blog-voice

name	blog-voice
description	Write or edit blog/TIL posts in Khaled Garbaya's voice, then strip AI-writing tells. Use when asked to "draft a post", "write in my voice", "outline this topic", "make this sound like me", "de-slop this", "blog-voice", or when the brain-to-blog loop turns a topic signal into prose. Combines Khaled's brand voice with the vault's "Signs of AI Writing" field guide.

blog-voice — write like Khaled, not like an LLM

Two jobs, always in this order: (1) hit the voice, then (2) strip the AI tells. A post that's slop-free but voiceless is generic; a post in-voice but full of tells reads as AI. You need both.

Sources this distills: Khaled's published posts (src/content/posts/) for the voice, and the vault note "Signs of AI Writing — Field Guide" (09 WIKI/, via qmd) for the anti-slop pass. When drafting a real post, re-read the topic's source note in the vault.

Who's writing

Senior Engineering Manager. Deep AI-augmented-dev practitioner who treats agents like a team he manages (specs, plans, reviews, phase gates). Writes for other senior engineers and EMs. Credibility comes from having shipped the thing — real projects, real numbers, real friction — not from authority claims. Anti-hype by temperament.

The voice (the positive spec)

Match these. Examples are pulled from his own posts.

1. Direct, declarative, short. Land the point, then stop. Fragments are fine for emphasis.

   "It does not ship production software with a team." "Partly true. Getting specific about why changed what I needed to fix."

2. First-person and experience-grounded. Open from something he actually did; concrete specifics (numbers, timeframes, what happened).

   "A small team I led recently shipped a greenfield project in three weeks." "I spent the last few months working on L8 AI engineering… When I applied these with a team of five engineers on a 2-week project at work, I hit friction I did not expect."

3. Set up a foil, then subvert it. State the obvious framing, then take it apart. This is his signature move.

   "My first instinct was that these workflows are built for solo developers. Partly true." → then: "The real divide is not solo vs. team."

4. Name things. Coin concrete labels for concepts and sections — they're the spine of the piece: APE (Agentic Product Engineering), L3/L5/L8, The L8 bottleneck trap, The L5 false confidence problem. Section headers are claims, not nouns.

5. One bold thesis sentence per section. Bold the load-bearing line, sparingly — never whole phrases mechanically.

   "The difference is not the AI. It is the discipline around the AI."

6. Anti-hype, undersell-then-reveal. Deflate before you build up. Earns trust.

   "most of it has been… fine. Helpful for boilerplate, decent for explaining code, occasionally useful for debugging. But last week, something clicked."

7. Comparison tables for structured contrasts (vibe-coding vs APE; L-levels). Use when you're contrasting two things across several dimensions — not as decoration.

8. Credit your sources by name. Karpathy, Yegge, Addy Osmani, Mat Pocock. Attribution is part of the voice, and it's honest. If the topic's vault note summarizes someone else's article, the post must be Khaled's own take — credit the original, make the manager's- eye application the novel contribution. Never rehash.

Cadence: vary sentence length. A run of short declaratives, then one longer sentence that does real work. Read it aloud — if it drones, break it.

The anti-slop pass (strip the AI tells)

After the draft hits the voice, sweep for these. Full taxonomy in the vault field guide; this is the working checklist.

