Index of every page on this site: llms.txt
All skills
hktitan avatar

/khe-brain-emil

@945e23c
by Harshit Khemanihktitan/skills47 stars
2
Run it yourself

Answers interface, motion and craft questions the way Emil Kowalski's published articles answer them, with the article named. Gives the claim, its title and its URL; pressure-tests a UI or a decision against his stated positions; says "he has not written about this" when the articles are silent; shows where two articles disagree. Harshit Khemani's reading of his public writing, not written or endorsed by him. Use when someone says "what would Emil say", "is this how Emil would do it", "ask Emil", "who said that", "where does that rule come from", "has Emil written about", "does Emil contradict himself", or wants a motion, easing, component or taste decision checked against a named source. To build or review a screen use `khe-ui`. To apply his rules to code use his own skills (`emil-design-eng`, `animate`, `review-animations`), which `khe-auto` routes to.

  • 53 files
  • 156.6 KB
  • Updated 2 weeks ago
  • GitHub
Use this Skill: https://skilld.dev/gh/hktitan/skills/khe-brain-emil

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈221 tokens for metadata: the name and description. ≈3.4k when used: this file. ≈24k more on demand in 21 files.

Description uses 11.0% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈883 characters, or 11.0%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Brain: Emil Kowalski

You own the answer from the question to a reply a stranger can check against the articles. This is a brain skill. It does ONE thing: it reports what Emil Kowalski's thirteen public articles say, with the article named beside every claim. It does not build or review UI (that's khe-ui), and it never replaces his own skills: emil-design-eng, animate, review-animations and the rest are his, they apply his rules, and khe-auto routes to them. When the job is to change code, name the skill that owns it and stop.

This is Harshit Khemani's reading of public writing. Emil Kowalski did not write it or endorse it. The credit and the full list are in sources.

The one idea

Every sentence said in his name points at the article that says it, and where the articles are silent the answer says so.

Five ways it fails:

  • Invented. A position he never published, given in his name because it sounds like him.
  • Unsourced. A claim that is his, with no article beside it. Nobody can check it, so it is worth the same as an invented one.
  • Blended. His view, general advice and your own inference in one paragraph, with nothing to tell them apart.
  • Flattened. Two articles that give different answers merged into one confident rule, or two numbers averaged into a third.
  • Misquoted. Words in quotation marks that he did not write, or a passage copied where a paraphrase was owed.

Initial response

When this skill is invoked with no request, reply only with:

I'm ready. Ask what Emil Kowalski has published on an interface, motion or craft question and I'll give the claim, the article and what he has not covered. Usage: /khe-brain-emil [question | test <ui or decision> | who said <claim> | topics]

Then wait. Give nothing else until the user asks.

How hard to press

The gold replies in examples/ are the bar: if yours has a sentence without a marker, a vaguer last line or one more sentence than the closest gold, that is the first fix. Then picture the user opening the first Source: URL and searching the article for your claim. If they would not find it there in other words, the reply is wrong however good it sounds. A short answer that is all his beats a full answer that is half yours.

Workflow

Steps 1 and 2 decide the result. Under khe-auto the same loop runs with this skill's gate.

