Skill study

Study: shapes, not sentences

"Copy Paul Graham's style" was the most repeated instruction in the chat log this plugin came from, and the paid product answered it with adjectives ("plain, one thought, no structure") and then wrote 120-word essays. A style is learned from examples and stored as shapes: the moves a writer makes that another writer can make with different material. Sentences stay with their author.

1. Read

Through the browser capability, logged in as the user: the creator's LinkedIn recent-activity page (posts tab) and X profile. Collect 20–40 posts with dates and engagement. Sort by engagement relative to their own median; the top quartile is what to learn from, the bottom quartile is what to avoid. Save the raw pull to inbox/study-<handle>-<date>.md (private; it's for analysis, never for reuse).

If the browser is unavailable, search results expose first lines and comment counts; that is enough for openers and cadence, not for structure. Say which you had.

2. Extract shapes

For each top-quartile post, name the moves. Use the vocabulary in references/shape-vocabulary.md so shapes are comparable across creators:

• Opener: event / claim / number / question-to-self / list-promise / "one of…"
• Body: single observation / story → lesson / argument-you-lost / numbered steps / contrast (then vs now) / product episode
• Specificity: what only they could know — exact number, named person, dated event, their own mistake
• Ending: flat last line / callback to opener / one open question / CTA with price / p.s.
• Length band, paragraph shape, sentence-length spread, punctuation habits (em dashes? parentheses? "p.s."?)
• Cadence and mix: posts/week, text vs image, how often they sell
• What they never do (from the bottom quartile and from absence): hashtags? questions? emoji?

Write shapes/<handle>.md in the vault: 5–10 shapes with a one-line description and paraphrased evidence ("opens on a lost argument: 'One of the debates I've had for 15+ years (and pretty much always lost)…'"; a quoted fragment under ~15 words is fine as a pointer, a whole post is not). Add the creator to shapes/index.md.

3. Use

write reads shapes/ when the user asks for a shape ("PG-style comment", "a Dharmesh-style episode post") or when a lane in strategy.md names a reference creator. The shape supplies the moves; the user's story supplies every word of content; voice.md supplies the sentence-level voice. If a draft could be mistaken for the studied creator's post, it has failed: the reader should recognize the user.

3b. Public teardown (the creator-spotlight loop)

A teardown written for the user's own vault is a shapes catalog. A teardown written for the public is also distribution: the creator shares it ("someone studied my posts"), their audience reads it, and the page is a third-party-looking citation for AI search. The paid product's founders did this by hand as cold email with a personalized analysis; the plugin does it weekly as a page (templates/scheduled-tasks/creator-teardown-weekly.md).

A public teardown tells the story in pictures and keeps the text short: 400 to 600 words, first person, in the user's plain voice (the landing page docs/index.html is the model), run through a humanizer pass before handover. It copies the landing page's design and nothing else (the rule for every page on the site, written down in docs/DESIGN-NOTES.md and checked by tests/test_claims.py): section headings with the same coloured line icons, img class="shot" screenshots, ul class="rows" for small comparisons, plain tables. No metadata line under the title (no row of facts joined by middle dots; put those facts in a sentence), no legends with coloured bullets, no stat tiles, no cards. It carries (a) one beeswarm chart in the style of Linear Insights: one column per shape, every post a small dot on a log scale with dashed gridlines and the axis on the right, the median as a dotted line, one muted hue per column and the column name underneath, no legend. It is drawn by a small inline script that lays the dots out for the reader's actual width (so it stays readable on a phone), shows a preview card the moment the pointer is near a dot (the kind, the opening of the post as the feed shows it, its numbers, a link), opens the post on LinkedIn on click, and on touch shows the card on the first tap and opens on the second. Each dot needs the post's permalink: on the activity page open the post's control menu, choose "Copy link to post", and read the link by wrapping navigator.clipboard.writeText in the page first (never read the user's clipboard); the slug in the link repeats the post's first words, which checks that the link belongs to that post. Leave the preview text out for a post about people losing their jobs or anything else sensitive. (b) any cross-post measure such as comments per 100 reactions as a rows list, computed on posts at or above the median so tiny posts do not win on small denominators, with group sizes; (c) two or three screenshots of the creator's posts exactly as the feed shows them (collapsed text, the image, the reaction bar), from their activity page, with any third person's name and face and the user's own avatar blurred, never a post whose image is AI-generated (a content-credentials badge on it), stanley-vault image-clean on every file; (d) a collapsible table of every post with its numbers and your own short label. Images go to docs/creators/img/<slug>/. Never claim what the comments said unless you read them. Post times come free with the permalinks: the number at the end of a LinkedIn activity, share or ugcPost id, shifted right by 22 bits, is the posting time in milliseconds since 1970, so the teardown can compare the first and second post of the day, weekdays, or hours without guessing. Render the page at desktop and phone width and look at it before committing. Then: Markdown with front matter Markdown with front matter (layout: default, title, description (one line), creator, platform, followers read this run, posts_read, date); the site's Jekyll layout styles it, sections: what the numbers say (median and top three, read this run, nothing from memory), the shapes that worked with the screenshot that shows each, what got ignored, and "what a founder in could borrow this week". Say what did not work for them too; a teardown that is all praise is an ad. Never a sentence of theirs typed into the text (the screenshots carry their words, only as far as the feed shows them before "more"), never a private detail, never a guess about their business. Saved to docs/creators/<slug>.md in the repo named in my-human.md (teardown_repo:), linked from docs/creators/index.md, committed, not pushed. When the teardown repo is the Open Stanley repo itself (it contains open-stanley/skills/study/references/), add the same findings as a shapes-only entry at the top of open-stanley/skills/study/references/creators-latest.md in the same commit, replacing that creator's older entry there: tags from the shape vocabulary, the numbers read this run, no sentence of theirs. That file ships with the plugin, so every user's write and rituals get the finding at the next version bump. Any other repo: the vault's shapes/<handle>.md is where it stays, and /open-stanley:learn can propose it upstream. Run the teardown and the note through a plain-language pass before handing them over (no em dashes, no punchline lines, no bold lead-in lists, no claim the numbers don't show). The two-line note to the creator is drafted and left for the user's yes. Its "one thing surprised me" line must be a pattern across posts that the creator cannot see on any single post's analytics page (a whole class of posts landing above or below their median, a shape that changes comments per reaction), never "post A beat post B" or "your best post was X".

4. Ships with

references/creators-latest.md: one entry per creator from the weekly public teardowns, newest first; where a creator appears in both files, this one wins. references/creators-2026-09.md: shapes for Paul Graham (X), Dharmesh Shah, Yamini Rangan, Yuriy Zaremba, and Vitalii Dodonov, derived from live reads of their feeds on 8 Sep 2026. Re-study every quarter; shapes drift with platforms.