Everything Claude Code: the 27,000 token audit, the five keeps and the raid checklist

TL;DR

• Everything Claude Code (ECC) is a repository to raid, not to install whole: a full install parks about 27,000 tokens of standing context in a 200,000 token window, about 14 percent, before your first message.
• The bill is metadata. The 286 skill names and descriptions weigh 85,000 characters, about 21,000 tokens, and the 68 subagent descriptions add about 5,000 more. Skill bodies load on demand; descriptions do not.
• Five pieces earn their place: AgentShield, the settings block, the instinct system, the three context-accounting skills, and the selective installer that fetches the first four without the rest.
• Leave the 68 agents as a set, the 94 slash commands the README itself calls temporary, the 23 hooks minified into one 40 KB file, and 15 MB of documentation in 22 languages.
• The licence is MIT. Copying five directories is the intended use. Everything below was counted on version 2.2.1 on 2026-09-06, with 110 open pull requests waiting, so recount before you quote.

What the measurements say

The repository is larger than the numbers people repeat about it. The counts circulating online said 214,000 stars, 64 subagents, 262 skills and 84 slash commands. On the day of the clone, the README and the directories said 250,466 stars, 68 agents, 286 skills, 94 slash commands, 122 rule files and 23 hooks s2. The repository also carried 37,679 forks, 51 open issues, 110 open pull requests and 2,631 commits s1. None of the quoted figures were wrong when they were published. They were stale within days, which is the first lesson: this repository moves faster than anything written about it.

Size follows. The README runs 2,200 lines, the documentation folder is 15 MB in 22 languages, and a full clone pulls 88 MB s2. ECC also ships configuration for 19 different agent harnesses, so a Claude Code user downloads 18 configurations they will never load s1.

The question that matters is what a full install costs every session, and the answer lives in the descriptions, not the bodies. Claude Code loads the name and description of every installed skill and every subagent into context so the model can decide when to use them. Summing the 286 skill names and descriptions gives 85,000 characters, about 21,000 tokens; the 68 subagent descriptions add about 20,000 characters, about 5,000 tokens; the repository's own instruction file adds about 1,000. The standing total is about 27,000 tokens, 14 percent of a 200,000 token window s3. That 14 percent is paid before your files, your diffs, your tool results and the conversation itself, and every compaction pays it again.

The repository knows this. Its README carries the line "optimize the context window, persist everything else" and ships 286 skills under it s2. The Reddit thread asking whether the project is worth it has a top reply saying the token usage of the descriptions is minimal s8. The measurement above is the test of that reply, and 27,000 tokens is not minimal.

Five pieces survive the audit.

AgentShield is a security auditor for agent configurations. It ships inside ECC and as its own npm package, where it took 5,539 downloads in the last week, with six dependencies, under MIT s7. One command audits a directory of agent files, hooks and server definitions. Its limit is a date: it was last published in March 2026 while the repository was updated the day before the measurement, and its README still describes the ecosystem at 42,000 stars against 250,466 s7. Run it; do not trust it to know about anything shipped after March.

The settings block is a ~/.claude/settings.json fragment: model set to Sonnet, thinking budget cut from 32,000 to 10,000, compaction threshold moved from 95 percent to 50 percent, subagents pushed onto Haiku. The README claims about 60 percent off model cost and 70 percent off the hidden thinking cost s2. Those are its figures, not ours. The trade is quality on hard reviews, which is why the block is worth copying with the model lines removed.

Instincts are the continuous-learning-v2 skill. The first version only watched at the end of a session; the current one watches before and after every tool call. Each observed pattern becomes an instinct with a confidence weight between 0.3 and 0.9, analysed in a background agent on Haiku so your main context pays nothing. Version 2.1 made instincts project scoped, the repository's own word for the problem it fixed being contamination, and an instinct seen in two projects is promoted to global. The tooling behind it is 2,200 lines with six commands, and instinct libraries export and import between machines s6.

The three context-accounting skills are a context budget, a compaction adviser and a token budget adviser. They show you the bill the rest of the repository runs up s3.

The selective installer reads manifests/install-profiles.json, which ships seven profiles: minimal, core, developer, security, research, opencode and full. The installer also takes component names, so you can name the four pieces above and nothing else s5. The profiles are coarse: even minimal carries the agents and the commands, so name components rather than picking a profile.

What stays in the repository. The 68 agents cost about 5,000 tokens and are mostly language reviewers and build fixers for stacks you do not write in; take the two that match your work s1. The 94 slash commands are called convenient entry points on the way to a skills-first surface by the README itself, and a folder of retired ones already exists s2. The hooks live in a single file of 291 lines and 40 KB holding 23 hooks, 8 of which run before every tool call and 7 every time the agent stops, each squeezed onto one line of roughly 33,000 characters you cannot read before you run it s4. That is executable configuration you have not audited, and no star count changes that.

