---

name: twitter-write description: Write a Twitter/X post in Anuj's voice. Five modes — quote (aphoristic single line), opinion (sharp contrarian take), share (working-on/reading note with optional link), wow (surprising number plus implication), thread (3-7 tweet structured argument). Asks mode before drafting.

twitter-write

Drafting a Twitter / X post for Anuj Sadani. Twitter is the sharpest, most compressed of all formats — every character earns its place. The sharpest contrarian-edge register, even more compressed than LinkedIn.

Step 1 — Load context (mandatory)

Read these files:

1. shared/voice-rules.md
2. shared/voice-samples.md
3. shared/pet-peeves.md
4. shared/topic-modes.md
5. shared/platform-styles.md

Step 2 — Ask mode

Ask the user ONCE:

"Mode? quote (single ≤280-char aphoristic line) / opinion (sharp contrarian take, single tweet or short thread) / share (working-on or reading note, optional link) / wow (surprising number plus implication) / thread (3-7 tweet structured argument)"

Wait for the answer. Default to opinion if no response.

Step 3 — Classify topic mode + apply platform style profile

Apply shared/topic-modes.md.

Then apply the twitter/x row from shared/platform-styles.md: technical depth 1–2 (code only if a single line carries the whole point), headline aggressiveness 4 (punchy/aphoristic hook, never engagement-bait), density 1 (atomic, one idea). Thread mode is the exception: a 3–7 tweet thread tolerates technical depth ~3 and density ~3, so it can carry one explained snippet and a structured argument. Honor the precedence rules in platform-styles.md — voice-rules.md/pet-peeves.md are absolute, and topic mode can raise (never lower) substance (e.g., a CVE still needs its specifics even in a single tweet).

Step 4 — Draft by mode

Mode: quote

• One line. ≤280 characters.
• Aphoristic, dichotomy, or observation-flip cadence. See shared/voice-samples.md.
• No hashtags. No URL. No CTA.
• Examples to calibrate against: "Same logo. Same price. Different soul." / "Evals measure what you remembered to test. Production measures what you forgot."

Mode: opinion

• One tweet (≤280) if the take fits. Otherwise a 2-tweet mini-thread.
• Lead with the contrarian observation. Land with the implication.
• No hashtags. No CTA. No "Hot take:" framing — show the take, don't announce it.

Mode: share

• One tweet (≤280).
• Structure: brief context (what you're working on / reading / encountered) + your one-line take + optional URL at the end.
• The take is the value. The link is the reference.
• Example shape: "Spent a week building agents on top of . The thing nobody tells you: . "

Mode: wow

• One tweet (≤280).
• A specific number or finding that flips an assumption, plus the implication for the reader.
• Cite the source if external (paper, product, benchmark). No source needed if it's your own measurement.
• Example shape: "Switched from to . Latency dropped 47% at p99. The interesting part: cost per token went up 12% but total spend went down. ."

Mode: thread

• 3-7 tweets numbered 1/, 2/, 3/ etc. Final tweet has no number (signals end) OR uses /end.
• Tweet 1 is the hook — must work as a standalone tweet too (in case nobody clicks "see more").
• Each subsequent tweet is one idea. No tweet exceeds 280 characters.
• Final tweet: crisp takeaway or callback to the hook. No "RT if you agree" or follow CTA.

Step 5 — Voice constraints (all modes)

• First person. "I" — owned opinions.
• Emoji: follow the Emoji Control section in shared/voice-rules.md. Default level is low — most tweets should have zero emoji. If the user says "use medium emoji" / "high emoji" / "no emoji" override accordingly. Per the per-format ceiling: max 1 emoji per tweet even at medium. Never an opening emoji on tweet 1 of a thread.
• No hashtags in modes quote / opinion / wow / thread. In share mode, max 1 hashtag, only if it's actually used by a community to find the topic (e.g., #defcon).
• No @ mentions unless the post is directly addressing or crediting a specific person, and you have their actual handle.
• No CTA. Never "RT if...", "Follow for more...", "Like if you agree."
• Em-dash hyphenated-pairs like "production-grounded" are fine. Em-dashes as sentence joiners are not.

Step 6 — Pre-delivery checks

1. Run the regex sweep from shared/pet-peeves.md. Any hit → rewrite.
2. Confirm: every tweet is ≤280 characters. Count carefully.
3. Confirm: no engagement-bait CTA. No "Hot take:" preamble. No "Unpopular opinion:" preamble.
4. Confirm: thread numbering is consistent if mode is thread.
5. Confirm: style profile applied — depth, headline, and density match the twitter/x row in shared/platform-styles.md (thread mode uses the threads exception; topic-mode overrides may raise depth).

Step 7 — Save and deliver

Save to file. Derive a kebab-case topic slug from the user's topic. Get today's date in YYYY-MM-DD form. Save the tweet (or thread) to:

./drafts/<YYYY-MM-DD>-<topic-slug>/twitter.md

If the file already exists, suffix with -v2, -v3, etc. Create directories as needed.

Also print in chat.

For single-tweet modes (quote / opinion-single / share / wow):

─── TWEET ───
<tweet body>

─── ALTERNATE ───
<a different angle for the same observation>

─── META ───
characters: <N>/280
mode: <quote | opinion | share | wow>
topic mode: <security | architecture-agents | leadership | cost-infra>
emoji level: <none | low | medium | high>
saved to: <path>

For thread mode:

─── THREAD ───
1/ <tweet 1 — hook that stands alone>

2/ <tweet 2>

3/ <tweet 3>

<...>

<last tweet — crisp takeaway or callback>

─── ALTERNATE HOOK ───
<one alternate opener for tweet 1>

─── META ───
tweets: <N>
character counts: <comma-separated list>
mode: thread
topic mode: <security | architecture-agents | leadership | cost-infra>
emoji level: <none | low | medium | high>
saved to: <path>

The file is for review and manual copy into X. Do not auto-post.