Ravli

Testing AI tools so you don't have to. Prompts, setups and agents that actually work

[Ravli @Ravlizk]

Anthropic ships almost 40 in-house plugins for Claude Code. These 10 are the ones to install first

THE SETUP:

→ code-review puts several agents on every PR and scores each finding by confidence, so the real issues reach you and the nitpicks don't

→ claude-security hunts vulnerabilities in your own code and makes agents argue with each finding before it reaches you

→ security-guidance flags injection, XSS, SSRF and hardcoded secrets right when Claude writes them

→ feature-dev splits a feature into explore, architect and review steps with a separate agent for each

→ code-modernization reads legacy code, writes down the business rules hidden in it and plans the upgrade

→ ralph-loop keeps Claude on one prompt in a while-true loop until the job is finished

→ claude-code-setup scans your repo and tells you which hooks, skills and MCP servers it needs

→ hookify turns a repeated mistake into a hook that stops it next time

→ skill-creator writes skills and runs evals on them

→ playground builds a one-file HTML tool with sliders and live preview

stop writing one huge system prompt and hoping Claude remembers it

install: /plugin install name@claude-plugins-official
full list in the reply ↓

[Gipp 🦅 @gippp69]

https://t.co/gTEKJ3i4XU

[Ravli @Ravlizk]

@Vaprolus I think i kinda need to rethink my money magment hahahaha

[Ravli @Ravlizk]

@Amos_FACT GM

[Ravli @Ravlizk]

@lateascent Ithink we all MUST do this, it's not even a question now

[Ravli @Ravlizk]

@Qyravex It should, trust me, it's as easy as take a candy from a baby hahahaha

[Ravli @Ravlizk]

This is f*cking insane. Claude Code is quietly turning into a jailbreak tool for the stuff in your living room

this week alone:

→ a TV that showed an ad before Netflix: 25 Xiaomi packages removed, new launcher. 5.1M views
→ a Kindle over USB: Amazon bloat gone, Tailscale added
→ a 2017 TV in 5 minutes: ads blocked, faster UI, apps straight from GitHub
→ a TV and a PS5 in one night: 37 tracking domains blocked, 14 preinstalled apps hidden

same move every time: connect the device, describe what's annoying, let Claude read it and fix it

and if you want to go past TVs, that's where REA (Reverse Engineer Anything) comes in: an open-source repo that lets your agent take apart app behavior down to the binary

full course in the quoted post

which device in your house goes first?

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[Ravli @Ravlizk]

**Solid cost-optimization pattern, with the usual caveats.**

The split (cheap model for bulk reading/architecture mapping via OpenCode + Opus 5.5 for planning and implementation) is a practical way to avoid burning the expensive model’s context and tokens on raw repo ingestion. OpenCode being free/MIT and model-agnostic makes the “only pay for the model” claim accurate; DeepSeek V4 Flash / GLM-5.3 Flash (and their recent free or very low-cost tiers on some OpenCode setups) are reasonable for the documentation pass. The explicit “read the architecture notes first and propose a plan” prompt, plus the license reminder, are the parts that actually make it workable.[](<https://t.co/jQ4mX3T0R0>)[](<https://t.co/C0RksYAjXY>)[](<https://t.co/oLh0fs83pF>)

### What works well
- Architecture map (folder structure, data flow, key modules) written into a `docs/` folder is a high-leverage intermediate artifact. It compresses the repo for the expensive model and gives you something reviewable before code is generated.
- Forcing a plan-first step reduces the “write the whole thing immediately and hope” failure mode.
- MIT/permissive licenses make the “rebuild my own version with these changes” path cleanest. Checking the license file first is the non-negotiable part.

### Practical limits
- Flash-tier models are good at surface-level mapping but can miss subtle invariants, error paths, or implicit contracts. Treat the generated docs as a draft and spot-check critical modules yourself (or have a second pass).
- “Build your own version” is strongest as a clean-room exercise guided by the notes + your requirements, not as near-verbatim recreation. Even with MIT, subtle design choices, performance characteristics, and edge-case handling often don’t transfer cleanly.
- Opus 5.5 is strong for agentic coding, but you still need to verify the output (tests, types, behavior) rather than treating the first generation as final.
- Context and cost savings are real for large repos, but the quality of the final result is still bottlenecked by how good the intermediate architecture notes are and how carefully you review the plan.

Overall it’s a reasonable workflow if you’re disciplined about the review steps and licenses. The “stop asking the expensive model to read the whole repo cold” advice is the highest-signal part.

[Ravli @Ravlizk]

@nar96265 hi guys

[Ravli @Ravlizk]

@0xSeinz At 95% per step, \(0.95^9 \approx 63\%\) makes long unsupervised chains fragile. Tao’s point is that independent verification (human gates or hard checks) breaks the pure \(p^n\) product into shorter, restartable segments.

How many handoffs do you typically run before a check?

[Ravli @Ravlizk]

@Qyravex Grok AI is for sure Top-3/Top-2 AI nowadays

[Ravli @Ravlizk]

full talk: How Anthropic uses Claude Code, Daisy Hollman at NDC Copenhagen
https://t.co/ROjbcsnb6Y

the worktree setup starts around 50:25

[Ravli @Ravlizk]

This is f*cking insane. Anthropic engineer Daisy Hollman (nearly 10 years on the C++ standards committee):

"I didn't feel like I could operate efficiently with just 16 permanent work trees. So now this goes all the way to Z."

26 git worktrees, A to Z. each one has a long-running Claude agent with its own name, its own memory and one job: keep that worktree moving

in this 1-hour NDC Copenhagen talk she shows how the Claude Code team actually works

worktree → persistent agent → /rename + /color so you know which window is which → you review what each one reports

her math: manage 20 Claudes instead of 5 and you get close to 4x the work done

the full talk is free on YouTube and has 36K views

link in the reply ↓

[Ravli @Ravlizk]

full collection from the Anthropic team: https://t.co/d03xjrzXW9

install any of them with /plugin install name@claude-plugins-official

[Ravli @Ravlizk]

Anthropic ships almost 40 in-house plugins for Claude Code. These 10 are the ones to install first

THE SETUP:

→ code-review puts several agents on every PR and scores each finding by confidence, so the real issues reach you and the nitpicks don't

→ claude-security hunts vulnerabilities in your own code and makes agents argue with each finding before it reaches you

→ security-guidance flags injection, XSS, SSRF and hardcoded secrets right when Claude writes them

→ feature-dev splits a feature into explore, architect and review steps with a separate agent for each

→ code-modernization reads legacy code, writes down the business rules hidden in it and plans the upgrade

→ ralph-loop keeps Claude on one prompt in a while-true loop until the job is finished

→ claude-code-setup scans your repo and tells you which hooks, skills and MCP servers it needs

→ hookify turns a repeated mistake into a hook that stops it next time

→ skill-creator writes skills and runs evals on them

→ playground builds a one-file HTML tool with sliders and live preview

stop writing one huge system prompt and hoping Claude remembers it

install: /plugin install name@claude-plugins-official
full list in the reply ↓

[Gipp 🦅 @gippp69]

https://t.co/gTEKJ3i4XU

[Ravli @Ravlizk]

@kelvoroz I think you kinda right😅😅

[Ravli @Ravlizk]

@vayrenrose shit just got real

[Ravli @Ravlizk]

@0xboan looking interesting sent me promt hahaha