Loading today’s stories
Build1 distinct publisher2 min readPublished
Nobody tells you to uninstall an agent skill, so the pile only grows. One author finally priced his: the cost tracks the size of the install, not anything he asked the model to do.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
build
Config rot has a latency bill: a 70-line weekly audit for agent environments1 distinct publisher
product
A 2x LLM bill is not a bug report: token spend is an observability problem1 distinct publisher
build
Thirty minutes a day, and none of it from letting the agent write Swift1 distinct publisher
build
255 tools, 71,929 tokens: the standing charge hidden in your MCP config1 distinct publisher
The resident part is a router table. The model cannot select a component it was never told about, so every description has to be in the window at turn one, and the total therefore tracks how many things you installed rather than how many you use [4].
Divide it out and no single entry looks worth arguing about. 9,857 tokens across 160 components is about 62 tokens each [1]; skills average 70, agents 53, commands 26 [2]. That is how an install reaches 160 without a decision ever being made.
The measurement has two soft edges, and the author flags both. It is characters divided by four, so 9,857 tokens is really 39,428 characters of frontmatter text [9][5], and his position is that a real tokenizer moves the absolute figure a few percent and reorders nothing [9]. The scan also walks only three glob paths under the config directory [10]. Everything he lists as stacking on top of it (the system prompt, tool definitions, each MCP server's tool schemas, CLAUDE.md, and the files you actually needed to read) sits outside the 9,857 [8]. Treat it as a floor for one directory rather than a per-session total.
Trigger cost shows a skewed distribution: 322,990 tokens of skill bodies across 107 skills, median 1,932 [13]. The mean is 3,018, about 1.6 times the median [7], and the tail ends at continuous-learning-v2, 56,453 tokens per trigger, roughly 28% of a 200k window in one load [14][6]. The second heaviest description, meanwhile, is token-budget-advisor, which charges 209 tokens of permanent residency for the offer to save you some [11][12].
Some of that residency buys little. According to the author, continuous-learning-v2's CLI works by hand, but the automatic session observation that was the reason to install it depends on hooks hard-wired to a full plugin install and silently missing from a selective one [15]. The ck skill's session-start.mjs hook is not wired up either, so its cross-session memory never loads [17]. Both descriptions still load into the window at turn one, even though neither points at code that runs.
For the figure to transfer to your machine, your component count has to be in this range and your skill authors have to write descriptions of this length; the heaviest here is 244 tokens [11]. A twelve-skill install with one-line frontmatter is not paying 5% of a 200k window [6]. The mechanism transfers regardless, and the only quantity it gives you to turn is the count.
Ranked by verification strength, evidence, and original report placement.
The author measured the cost of that config at 9,857 tokens, paid on every session before typing a character.
The author describes 9,857 tokens as about 5% of a 200k context window, gone before anything happens.
Tokens in the author's script are estimated as characters divided by 4; he states a real tokenizer moves the absolute numbers a few percent and changes no ranking, which is why he avoided the dependency.
The measuring script is one standard-library file that defaults to ~/.claude and scans three glob paths: skills/*/SKILL.md, agents/*.md and commands/*.md.
delivery-gate, gateguard and safety-guard ship scripts that nothing registers in settings.json; they advertise automatic enforcement, and what was actually installed is documentation.
The ck skill's session-start.mjs hook is not wired up, so the cross-session memory never loads it.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 29, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Reproducible method, unreplicated result
The arithmetic is unusually checkable — the script is printed in full, the output table is shown, and the totals divide out consistently — but every figure describes one person's home directory, and nobody has re-run it. The premise doing the most work, that all 160 descriptions stay resident all session, is asserted from reasoning about how the model picks components, with no vendor documentation or prompt trace behind it. Strong on transparency, thin on independent confirmation.
One repo, one home directory
What can actually be observed spreading is small: an MIT-licensed script on GitHub and exactly one install that was measured and then pruned. There are no stars, forks, downloads, other people's numbers, or any sign that a second developer has priced their own config. Treat this as a technique released, not a technique adopted.
Framing outruns the sample, then apologises
Slightly overstated, and mostly in the generalising rather than the numbers. 'Rent you pay forever' does real rhetorical work on a figure the author himself calls not a catastrophe, and one laptop becomes a claim about what skills cost in general. Against that, the post volunteers its own weaknesses — the rule-of-thumb tokenizer, the modest 5% share, the admission that a well-aimed paragraph often beats code — which is more self-correction than this genre usually offers.
Self-published, with a repo and a sequel attached
No editor stood between the measurement and the headline. The post ends by pointing at the author's GitHub project and teasing a longer write-up covering a 284-skill pack cut to 104 and a 'rules-library trap' — so a bigger number and a more scandalous set of broken components both serve the author. Offsetting that: the licence is MIT rather than a lead-capture, the code is short enough to verify, and the argument is aimed at deleting things rather than buying them. The components named as dead on arrival were not asked for comment.
Single account, open method
Our confidence sits in the middle for a straightforward reason: the reasoning is fully exposed but the sample is one, the publisher is one, and the platform behaviour underneath has not been confirmed anywhere else in this reporting. The internal consistency is good enough that we would expect the ranking of costs to hold on someone else's machine; the specific totals should be treated as one data point until a second person runs the script.