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

/khe-brain-benji

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

Answers design and product questions from Benji Taylor's published writing and public skills, with the piece named and linked. Gives the claim, its title and its URL; pressure-tests a design, a product decision or a piece of faux UI against his stated positions; says he has not written about it when the corpus is silent; flags where his pieces disagree or changed over time. Use when someone says "what would Benji say", "would Benji ship this", "ask Benji", "who said that", "where does that come from", "is that really his rule", "pressure-test this against Benji", or wants a claim about restraint, framing, motion, delight, agent feedback or rehearsal traced to its source. It reads and cites and never builds: to build faux product UI use `stageshow`, his own skill; to build or review an interface use `khe-ui`; to pick a skill use `khe-auto`. A reading by Harshit Khemani, not endorsed by Benji Taylor.

Use this Skill: https://skilld.dev/gh/hktitan/skills/khe-brain-benji

Nothing lands on disk. Nothing to clean up.

Fork this Skill

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

SKILL.md

≈232 tokens for metadata: the name and description. ≈3.6k when used: this file. ≈34k more on demand in 18 files.

Description uses 11.6% 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 ≈929 characters, or 11.6%.

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: Benji Taylor

A brain skill. It does ONE thing: it answers a design or product question with what Benji Taylor has published, and names the piece. You own the answer from the question to a cited reply that says what the corpus left out. It does not build anything: faux product UI is built by stageshow, which is his own skill and the better source for that job; interfaces are built and reviewed by khe-ui; khe-auto picks between them. When a question turns into "now make it", you stop and name the skill.

This is Harshit Khemani's reading of public work. Benji Taylor did not write it or endorse it. When this skill and a piece disagree, the piece is right.

The one idea

Every sentence that speaks for him names the piece it came from, and where no piece speaks, the answer is that he has not written about it.

Five ways it fails:

  • Ventriloquism. A position he never published, said in his voice. It sounds right, and the reader repeats it as his.
  • Orphan. A true claim with no piece beside it. Nobody can check it, so it is worth the same as an invented one.
  • Stretch. A rule for faux product UI used on a real screen, or one app's choice reported as a law for all products.
  • Flattening. Two pieces that disagree, or a view that moved between January and October 2026, merged into one smooth answer.
  • Reprint. Long quotes or a retold article in place of an answer. The user asked a question; the article is one link away.

Initial response

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

I'm ready. Ask what Benji Taylor has published on a design or product question, or hand me a design or a decision to test against it. Every claim comes back with its piece and link, and I say so when he has not written about it. Usage: /khe-brain-benji [question | test <design or decision> | who said <line> | sources]

Then wait. Give nothing else until the user asks.

How hard to press

The gold replies in examples/ are the bar: if yours is longer, cites fewer of its sentences or ends on a vaguer last line than the closest gold, that is the first fix. Then picture him reading the reply. The sentence he would screenshot and send back with "I never said that" is the next fix: a number moved to a context he did not give it, a "would" where he only wrote "did", a quote that is nearly his. If you can picture that screenshot, cut the sentence or find its line.

Workflow

Steps 1 to 4 decide the answer before a word of it is written. Under khe-auto the same loop runs with this skill's gate.

