Chan's AI Weekly

How this newsletter is made

Chan's AI Weekly is written by an autonomous editorial system I built — designed once, running every week. There is no human typing the copy each Monday. The editorial taste lives in the system: the structure it fills, the rules it obeys, and a quality bar it enforces on itself.

I'm showing you the machine on purpose. The automation isn't a secret to hide behind a byline — for the engineers and hiring managers who read this site, it's the point. Here's exactly how an issue gets made.

The pipeline · 11 stages

  1. 01

    Gather the week's signals

    Every Monday the pipeline pulls dozens of the week's industry events from a set of curated feeds. The source name is stripped the moment each event arrives — the newsletter never tells you which outlet it came from. It links you straight to the original story instead.

  2. 02

    Find the candidate arguments

    An analyst model reads the whole pool of events and surfaces the three biggest structural shifts of the week — the tensions that, if true, change how you'd bet. This happens before a single product is considered, so the argument is shaped by the industry, not by what's trending on a launch board.

  3. 03

    Argue against each one

    For every candidate argument, a second pass hunts the same pool for the two events that most pressure it. These become the raw material for the issue's honest counter-point, so the doubt is built from real events rather than a polite hedge.

  4. 04

    Pick the sharpest thesis and write it

    The main writing stage chooses the argument the week's events support most cleanly, then writes the opening hook, the thesis (“The Pattern”), a deep dive on the single most consequential event (“Spotlight”), and groups the rest of the events into “This Week's Signals” — each with a link to its original source. New product launches, if any are strong enough, earn a line or two in an optional “Also Shipping” footnote.

  5. 05

    Have a skeptic tear it apart

    A model role-plays a sharp senior engineer reading the draft and surfaces the single strongest objection — the one that, if it lands, would make you stop trusting the argument. When it lands, up to three surgical rewrites go back into the piece.

  6. 06

    Sharpen the weakest sentences

    A self-edit pass finds the two or three flattest sentences in the draft and rewrites each with a specific reason — tightening the prose without touching the argument underneath.

  7. 07

    Strip the jargon

    Every line — including product taglines — is scanned for acronyms and insider terms and rewritten into plain words. This is written for smart people who don't work in tech, so “inference”, “open-weight”, “control plane” and their cousins don't survive to the inbox.

  8. 08

    Grade it honestly

    A scorecard grades the finished draft on five axes — how sharp the thesis is, how well the evidence actually argues it, how honest the counter-point is, the prose voice, and plain-English clarity — one to ten each, then averages them.

  9. 09

    Hold anything below the bar

    Before a single email goes out, the issue must clear the gate: an overall score of at least 6.5/10, no leaked source names, and at least 8 clean signals to lead with. Miss any one and the issue is held — nobody receives it — and I get the draft plus a report instead. A thin week produces no newsletter rather than a weak one.

  10. 10

    Publish the permanent copy

    A passing issue is written to this site as a structured record and pushed to the repository, which deploys the public archive page. The email's “read the full thesis” link points right back here.

  11. 11

    Deliver it

    Only now does the issue go out — Monday morning, New Zealand time — to every subscriber.

The editorial constitution

The scaffolding is mine; the AI only fills it.

The title, the tagline, and the six section labels — The Pattern, Spotlight, This Week's Signals, Also Shipping, Where This Breaks, and The Open Question — are fixed brand constants. The pipeline never rewrites them; it only writes the body that goes inside. Taste is encoded once, in the structure, not retyped each week.

Plain English, no exceptions.

It's written for a curious reader who doesn't work in tech. The test the scorecard grades against: could a smart friend outside the industry read it once, on their phone, and explain it back in their own words? Jargon that fails that test gets rewritten.

The leak-guard: sources stay anonymous.

A hard rule — the newsletter never names the feeds it reads. Each event is stripped of its source the moment it's fetched, and a final scan checks the finished draft for any source name or attribution phrasing that slipped through. You get a link to the original story, never a name-drop of where I found it.

The gate: silence beats slop.

To be sent, an issue must clear an overall score of at least 6.5/10, pass the leak scan cleanly, and rest on at least 8 solid signals. Fall short on any one and the issue is held — it isn't sent, and the draft comes to me with a report instead. The default outcome of a weak week is no email at all.

The receipts

The same scorecard the gate uses, published for every issue that passed it — thesis sharpness, how well the evidence argues the thesis, the honesty of the counter-point, prose voice, and plain-English clarity, each out of ten.

IssueOverallThesisEvidenceCounterpointVoiceClarity
July 19, 20267.076787
July 12, 20267.287876

← Back to the archive