5 Things To Steal From Everything Claude Code

Everything Claude Code, usually shortened to ECC, is a Claude Code configuration repository published by Affaan Mustafa after he won an Anthropic hackathon and shipped a product built entirely with agents. It now carries roughly 250,000 stars on GitHub.

The counts quoted about it disagree, and the disagreement is instructive. The video that carried ECC to a wide audience, published four days before this measurement, counted 64 agents and 262 skills. The repository's own README, on the day we cloned it, said 68 agents and 286 skills.

Source	Agents	Skills
The widely shared video	64	262
The repo's README, today	68	286

Those are not rival counts. They are older ones. The repository moves faster than the videos made about it, which is a reasonable measure of how much any single number is worth. The question underneath is the one asked plainly on Reddit and never really answered: not whether ECC is impressive, but whether it is worth installing.

ECC describes itself as an agent harness performance optimization system covering skills, instincts, memory, security, and research-first development. Its README compresses the loop it installs into seven words: plan, test, implement, review, verify, remember, improve. Install it once and that becomes how your agent works.

Counted in a fresh clone, the repository contains:

Component	Count
Agents	68
Skills	286
Slash commands	94
Rule files	122
Hooks (all in one file)	23

The README that describes all of it runs 2,200 lines. The documentation folder alone is 15 MB, translated into 22 languages. Cloning the whole repository pulls down 88 MB. It is not Claude Code specific either: ECC ships configuration for 19 different agents and editors, from Codex and Cursor to Zed and Kimi.

This is a real body of work, and the licence is MIT, which means every piece of it is yours to copy. The problem starts only when you install all of it.

A skill's body is read only when the skill fires. Its name and its description, though, sit in your context from the first token of every session. The same is true of every subagent. That metadata is the price of admission, and it is the part the repository tours skip.

Measured on the fresh clone:

What loads every session	Size	Approximate tokens
286 skill names + descriptions	85,000 characters	21,000
68 subagent descriptions	20,000 characters	5,000
The repo's own instruction file	—	1,000
Total standing context	—	~27,000

Against a 200,000 token context window, a full install occupies about 14 percent before you type a word. Fourteen percent sounds survivable until you remember what shares that window: your files, your diffs, your tool results, and the conversation itself. Every skill you will never call competes with the code you are actually working on.

The repository knows. Its own README carries the line "optimize the context window, persist everything else," and then ships 286 skills. On the Reddit thread asking exactly this question, the top reply says the token usage of the descriptions is minimal. Twenty-one thousand tokens is not minimal.

Two honest caveats. Tokens here are estimated at four characters each, so the real figure moves with the tokenizer. And none of this says the skills are bad. It says they are not free, and that most of them are not yours.

The first two pieces worth taking need no install of the repository at all.

AgentShield is a security auditor for agent configurations. It ships inside ECC and also lives on npm as its own package, where it took 5,500 downloads in a week. One command points it at a directory; it audits your agent files, your hooks, your servers, your permissions and your secrets, and it takes flags for output format, severity floor, and safe automatic fixes. Its README explains why it exists with numbers from January: 12 percent of one agent skill marketplace turned out to be malicious, which is 341 skills out of 2,800, and one breach compromised 1.5 million API tokens. Those are the package's claims, not our measurements.

Its limit is a date. AgentShield was last published in March while the repository itself was updated the day before we measured, and its README still describes the ecosystem at 42,000 stars against 250,000 today. The scanner is six months behind the thing it ships with.

The settings block costs nothing to try. The README publishes four changes:

Setting	Change
Default model	Set to Sonnet
Thinking budget	32,000 down to 10,000
Compaction threshold	95 percent down to 50 percent
Subagents	Pushed onto the cheapest model

The README claims about 60 percent off the model cost and 70 percent off the hidden thinking cost. Those are its figures, not ours, and they describe a trade: a smaller default model and cheap subagents will cost you on the hard reviews.

Instincts are the one idea in this repository we have not seen elsewhere. An instinct is a small learned behaviour with a confidence score. Not a rule you wrote, but a pattern the system noticed you repeating and wrote down for you.