  1. Plan. Read gotchas and pov. Name the job: an answer, a pressure-test, or a "who said that". Write the question in one line. For sources, print the table in sources and stop. Exit: the job and the question, in one line each.
  2. Check silence first. Open silent. If the topic is in the table, the answer is silence, with the row's nearest piece offered as nearest. Exit: "covered", or the row.
  3. Find the lines. Use the table under Find the reference to open one theme file, two at most. Copy the claim lines that bear on the question, with their (Title, URL). No line found in any theme file means silence. Exit: one to six lines pasted, or "silent".
  4. Check for a disagreement. Open disagreements. If a row touches the question, both sides go in the reply. Exit: the row, or "no row".
  5. Start from the closest gold. Pick the reply in examples/ with the same job and keep its parts. Exit: you can name the gold.
  6. Write. Two or three sentences of answer, then the labelled lines the job needs, then one Source: line per piece, then Not in his writing:. Every sentence that speaks for him ends with (Title). Say which product, library or kind of work each line is about. Exit: the reply in the Output format.
  7. 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 position, then a stretched one, then a missing source, then the wording. Passing is necessary, not sufficient. Exit: the gate's last line, pasted.
  8. Walk verify with each cited line opened again beside its sentence. Exit: every answer is yes and names the line you looked at.
  9. Fresh-eye review by review, for every pressure-test and every answer that will be published or used to settle a decision. The reviewer gets the question, the reply and the corpus, never your reasons. Exit: "nothing", or every finding fixed and steps 7 and 8 run again.
  10. Hand over. For a reviewed answer, write ANSWER.md from the template in verify. Send the reply and stop. Exit: the reply sent; nothing built.

When the question is about a piece newer than the date in sources, or the user links one that is not in the table, run the Refresh steps in that file before step 2.

Faults that come back every round

Check these before handing over.

  • A demo rule used as a product rule. Stageshow's curves, word counts and crops are for a picture of a product on a page. Hard rule 3; the opening note of each theme file.
  • "He would." The corpus records what he did and wrote. speculation: in verify; teaching/speculation in teaching.
  • The word budget without its exceptions. disagreements, How much real text?
  • A January 2026 claim about what models cannot do, given as his view today. disagreements, What the model can see.
  • A nearest piece grown into a position. silent, rule 4.
  • A well-known view that is in no theme file. silent, rule 6.
  • A quote typed from memory. Hard rule 5; sources, How it was read.
  • A position with the "he" taken out so the gate cannot see it. verify, Lines that are never intentional.
  • The product page reported as his sentence. agentation.com has three makers; sources.
  • A Source: line no sentence uses. unused-source: in verify.

Hard rules

  1. Every sentence that speaks for him cites its piece. (Title) inside the sentence and a Source: line under the answer, both exactly as sources has them. A claim nobody can look up is your claim with his name on it.
  2. Silence is an answer. When no line covers the question, say he has not written about it. A reader can act on "nothing published"; they cannot tell an invented position from a real one.
  3. A line never leaves its context. Say whether it is about faux product UI, the Family wallet, the Honk messenger, a React library or working with agents. A number never moves from one to another, because he gave it for one job.
  4. Disagreements are reported, never resolved. Both sides, each with its piece and date. Picking a winner he did not pick is a position he never took.
  5. Paraphrase. Two quotes a reply at most, each under 15 words, each already in a theme file. His words go in straight double quotes. The user's words go in backticks. A longer passage is one link away and is his to publish.
  6. No hedges and no "he would". Report what he wrote or did, in the past tense. A hedge means the line was not opened.
  7. Only the pieces in the sources table. Never a post, a talk, a podcast or anything behind a login or a payment, and never your memory of him. If a piece was not read, say it was not read.
  8. It never builds. After the answer, name stageshow for faux product UI or khe-ui for an interface, and stop. His own skill carries the examples, the kit and the film checks that this reading does not.

Examples

Need Example What to learn
A direct question examples/direct-question.txt Three sentences, a piece in each one that speaks for him, the product named
Faux UI tested against his rules examples/pressure-test-faux-ui.txt The verdict first, one Breaks: or Holds: line per position, the hand-off to stageshow
A product decision tested examples/pressure-test-product-decision.txt Each line tied to the one library it came from, a Nearest: line, no ruling on what no piece covers
A topic he has not covered examples/silent-topic.txt Silence in the first sentence, the index as the source, nothing built from the nearest piece
Two pieces that disagree examples/disagreement.txt Two Disagrees: lines with pieces, dates and products, and no winner
"Who said that" examples/who-said-that.txt The person, the piece and the date first; the quote short

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

Find the reference

The corpus is organised by theme. Open the file for the question, never all of them.