Measurements

What people quote against what the repository held on 2026-09-06, version 2.2.1:

Inventory of a full install:

Standing context of a full install:

Protocol: clone affaan-m/ECC at version 2.2.1 on 2026-09-06. Count directories for agents, skills, commands and rules; count hook entries in hooks/hooks.json. Concatenate the name and description fields of every SKILL.md and every agent file, measure characters, convert to tokens. Compare against the figures circulating online and against the README of the same day. Nothing here measures answer quality; it measures what a full install costs before the first prompt.

Do this Monday

• Run npx ecc-agentshield against your ~/.claude directory and read every finding before touching anything else in this list.
• Open hooks/hooks.json in the ECC repo and confirm you cannot read a 33,000 character line; decide from that whether any hook from it enters your setup.
• Count your own standing context: concatenate the descriptions of every skill and agent you have installed and check the character total against the 85,000 that 286 skills cost here.
• Copy the settings block but strip the model lines; keep the compaction threshold at 50 percent for one week and note whether your sessions lose anything.
• Install continuous-learning-v2 alone with the installer's component names, not a profile, and let it run for five working days before judging it.
• Add the three context-accounting skills and read the context budget output at the start of your next long session.
• Pick at most two agents from the 68 that match your actual stack and ignore the rest.
• Put a recount date in your notes: the repository had 110 open pull requests on 2026-09-06, so every number above has a shelf life.

Go further

• Read the README's own line "optimize the context window, persist everything else" next to the 286 skills it ships, and decide which half you believe: ECC README.
• The instinct lifecycle, from a 0.3 weight observation to a promoted global skill, is documented end to end in the skill file itself: continuous-learning-v2.
• Study the seven install profiles and the component-name path before installing anything; it is the single file that makes the raid possible: install-profiles.json.
• AgentShield's npm page shows the 5,539 weekly downloads and the March 2026 publish date side by side; that gap is its limit: ecc-agentshield.
• The 23 hooks, 8 pre-tool and 7 on stop, are visible in one file; open it before you let it run: hooks.json.
• The Superpowers versus ECC debate is about two different objects, a workflow discipline against a configuration warehouse, and the thread shows why people conflate them: Superpowers vs ECC on Reddit.
• A working developer's setup after months of use is a useful contrast to a 286 skill install; note how few pieces it contains: My Claude Code workflow after months of daily use.

Sources

• affaan-m/ECC on GitHub, GitHub. Why read it: the live counters (stars, forks, issues, pull requests) are the only way to know how stale this pack has become.
• ECC README, GitHub. Why read it: 2,200 lines that state the component counts, the settings block claims and the admission that commands are temporary.
• ECC skills directory, GitHub. Why read it: the 286 folders whose descriptions make up the 85,000 character standing bill; the three context-accounting skills live here.
• ECC hooks.json, GitHub. Why read it: one look at the 291 line, 40 KB file settles whether you will run unreadable hooks.
• ECC install-profiles.json, GitHub. Why read it: the seven profiles and the component-name install path that let you take five pieces and leave the rest.
• continuous-learning-v2 skill (instincts), GitHub. Why read it: the full design of instincts, weights, project scoping and promotion to global.
• ecc-agentshield on npm, npm. Why read it: download count, dependency count and publish date in one place; the date is the caveat.
• Is everything-claude-code really that good?, Reddit. Why read it: the thread whose top reply the measurement in this pack was designed to test.
• Superpowers vs Everything Claude Code (ECC), Reddit. Why read it: the comparison people keep asking for, and the reasons it is a category error.
• My Claude Code workflow after months of daily use, Reddit. Why read it: a lean daily setup to measure your own install against.

FAQ

Does the 27,000 token figure apply if I use a bigger context window?

The characters do not change: 85,000 for skill descriptions and about 20,000 for agent descriptions. Only the percentage moves. The 14 percent figure is against a 200,000 token window; recompute it for yours.

Can I install just the five pieces without cloning 88 MB?

Yes. The installer reads manifests/install-profiles.json and accepts component names, so you can name AgentShield, the settings block, continuous-learning-v2 and the three context-accounting skills directly. Avoid the minimal profile; it still carries the agents and the commands.

Is AgentShield safe to rely on?

It is useful and it is behind. Last publish March 2026, with a README describing a 42,000 star ecosystem against 250,466 at measurement time. Run it as a first pass, not as a verdict.

Why not just take the hooks too?

Twenty-three hooks live in one 40 KB file, each on a single line of roughly 33,000 characters, with 8 running before every tool call and 7 on every stop. You cannot review that before you execute it, and the licence does not change the risk.