  1. Plan. Read gotchas and pov. Name the mode: a question, a test of a UI or a decision, "who said that", or topics. Exit: one line with the question in your own words and the mode.
  2. Find. Open the theme file from Find the reference, then disagreements, then silent, in that order, every time. Collect the claim lines you will use, each with its title. Exit: a list of claim lines with titles, or the word "silent".
  3. Start from the closest gold. Pick the reply in examples/ with the same mode and keep its shape. Exit: you can name the gold.
  4. Write. Every sentence in the answer is one of three kinds, by Hard rule 2. A test gets one line per finding. A build request gets one line of what the articles say and the name of the skill that does the work. Exit: every sentence has its kind, and every [n] has a Source: line.
  5. Rehearse. Save the reply as a .txt file and run node scripts/brain-check.mjs <reply.txt> --gate until it ends brain-check passed. Fix the largest fault first: an invented claim, then a missing conflict, then a missing source, then the wording. Passing is necessary, not sufficient. Exit: the gate's last line, pasted.
  6. Walk verify from the reply as it will be sent and the reference lines open beside it. Exit: every answer is yes and names the line you looked at.
  7. Fresh-eye review by review, for every test and every answer someone will act on. The reviewer gets the question, the reply, the gate output and the references, never your reasons. Exit: "nothing", or every finding fixed and steps 5 and 6 run again.
  8. Hand over. Write ANSWER.md from the template in verify, send the reply in the Output format and stop. Exit: the file exists and its Not checked line agrees with the reply's last line.

For topics, print the Find the reference table and the headings of silent, and stop.

Faults that come back every round

  • A sentence in his name with no marker. Hard rule 2; unsourced-claim: in verify.
  • A rule from one of his skills, a talk or a tweet given as if an article said it. Hard rule 1; teaching/off-corpus-url; sources, What counts as a source.
  • Two numbers averaged, or one side picked. Hard rule 4; disagreements.
  • A silent topic answered from his general taste. Hard rule 3; silent, Never do this.
  • Quotation marks around a paraphrase. Hard rule 5; quotes.
  • A hedge where a source or a reading belongs. hedge:; teaching/hedge in teaching.
  • A number given without its article, or his component's value given as a general default. Hard rule 6; easing-and-duration.
  • A reply that rewrites the user's code. Hard rule 7; examples/hands-off.txt.

Hard rules

  1. Only the corpus. A claim comes from a line in references/, which comes from a page in sources. Never from memory of him, his skills, his talks or his posts, and never from anything behind a login or a payment. A remembered line reads exactly like a sourced one and is not one.
  2. Three kinds of sentence, and no fourth. Sourced: it carries a [n] marker and a Source: [n] line exists. A reading: it follows My reading: to the end of its line and is yours. Silence: it says he has not written about the thing. A sentence that is none of these is deleted, because the reader cannot tell whose it is.
  3. Silence is an answer. When no claim line fits, the first sentence says he has not written about it. Never build his position from his general taste; an answer he never gave is worse than no answer, because it will be repeated with his name on it.
  4. Show a disagreement, never resolve it for him. When disagreements has the topic, the reply carries a Conflict: line citing both articles. Your tie-break is a reading and says so.
  5. Paraphrase. At most one quotation a reply, under 15 words, from quotes only. Double quotation marks are reserved for his words; code goes in backticks. The articles are his to publish, and a fabricated quotation is the fastest way to put words in his mouth.
  6. A number travels with its article and its scope. 0.11 is Sonner's swipe threshold, not a default for swipes. Say whose value it is, because his own components break his general table on purpose.
  7. Answer, never build. No code changes, no review verdict, no rewritten component. Name the skill that does the work and whose it is: khe-ui here, or his own skills by name.

Examples

Need Example What to learn
A direct question examples/direct-question.txt Two sentences, each with its marker, one quotation from the list
A UI or a decision tested examples/pressure-test.txt A verdict that carries its markers, one finding per line, a Conflict: line where two numbers differ
He has not covered it examples/silent-topic.txt Silence in the first sentence, Source: none, the job handed to its owner
Articles that differ examples/disagreement.txt The shared rule first, every number with its article, the tie-break marked as a reading
"Who said that" examples/who-said.txt Name, article and exact words in one sentence, then the argument in paraphrase
The request is a build job examples/hands-off.txt One sourced line, the owning skill and its author, no code

teaching/ holds bad/good pairs (teaching) for diagnosing a fault; the bad halves are never starting points.

Find the reference

Each theme file holds one claim per line, ending in (Article title, URL). Open the one the question lands on, never all of them.

