X post engine

x-post-engine — session receipts → post drafts

The playbook (docs/strategy/founder-content — the 6 formats): build log · documented failure · value post · receipt · contrast hook · milestone chapter. Formats are commodity; the INVENTORY is the moat. This skill's job is extracting the inventory from real work and drafting; the angle call stays with Max — that is the one non-delegable step. Never post anything.

VOICE (applies to everything this skill produces): first person singular, always — "I", never "we"; it's Max talking to his audience and it's intimate. The agents are "my agents", the company decisions are "I decided". Written like he talks: contractions, short sentences, direct address ("you"), asides allowed. Read each draft aloud in your head — if a sentence can't be SAID naturally, rewrite it. Best practices still hold underneath (numbers, hooks, receipts); verbal delivery, disciplined structure.

Inputs

The current session's events, or a named event/topic. Best sources: what shipped, what broke, what a number proves, what an agent did that humans find surprising.

Before drafting, read the Patterns section of docs/strategy/x-vault.md if it exists (only Patterns — the corpus below it is raw evidence, skip it). Patterns are market-validated structure (hooks, pacing, number use); apply them to the framing, never let them override Max's voice or the format rules below.

Information gain (optional — value posts + long-form only)

When a draft would land harder with an ORIGINAL founder story or real number that isn't this session's own receipt, invoke the information-gain skill (.claude/skills/information-gain) with the topic: it mines a niche YouTube transcript into a cited source brief (a real figure / failure / decision that exists nowhere else in writing). Weave in at most one, embed or link the source video, and only state numbers the transcript actually says. Skip this for build-logs and scars — those run on this session's own receipts, never mined material. Never let a mined number become a [FILL]-style fabrication.

Produce (append to docs/strategy/x-drafts/YYYY-MM-DD.md)

2-3 drafts, each tagged with its format + the target keyword. Rules per draft:

1. Keyword-first when the post has one (the GSC hack: X posts rank in Google — "GoHighLevel pricing —" as the opening words, not buried). Not every post has a keyword; build-logs and scars often don't. Never force it.
2. One number with context beats three numbers. "113 SEO pages in one day" is noise; "I typed 'build #1-8' and twelve AI agents shipped 113 pages to production before dinner" is a post.
3. Formats guide:
   • Build log: what happened today, present tense, unpolished, ends with what's next (people invest in unfinished things).
   • Scar: the failure with the embarrassing specifics kept in; the fix in one line at the end; never explain the lesson — let the reader get it.
   • Value post: steps + the WHY behind each step (advice without reasoning is a list); specific beats general ("how I got X" > "how to X").
   • Receipt: the screenshot IS the post — one line of text max. GSC/Bing graphs, revenue, the OG-card images from the calculators all qualify. Same graph monthly = a series people follow.
   • Contrast hook: two facts side by side, no explanation ("Solo founder. 12 AI agents. 113 pages shipped today."), reader resolves the gap.
   • Milestone: only with history to quote — quote the old post, add the new number. Flag candidates for later, don't force early.
4. Attach the visual when one exists — every draft gets a CREATIVE block, one of two types (never blur the line between them):
   • GENERATED — concept cards the loop renders itself (quote card, stat card, loop diagram): write an HTML card in docs/strategy/x-creatives/YYYY-MM-DD/<slug>.html and render with node scripts/x-creative-shot.mjs <html> [--size WxH]. In-feed posts: 1200x675 (16:9). X Article covers and inline images: 5:2 — use --size 1500x600. Styles that earn (from Max's bookmark corpus): document-as-image (a real-looking typeset doc/field-notes page — borrowed authority), workflow diagram with the REAL artifacts embedded and a payoff at the end, vertical phone-screenshot receipt. Dark stat cards for numbers. NEVER generate anything that imitates an authentic capture — no fake terminal windows, no fake analytics UIs, no fabricated numbers. Fabricated receipts break never-lies; that rule outranks engagement.
   • CAPTURE — real receipts only Max can take (terminal output, GSC/Bing graphs, Stripe, the product). Spec it exactly: which app/window, which content must be visible, what to crop, light/dark, and WHERE it goes (image 1 / image 2 / cover). "Plainness is the receipt" — no beautifying.
5. Paste-ready block — after each draft, a PASTE-READY fenced block Max can copy straight into X: Unicode sans-serif bold (𝗯𝗼𝗹𝗱) applied to AT MOST the hook line or the one key number — never whole sentences (screen readers read Unicode bold as garbage; sparing use is both taste and accessibility). NEVER bold the keyword line — Unicode chars aren't ASCII, bolding the keyword breaks its Google indexing. No underline in posts (renders unreliably). Line breaks exactly as they should post.
6. Plain words. No hashtags. No "excited to announce". No em-dash chains — X posts read spoken.
7. End the file with: "Pick one, adjust the angle, post. Repost winners in 6-8 weeks with updated numbers; quote-tweet anything that runs."

