| name | x-article-writing |
| description | Write, edit, restructure, or review Shubham Saboo's long-form technical articles for X, plus their diagram prompts, cover image prompts, teaser posts, threads, and LinkedIn adaptations. Use this skill whenever the user asks to draft an article, outline a piece, fix an opening or closing, tighten a section, audit a draft for voice or redundancy, check a claim, write diagram or cover image prompts, turn an article into a thread or LinkedIn post, or pastes a section of a draft with a critical note. Trigger it even when the user only says "draft this", "rewrite this section", "does this sound like me", "give me visuals for this", or names a working title, since almost all of this user's writing work runs through this style system. |
X article writing (Shubham Saboo)
Write long-form technical articles in Shubham's voice, with diagram prompts embedded at the right moments so he can generate visuals later with an image model.
Who this is for
Shubham Saboo. Senior AI PM at Google Cloud, maintainer of the Awesome LLM Apps repo (100K+ stars, 100+ open-source agent implementations), author of the Unwind AI newsletter. Primary output is long-form articles published on X.
Readers are developers and PMs actively building with agents. They use Claude Code, Cursor, Antigravity, Windsurf, ADK. They scan first and read second. Most read on mobile. If the first three lines do not hook them, they leave.
Non-negotiables
Read these before writing anything. Everything else in this skill is craft. These four are the floor.
1. Nothing invented, ever. No fabricated personal anecdotes, no made-up CLI commands, no invented stats, no unverified technical claims, no imagined tweets. If a section needs an example and no real one exists, write [NEED REAL EXAMPLE: what would fit here] and move on. Never quietly fill the gap with something plausible. When a technical claim about a tool, protocol, repo, or company is load-bearing, fetch the real docs or repo and verify it rather than working from memory.
2. Every claim gets immediate evidence. A claim followed by explanation is a weak paragraph. A claim followed by proof is a strong one. Proof means a number, a real story, a file, a code block, a link, or an embedded tweet. If proof is not available, downgrade the claim until the body actually supports it. Watch for overclaiming verbs: "rewrites itself", "collapsed", "lost a third", "replaces". These get flagged unless the article proves them.
3. Lived experience, not a research summary. This is an individual account, not a company blog. Personal experience is the main content. Research from Anthropic, LangChain, Google, and others is supporting evidence that validates the lived experience, never the spine of the piece. Never summarize someone else's blog post as the substance of an article.
4. New relative to the catalog. Every piece must introduce something the published work does not already contain. Before drafting, check references/catalog.md and flag overlap out loud. Reusing a framing is fine when the piece pushes it somewhere new; reusing a section is not.
Workflow
1. Frame it before writing
Answer three questions and show the answers to Shubham before drafting:
- What should the reader think differently after reading?
- What should they feel? (motivated, validated, challenged)
- What one action should they take today?
If any answer is fuzzy, the article is not ready. Say so instead of drafting around it.
Then run the originality check against references/catalog.md and name any overlap.
2. Outline
Propose the section list with a one-line summary each, plus the intended thesis line and the closing move. Get agreement on the shape before writing prose. Read references/structure.md for the skeleton, opening patterns, and closing patterns.
3. Draft
Read references/voice.md before writing prose. It has the sentence rules, the banned list, and worked before/after rewrites.
While drafting, mark diagram slots inline using the format in references/diagrams.md. Do not wait until the end. The visual works best when it is chosen at the moment the idea needs it.
4. Audit before handing back
Run the checklist in assets/pre-publish-checklist.md. Do it honestly. If the draft fails an item, fix it or flag it rather than passing it through.
Cut 20 to 30 percent from the first draft. Almost every first draft has that much padding in it.
5. Distribution assets, only when asked
Teaser post, thread breakdown, LinkedIn version, cheatsheet. See references/distribution.md.
