1,400+ files): 6.7s
none of them broke 10 seconds. none.
for context, writing a decent CLAUDE.md manually takes 45 minutes minimum. and you still probably miss half the relevant skills.
if you've used Ferra and wondered why the CLAUDE.md feels different from other generators — it's because @karpathy is basically a silent co-author of every single output
his coding philosophy is the first section that loads. code reads like prose. explicit over implicit. document *why*, not *what*. single responsibility per function.
we didn't just borrow the ideas, we hardcoded them as the foundation. before skills, before stack detection, before anything else — Karpathy principles go in first.
the man hasn't touched Ferra once and he's still shaped every CLAUDE.md it's ever generated
https://t.co/lKoVSrgy5f
the Web Performance skill it pulled has actual targets baked in
LCP < 2.5s, FID < 100ms, CLS < 0.1
next/image for everything
next/bundle-analyzer in the pipeline
lazy load all below-fold components
Claude now has these as constraints, not suggestions. it'll push back if you're about to do something that violates them
6 more skills in the same ZIP. all Next.js/React. nothing from Python or Go snuck in.
https://t.co/X8ZgRNQ2f1
didn't type anything. just clicked "Next.js SaaS" from the template list
one click filled in the whole description. hit Forge It.
30 seconds later the CLAUDE.md had: frontend/backend/UI/state/validation/performance all broken down. key design patterns. component organization. SOLID principles applied to the actual stack.
that's not a template. that's an architecture guide.