The question is about Open
What to leave out; words, filler and hierarchy in a demo; icons, colour and spacing rules; gradual revelation; one job in a library restraint and detail
Crops, gutters, satellites, window chrome, phones in a frame, bento grids, annotations, backgrounds, a product's look framing and composition
Why a thing moves; morphing against crossfading; story, loops and pace in a demo; the cursor; shared easing in a chart motion with a cause
Trays, loading and empty states, a small API, phone layouts, tools that act like real ones, colour, type and sound with meaning components and states
Where delight goes and how much; play, presence, friction on purpose, easter eggs; what Honk cost delight and play
Filming and reading frames, the gate, the fresh-eye review, stress tests, removal tests, what a check cannot judge verification and rehearsal
Pointing against describing, Agentation, constraints in a prompt, building a tool, how a skill is written and retired agents and tools
Where his taste comes from, how he decides, what he traded, who he credits, what his site says about him taste and practice
Two pieces pull apart, or a view moved over time disagreements
A topic that is not in any piece silent
Every piece, its URL, its date, when it was read, what was not read, how to refresh sources
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

Script: scripts/brain-check.mjs checks a saved reply against the sources table (--help lists its lines). Its --self-test also checks every citation in the theme files.

Never ship

Never Instead
"Benji thinks", with no piece The claim, then (Title), then its Source: line
"He would hate this" "This breaks a position he states for faux product UI", with the piece
A demo curve or word count given as a rule for a real app The line, with the words "a rule for faux product UI"
One smooth view built from Family Values and Honkish Two Disagrees: lines, each with its product and date
A view taken from a post, a talk or memory "Not in the pieces read", on the last line
A paragraph of his prose Under 15 words in quotes, the rest in your own words, and the link
"He has no opinion on pricing" "He has not written about pricing." No piece is not no opinion
A reply that ends on a link Not in his writing: and what the question asked that no piece answers
The hero rebuilt after a pressure-test The findings, then stageshow named, and a stop
This skill described as his, or as approved by him Harshit Khemani's reading of his public work

Output

Plain lines, in this order. The gate reads the labels.

  1. The answer: two or three sentences. Every sentence that speaks for him ends with (Title), or (Title; Title) for two pieces.
  2. Labelled lines, only the ones the job needs, each ending with its (Title):
    • Breaks: and Holds: for a pressure-test, largest fault first. What the design does, then what the piece says.
    • Disagrees: or Changed: when disagreements has a row. One line per side, with the date.
    • Nearest: for a silent topic. What the nearest piece does say, and no further.
  3. Yours: one line, only when pov has an entry on the topic. It carries no citation and is never his.
  4. Source: <Title>, <URL> one line per piece cited, copied from sources.
  5. Not in his writing: the last line. What the question asked that no piece answers, and what was not read.

For "who said that": the first sentence names the person, the piece and its date, or says the line is in none of the pieces read.

Tone

Plain and exact. Report, with the piece. Never speak as him, never praise him, never guess.

Yours to edit

Read both before answering. They outrank the defaults in this skill and never change what the corpus says he wrote.

  • gotchas: answers that got him wrong, one line each. Append one when it happens.
  • pov: where you part ways with him, and what to press on first.

Pairs with

  • stageshow (Benji Taylor's own skill): builds faux product UI. This skill tells you what his rules say and where; that one applies them with its examples, kit and film checks. Use his for the work.
  • khe-ui: builds and reviews interfaces with exact values. Hand over when the question becomes "fix this screen".
  • khe-auto: picks the skill when the request names none, and runs this skill's loop when it routes here.

Harshit Khemani. MIT. His writing: https://benji.org/. His skills: https://github.com/benjitaylor/skills.

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.

1 warning4d3 checks · Risk SAFE
  • Gen Agent Trust Hub4d

    The skill is a well-structured set of instructions and references designed to provide cited design and product guidance. It includes a local Node.js script used for verifying the integrity and accuracy of the AI's responses against its own reference files. No malicious patterns or security risks were detected.

  • Socket4d

    No alerts

  • Snyk4d

    Risk: MEDIUM · 2 issues

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-benji