Long-form mode (X Article / blog, 1,500-2,000 words)

When invoked with "article" (or by the weekly loop), ALSO produce ONE long-form draft in docs/strategy/x-articles/YYYY-MM-DD-<slug>.md. Rules:

1. One article = one build with its receipts. Never a listicle of tips; the piece walks through something real we shipped, numbers kept in, failures kept in. The short-post formats scale up: a scar or value post at essay length.
2. Structure: hook (a number or a contradiction, first two sentences) → the problem as lived → the build in chronological order → what the numbers say → honest limits (what didn't work / what we don't know yet) → what's next. The honest-limits section is mandatory — it's the never-lies positioning in long form and the section readers quote.
3. Vault Patterns apply to the hook, the section openers, and every number (one number with context per point). Plain words, spoken rhythm, short paragraphs (X Articles render wide — 2-3 sentences per paragraph max). The article body must be PASTE-CLEAN plain text: zero markdown symbols — no #, ##, **, *, backticks, or bullet dashes anywhere in the body (they don't convert when pasted into the X editor and they scream AI). Section headings are short plain lines on their own; the FORMATTING MAP lists them so Max can style them as headings in the editor. Wrap the body between ===== ARTICLE BODY (paste everything between these lines) ===== markers; file metadata, the map, and tweets live OUTSIDE the markers.
4. Title = the article's best number or contradiction in the first six words; keyword up front when one exists (X Articles rank in Google); subtitle carries the second-best hook. Always provide 3 alternates in a comment under the title: A search-first (keyword leads), B feed-first (number or scar leads), C contrarian one-liner. Max picks per surface.
5. End with one line pointing at SeldonFrame only if the article earned it — the build IS the pitch; never bolt on a CTA.
6. Word count 1,500-2,000. Under 1,200 = make it a post instead; over 2,400 = cut the weakest section. Drafts only — Max posts manually (X Articles need Premium+ and his editorial pass).
7. Formatting map + creatives: X Articles have a real rich-text editor, so no Unicode tricks — end the article file with a FORMATTING MAP (exact lines to bold / italicize / underline in the editor, cover image, and where each inline image goes) plus CREATIVE blocks per the post rules above (GENERATED cards rendered to docs/strategy/x-creatives/, CAPTURE specs for receipts).
8. Concept visual (the sauce, drawn): every article gets 1-2 GENERATED diagrams that let a reader grasp the core mechanism from the image alone — flowchart of the loop, stack diagram, production line. Rules: (a) REAL specifics embedded (actual file names, numbers, tool names) — a generic boxes-and-arrows chart is AI slop and forbidden; (b) pass the rebuild test: someone could reconstruct the system from the diagram; (c) @themaxthule on the card — screenshots travel without attribution otherwise; (d) 5:2 via x-creative-shot.mjs --size 1500x600; (e) place it right after the section that explains the mechanism. Corpus evidence: the 5-ads workflow diagram with real artifacts did 1.1k likes / 901 replies — the diagram WAS the post.
9. Steal-this-prompt block: near the end of the body (before the closing), a plain-text prompt the reader can paste into their own LLM/agent to reproduce the build, delimited by "Copy everything between the lines" lines. Must be honestly runnable — test-shaped, concrete, no pseudo-prompt filler (never-lies applies to giveaways). This converts readers into repliers ("it worked") and is the never-taxes distribution mechanic: give the sauce, sell the platform.
10. 3 supporting tweets per article (in the same file, SUPPORTING TWEETS section, paste-ready): these are quote-reposts OF the article across the week — each takes a DIFFERENT door in: one number hook, one scar/insight line, one contrarian or value take. Never three rewordings of the title. Spacing note per tweet (e.g. "day 2 evening", "day 4 morning").
11. Cadence: 5 articles/week, each anchored to a real build from the week — the session logs, docs/learnings/, and the loops Max pastes into Claude Code daily are all legitimate build material. If a receipt (number, timing, cost) isn't in the material, mark it [FILL: what's needed] for Max — never invent a receipt to complete an article.

Cadence memory

Track (top of the drafts file) which formats ran recently — rotate; the receipt series (GSC/Bing graph) is monthly, anchored to real data only. Articles are weekly at most — skip the week if no build earned one; a forced article reads forced.