Pinned Tweet
Most "bad AI output" isn't a model problem. It's missing guardrails.
Before I trust any LLM with real spreadsheets, docs, or code, I run 5 quick stress-tests:
1. Count In vs. Count Out
"Report exact input count vs. output count. Never silently drop or merge items."
2. Assumptions First
"List every assumption you're making and pause if 2+ valid interpretations exist."
3. Format Lock
"Keep leading zeros (00123), exact IDs, and YYYY-MM-DD dates as raw strings."
4. Anti-Yes-Man Pass
"Give me the 3 strongest reasons this approach breaks in the real world."
5. Diff-Only Output
"Show only what changed and why. Do not rewrite untouched sections."
Saves hours of silent bugs every single week.
What's one guardrail rule you put in almost every prompt?
The fastest way to catch an LLM skimming your long PDF or transcript:
Ask for exact quotes + section headers BEFORE asking for a summary.
My two-step prompt:
1. "Extract every number, date, and named constraint verbatim with its page/section."
2. "Now summarize ONLY using the extracted facts above. Flag anything missing."
Accuracy jumps immediately because the model can't hallucinate bridges between paragraphs it skipped.
How do you verify long-document summaries without re-reading the whole file?
2
2
31
Telemetry across 10,000+ developers exposed the real AI coding bottleneck:
Merged PRs jumped +98%. Human review time spiked +91%.
Writing code isn't the bottleneck anymore — reading AI diffs is.
3 PR rules that keep review queues from collapsing:
• Hard cap PR size (~250 lines changed, no drive-by refactors)
• Require an AI-generated "blast radius" note listing touched state, auth, and DB paths
• Make the agent write the failing regression test BEFORE generating the fix
Fast generation without fast verification is just technical debt on leverage.
Where is AI slowing your workflow down instead of speeding it up?
2
1
45
Everyone watching Microsoft Ignite is talking about MAI Code 2 hitting ~98% frontier coding performance at ~40% of the cost.
The bigger workflow shift isn't one cheaper coder model — it's specialized routing:
• Tiny router model (like Decision 1.1) handles triage & classification in milliseconds
• Fast workhorse tier (MAI Code 2 Flash / GPT-6.2 Sol) handles 80% of multi-file edits at half price
• Heavy frontier model only touches gnarly architecture & security edge cases
Using your most expensive model on every prompt is like hiring a staff engineer to rename CSS variables.
Are you already routing tasks across cheap + frontier models, or still running one default model for everything?
1
74
Logitech now gives staff a day off to play games. GTA VI arrives Nov 19.
The first GTA VI mission isn’t in Vice City. It’s getting your leave request approved.
HR is the final boss. #GTA6
news.logitech.com/press-rele…
1
7
401
Two break points. Four serves. No rally.
Serving at 5–4, 15–40 against Jodar, Zverev went: ace → unreturned serve → ace → unreturned serve. First set secured.
He landed 88% of first serves in that set and won 6–4, 6–4. That tiny sequence tells me more than the score: Jodar found an opening, then barely got to play.
What stood out to you in the ATP Finals opener?
If you enjoy match details beyond the score, give me a follow. 🎾
#ATPFinals
nittoatpfinals.com/en/news/z…
Before letting AI clean a spreadsheet, give it four traps:
• Product ID 00123 — must stay text.
• Date 03/04/2026 — ask which date format, don’t guess.
• Two rows with the same customer — not necessarily duplicate orders.
• Blank price — missing, not $0.
Keep the original file. Ask for a change log and a separate “needs review” sheet.
A neat-looking table can still be wrong.
PSA played and authenticated a copy of 2005’s event-only PokéPark Fishing Rally. It also authenticated a separate cartridge for transferring scores to a kiosk. Preserving the kiosk setup would show how people actually played.
psacard.com/articles/article…
Rafael Jodar made his tour-level debut in January. He opens his first ATP Finals against two-time champion Alexander Zverev, who won their only meeting. I’d watch how often Jodar gets a look at a second serve early.
nittoatpfinals.com/en/news/z…
IVE’s “BANG BANG” won Song of the Year and Best Music Video at the Melon Music Awards. Four years after “LOVE DIVE,” IVE have a second Song of the Year. Which of the two would you play first for a new fan? #MMA2026
melon.com/mma/result.htm?mma…
FURIA swept Spirit 3–0 to win back-to-back Hong Kong titles. For the replay, I’d start with Mirage. It was Spirit’s pick, and FURIA closed it 13–8 to win the final.
teapawt.net/BLASTPremier/status/21…
Five home runs in four APBC games for Tomoya Masaki. His three-run shot put Japan up 5–0 in the fifth of the final against Korea. If you’re catching up on the game at Tokyo Dome, start with that inning.
npb.jp/scores/2026/1115/jpn-…
BTS won Artist of the Year at the Melon Music Awards. CORTIS won Record of the Year. Artist is 60% digital, 20% judging and 20% fan votes. Record is decided entirely by the panel. #MMA2026
melon.com/en/mma/2026/info
Before your next YouTube Live, preview the center crop. A widescreen layout can lose its labels and guests in the Shorts feed. I’d keep the thing viewers need most in the middle and use the edges for context.
support.google.com/youtube/a…
Pixar pulled the stars closer to the camera in WALL·E. At realistic distances, they’d barely move in the frame. Bringing them closer made the characters’ motion easier to read. What other movie trick hides in plain sight?
pixar.com/wall-e
Try storyboarding inside your 3D scene. Blender’s Grease Pencil lets you sketch arrows for movement and mark eyelines in the shot. Then you can check the composition from the scene’s camera.
docs.blender.org/manual/en/5…
A two-strike pitch with the bases loaded is where I’d want an ABS challenge most. MLB teams start with two and keep the ones they win. The cap tap makes knowing when to trust your own read part of the game.
mlb.com/news/abs-challenge-s…
FURIA took Ancient 13–11 in Hong Kong, then KSCERATO opened Dust II with a clutch. If you’re catching up on the final, start with the two late B rounds on Ancient. YEKINDAR found the opening in Cave both times.
hltv.org/news/45993/live-upd…
The Bun talk at AI Engineer makes a good case for choosing projects you can verify. With a coding agent, I’d start with a bug I can turn into a failing test. That gives me something concrete to review when the agent says it’s fixed.
ai.engineer/code/2026/talks
70