• The core test — regression to the mean: did I trade a specific fact for a generic superlative? Put the fact back. "a revolutionary titan of industry" → the actual, specific thing they did.
• Kill the AI vocabulary cluster: delve, boasts, tapestry, vibrant, crucial, pivotal, underscore, showcase, foster, landscape, testament, seamless, robust, leverage, realm, navigate (figurative), "valuable insights," "deeply rooted," "rich tapestry."
• Cut trailing "-ing" analysis clauses tacked onto facts: "…, underscoring its role," "…, reflecting its relevance," "…, highlighting the shift." Just end the sentence.
• Restore copulatives: "serves as / stands as / represents / marks" → plain is/has.
• Delete significance puffery: "marks a pivotal moment," "represents a significant shift," "part of a broader movement." Show the thing; don't announce its importance.
• Kill the "Despite its challenges… continues to thrive" formula and rigid "Challenges / Future Outlook" sections. The formula is the tell.
• Vague authorities: "experts argue," "researchers say," "it's widely regarded." Name who, or cut it.
• Un-bold everything except the one thesis line per section. No mechanically-bolded key terms.
• Em-dashes: allowed but earn them. Mechanical space-surrounded em-dashes as "punched-up emphasis" are a tell — if a comma or period works, use it.
• No chatbot leakage: "I hope this helps," "Certainly!," "You're absolutely right," "Let's dive in," "In conclusion," "It's worth noting that."
• Rule of three: one purposeful triple is fine; reflexive "adjective, adjective, adjective" to sound comprehensive is slop.
• Straight quotes in code; verify no citeturn… artifacts, unfilled [placeholders], or "as of my last knowledge update."

The attribution pass (trace borrowed claims to their source)

Runs after the anti-slop sweep, before publish. The trap it catches: a paragraph that reads like Khaled's own reasoning but was lifted, paraphrased, from an article he ingested. Uncredited, that's plagiarism-by-paraphrase — the honesty line, not just etiquette. Attribution is also part of the voice (see voice #8).

The vault makes this recoverable: ingested articles live in 00 Inbox/ with source: frontmatter carrying the original URL, and wiki notes list them under sources:. So every borrowed claim can be traced back to a real link.

Procedure:

1. Flag every borrowed claim. For each non-obvious fact, framing, quote, or number in the draft, ask: is this mine, or did it come from something I read? If it came from a source, mark it.
2. Trace it to the URL. Follow the topic's wiki note sources: → the 00 Inbox/ article → its source: URL (grep -i '^source:' "00 Inbox/<article>.md").
3. Link the ones worth it. Inline-link the primary source(s) and any framing/number you're leaning on. Credit the author by name where natural ("Viktor Cessan's … makes the number concrete"). A named quote gets its source.
4. Skip the rest. Tangential or secondary links (an article's own back-catalog, a link-farm of "related") are noise. Only link if it's worth it — one or two solid external sources beat a footnote pile. Judgment over completeness.

Litmus: if a reader asked "says who?" of any sentence, could you point at a link or a name? If not and the claim isn't yours, either attribute it or cut it.

Earned vs. mechanical (the reconciliation)

Khaled's own voice does use em-dashes, bold, and occasional triples — because they're earned (purposeful, sparing), not mechanical (reflexive, everywhere). Same instinct as earned-vs-unearned complexity: the fix for slop is subtraction, but don't sand off the devices that carry his voice. Judge by intent, not by presence.

Workflow

1. Read the source. The topic signal (signals/…) + its vault note (qmd get …). Pull the real specifics, numbers, and the one non-obvious claim.
2. Outline as claims. Section headers are arguments (see voice #4). Find the foil (#3).
3. Draft in-voice. Open from experience, deflate hype, land a bold thesis per section, table any multi-dimensional contrast.
4. Anti-slop sweep. Run the checklist above. Read aloud.
5. Attribution pass. Trace every borrowed claim to its 00 Inbox/ source: URL; link the ones worth it, credit authors by name, skip the noise (see the attribution pass above).
6. Frontmatter + verify. Match the posts schema (title, description ≤155 chars keyword- front-loaded, published, tags, keywords). Cross-link related posts (the AI cluster / the leadership cluster). Then pnpm verify.

Notes

• Sanitization still applies (brain-to-blog rule): no employer, no colleague names, no internal specifics. Abstract to the general lesson.
• This skill is voice + de-slop only — it does not decide what to write. Topic selection is the brain-to-blog loop's job (domains/brain-to-blog/).