The first version only watched at the end of a session. The current one watches before and after every tool call, which is why it catches anything at all. Each instinct carries a weight between 0.3 and 0.9, and the analysis runs in a background agent on the cheap model, so it does not eat your main context to do its job.

Version 2.1 made instincts project scoped, so your React patterns stay in the React project. The repository's own word for what that fixed is contamination. When the same instinct appears in two different projects, it is promoted to global: the system decides what is universal by watching, instead of asking you.

This is not a stub. The tool behind it runs 2,200 lines with six commands, and the path it describes goes from instincts to clusters to a real skill. Instincts are also portable: export your library, hand it to someone on your team, and they import it, which is the first sane answer we have seen to onboarding a person onto an agent setup.

The string attached is the hook runtime. All of this rides on it, and the hook runtime is the exact part of the repository you are made to opt into on purpose. Refuse the hooks and you do not get instincts. A confidence score is also a heuristic, not a measurement.

Keep number four is three skills that account for context instead of spending it: a context budget, a compaction adviser, and a token budget adviser. The repository that costs you 14 percent also ships the tools to see it. The compaction rule alone is worth taking: compact after research and before you implement, after a milestone, after debugging, after an approach fails, and never in the middle of an implementation, where you lose the file paths and the half-finished state.

Keep number five is the installer, which is what makes the other four reachable. It ships seven profiles, and the smallest one says in its own description that it leaves the hook runtime out entirely. You can also skip profiles and name the skills you want one by one, and an adviser takes a plain sentence about your work and returns the components that match, with a preview before anything is written to disk. The installer also refuses to guess about hooks: any install that would put the hook runtime on your machine stops, prints what those hooks can do, and waits for a yes or no.

Install that way and the bill stops being 14 percent and becomes a fraction of it. We are not putting a figure on that fraction, because we did not measure every combination. The floor is honest, though: even the smallest profile still carries the agents and the rules, so naming the skills by hand is the only way to pay for exactly what you use.

The 68 agents as a set. They cost about 5,000 tokens of standing context, and most of them are language reviewers and build fixers for C++, C#, Django and Kotlin, for stacks you do not write in. Take the two that match your work. The set is not a package deal.

The 94 slash commands. The README already treats them as temporary, calling them convenient entry points while the repository moves to a skills-first surface, and there is a folder of retired ones. Commands are also the easiest thing in this ecosystem to write yourself.

The hooks file. Every hook in this setup lives in a single file: 291 lines, 40 KB, 23 hooks, of which 8 run before every tool call and 7 run every time the agent stops. The code is not in a script you can open. Every hook is squeezed onto one line, 33,000 characters of it, with the longest single line running to 2,000 characters.

Be fair about it: the descriptions beside them are readable and some are good. One blocks the agent from editing your linter config and says why, steering the agent to fix the code instead of weakening the rules. That is a hook worth having. The problem is not the intent. The code is public and it is MIT. The problem is that the intent is the only part you can read, and the repository says it first: hooks run shell commands, and you should treat them as executable configuration.

The ballast. Configuration for 19 harnesses and 15 MB of documentation in 22 languages, for a tool you run in one language, in one place.

Raid it, do not install it. The five pieces that survive the audit are the scanner, the settings block, the instinct system, the three skills that count your context, and the installer that gets you the first four without the rest.

The number that decides it is 27,000 tokens of standing context for a full install against a window of 200,000. The licence makes the raid legitimate: it is MIT, and copying five directories out of this repository is the intended use, not a workaround.

On Reddit people keep asking whether to pick Superpowers or ECC. They are not the same kind of thing. One is a discipline for how the agent works; the other is a parts bin, and the answer for a parts bin is to take parts.

Two limits on all of this. Everything here was measured on version 2.2.1; the repository was updated the day before, and it has 110 open pull requests waiting, so it moves. And none of this measures whether the skills make the answers better. It measures what they cost you to have. Those are different questions, and only one of them fits in a video.