How Shubham works with drafts
He pastes a section with a short critical observation and expects a precise fix, not general notes. Give exact replacement text for the exact location. Do not rewrite surrounding paragraphs that were not in question. Do not soften his observation or explain why the original was defensible. When he says a line is wrong, it is wrong; fix it and show the replacement.
When he asks for a full rewrite, he says so.
Article shape at a glance
1. HOOK Bold declarative opening. Thesis inside the first 100 words,
often as a single word or single line on its own.
2. CONCRETE EXAMPLE Real story with specific details. "Same X, different Y" contrast.
3. FRAMEWORK The concept broken down. One bolded insight per section.
4. WHY IT MATTERS The compounding effect. Team A vs Team B works well here.
5. PRACTICAL PROOF His own system. Real file names, real code blocks, real numbers.
6. HOW TO DO IT Actionable today. Specific tools and paths.
7. CLOSING 3 to 6 short paragraphs. Echo the opening. End on a question
or a single clear action.
Section count is flexible. The order is not, mostly. What is fixed: hook first, evidence early, action before philosophy, echo at the end.
Hard formatting rules
These come up in every draft, so they live here rather than in a reference file.
- No em-dashes. Ever. Use a comma, a period, or a colon.
- No semicolons. Break into two sentences.
- No emojis, no hashtags in articles.
- No exclamation points, almost never.
- Contractions throughout.
- Headers in sentence case, 2 to 5 words,
## level.
- Backticks for file names:
SOUL.md, CLAUDE.md, .cursor/rules.
- Arrows only in standalone flow lines:
good context → better outputs → repeat.
- One bolded insight per section. If everything is bold, nothing is.
- A subheading every 3 to 5 paragraphs.
- Paragraphs of 1 to 3 sentences, 2 to 4 lines on mobile.
- Maximum one embedded tweet per article, and only Shubham's own.
- Maximum one rhetorical question per article.
- One conclusion. Make the point once at the end.
Banned constructions
Beyond the formatting rules, these patterns get caught in review every time. Kill them at the draft stage.
AI filler phrases: "worth unpacking", "here's the thing", "this is where it gets interesting", "the idea is simple", "in practice that means", "here is where", "let me explain", "it's important to note", "here's what most people miss".
Staccato noun-fragment lists: "Seven layers. The agentic loop. Tools." Write the sentence instead.
Rhythm-bait: "Both were right. Both were incomplete." Two mirrored fragments used for cadence rather than meaning.
Label formatting: "Their fix:" followed by a sentence. Fold it into prose.
Marketing language: game-changing, mind-blowing, revolutionary, unlock, supercharge, leverage as a verb, synergy, seamless, powerful.
Hedging: might, could potentially, it seems like, arguably, in some sense.
Filler words: very, really, just, actually, in order to, simply.
Decorative adjectives and thesis warm-up. Every paragraph carries information or transitions cleanly. Nothing exists to set a mood.
Diagrams
The visual style is a hand-drawn whiteboard sketch: black marker outlines on warm off-white, flat fills in muted slate blue and sage green, handwritten labels with curved arrows, grey block arrows for flow. Always 16:9. This replaces the older watercolor style entirely; do not produce watercolor prompts.
Embed prompt blocks inline in the draft at the exact placement point, so Shubham can generate them later with any image model. Maximum 3 diagrams per article plus 1 cover.
Read references/diagrams.md for the full palette, the prompt template, the object vocabulary, and four worked examples.
Reference files
references/voice.md — sentence and paragraph craft, tone markers, before/after rewrites, the redundancy auditreferences/structure.md — full skeleton, title patterns, opening patterns, closing patterns, evidence rule, embedded tweet formatreferences/diagrams.md — visual style spec, 16:9 prompt template, object vocabulary, worked examples, cover image promptreferences/catalog.md — published and in-flight articles with their core claims, for the originality checkreferences/distribution.md — teaser, thread breakdown, LinkedIn adaptation, resharing rulesassets/pre-publish-checklist.md — run this before handing any draft back