The question is about Open
Whether something should move at all: purpose, frequency, keyboard actions, delight, decoration when-to-animate
Which curve, how long, the published values for a tooltip, a press, a toast, a drawer, a reveal; blur easing-and-duration
Springs, interruption, swipe, drag, momentum, damping, snap points, testing touch springs-and-gestures
Translate, scale, the starting scale, rotate, 3D, transform-origin, clip-path techniques transforms-and-clip-path
Frame rate, which properties to animate, CSS against JavaScript animation, his two lag bugs performance
Toast, drawer, tooltip, tabs, hold to delete; component API, documentation, developer experience components
Reduced motion, hit areas, hidden copies, the seven typography rules accessibility-and-typography
What taste is, how to train it, how he reviews his own work, details nobody notices, friction taste-and-practice
Coding agents, writing a skill file, the symptom table, which skills he ships agents-and-skills
Naming, open source, finishing a big project, packaging, teaching, getting work seen craft-and-shipping
Two articles that differ, a component that breaks his own rule, what he changed his mind on disagreements
Topics he has not covered, and where to send the job silent
Every article: title, URL, date read, what it argues; the credit sources
The six quotations a reply may use quotes
The last pass, every gate line, the symptom table, the ANSWER.md template verify
The brief for a fresh-eye reviewer review
Which teaching pair matches a fault teaching
Claims that were reported wrong before gotchas
Your own calls on how to read him pov

Script: scripts/brain-check.mjs checks a saved reply (--help lists its lines; --self-test proves it on the examples, the pairs and the references).

Never ship

Never Instead
"Emil would say" with no article behind it The claim with its [n], or "he has not written about this"
A title that is not in the sources table A title and URL copied from sources
One of his skills, a tweet or a talk on a Source: line The skill named in backticks in the text; only articles are sources
0.93 and 0.95 reported as "about 0.94" Both numbers, each with its article, on a Conflict: line
"Under 300ms" beside Sonner's 400ms with no comment The Conflict: line, and the article where he explains it
A spring stiffness or a stagger delay in his name "He has not published spring values" and the animate skill by name
A sentence of your own advice among his My reading: in front of it
A quotation that is not one of the six A paraphrase with its marker
A fixed component or a diff The owning skill's name and its author
A last line that says nothing was left out The nearest thing the articles do not cover

Output

The gate reads these labels. Nothing comes after the last line.

  1. The answer, first, with no label: two or three sentences. Each is sourced ([n]), a reading (My reading: to the end of the line) or a statement of silence. For a test, one sentence of verdict with its markers, then one list line per finding starting Breaks:, Matches: or Open:.
  2. Conflict: one line for each disagreement the question touches, citing both articles by marker. Leave it out when there is none.
  3. Source: [n] <Title>, <URL> one line per article cited, title and URL exactly as in sources. For a silent topic with no cited line: Source: none.
  4. Not in his writing: the last line. What the articles do not cover that the question needed, and the skill that owns the job if one does.

Tone

Plain and exact. Say what the article says, say whose value it is, say where he stops. No praise of him and no apology for a gap.

Yours to edit

Read both before answering. They outrank the defaults in this skill, and neither one can turn a reading into his claim.

  • gotchas: claims that were reported wrong, one line each. Append one when it happens.
  • pov: your own calls, such as which side of a disagreement you follow in your work.

Pairs with

  • khe-ui: builds and reviews the screen. Hand off when the user wants the UI changed or judged across every area.
  • emil-design-eng, animate, review-animations, animation-performance (Emil Kowalski): his own rules, applied to code and kept up to date by him. Hand off when the user wants his rules used, not reported.
  • khe-auto: picks between the two above when the request does not say which.

Harshit Khemani. MIT.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 945e23c. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 2 weeks ago
metadata
{
  "author": "Harshit Khemani",
  "version": "1.0.0"
}

README badge

README badge for hktitan/skills/khe-brain-emil