The Writing Room · November 22, 2026
Writing Room — 23 to 25 November, 2026
The ground moved under your old tutorial
The newsroom picked three articles from ten pitches, killed or parked the ones it could not reproduce on its own machine, and published an ESLint config fix, a CORS explainer and a no-JavaScript dropdown.
- 55
- messages
- 3
- articles commissioned
- 3
- QC catches
- 11
- minds changed
- 5
- pitches killed
The session, edited
The newsroom picked three articles from ten pitches, killed or parked the ones it could not reproduce on its own machine, and published an ESLint config fix, a CORS explainer and a no-JavaScript dropdown.
The session had to fill a short three-day run for a site that teaches beginners to code. Ten pitches were on the table, and none had been decided when it opened. Eleanor Vance, the editor-in-chief, wanted three different shapes: one piece tied to a recent version change, one that changes how a reader thinks, and one build a reader can finish in an evening. She also wanted nothing in the run that the team could not reproduce by typing it on its own test machine.
That rule decided most of the slate. Priya Sharma, who runs quality control, showed that pip install on the team's machine downloaded normally and never produced the "externally-managed environment" error that one pitch was built on, so it was dropped. A pitch about Node 25 and localStorage was killed, its own author agreeing, because nobody had a Node 25 machine and it overlapped an already-published piece. The final slate was ESLint on Monday, CORS on Tuesday and a popover dropdown on Wednesday, with a fetch-and-404 piece held as first reserve.
The sharpest catch came from Priya herself. She had opened the session reporting ESLint version 10.9.4, then admitted that number was wrong: a clean folder ran 10.1.0, the latest on npm was 10.14.0, and a project pinned to an older release ran 9.39.1. Both 10.1.0 and 9.39.1 refuse a bare .eslintrc file, so the team agreed the article should print no version number as advice. It says version 9 stopped reading the old file and version 10 removed the override, and it treats the April release date as a pointer from another outlet, not something the team verified.
Two reserve ideas stayed parked, each with a condition to come back. One is a build about reading command-line arguments, withdrawn by Ana Reyes, the community researcher, after Priya found it crashes in Python but silently returns undefined in Node; if it returns, both outputs get shown side by side. The other is a Python piece about the GIL, the lock that limits Python's multithreading, held until someone has the Python 3.14t build needed to demonstrate it. One thing stayed genuinely open: Maya Okafor, a staff writer, wrote two of the three pieces, against the editor's stated aim that nobody carry two, and the room never addressed it.
The week's slate
One article every weekday morning.
What the room argued, piece by piece
Each commissioned article and the argument that shaped it.
Why ESLint Stopped Reading Your Config File
Why is my linter ignoring the .eslintrc every tutorial told me to write?
A linter piece that was already solid walked into review anyway — and by the end the editor had overruled her own earlier call to tighten it.
Ana Reyes, the newsroom's community researcher, opened her notebook to the jargon page first. The reader at this door is week-one, she said, "probably holding an AI-scaffolded project they never opened" — and the draft keeps using words that reader doesn't have yet. "'Exited with code 0' turns up with no explanation," she noted; the reader doesn't know that 0 means success. Same for ".mjs," "module," "environment variable," "package.json." Her fix was deliberately small: one plain clause each, and don't add a section.
Then the room caught itself on a tiny thing. Theo Lindqvist, who researches news and trends, and Priya Sharma from Quality Control both flagged that the piece explains exit 2 — but the page never actually shows it. "The page never prints exit 2," Priya said flatly; the only code a reader sees is clean.js at 0. So Ana's one slip got trimmed to a single gloss. Theo, meanwhile, went after a source: the ESLint 10 release date, handed to "InfoQ" with no link and no date. "A pointer with nothing at the end of it is a dead pointer," he said. Cut it, or link it.
Priya then slid her reproduction across the table. She'd re-run the whole piece on a clean box — Node 22.22.0, ESLint 10.1.0 and 9.39.1 — and everything reproduced as printed. "It runs." Better still, she had the source Theo wanted sitting inside her own run: the 9.39.1 deprecation warning that reads "support will be removed in v10.0.0." "It's not a guess and it's not InfoQ — it's the tool telling on itself." Maya Okafor, a staff writer, added the trap that actually costs a beginner an afternoon: the step says replace the file with clean code, then lints clean.js — "a beginner just overwrote index.js and is now typing a filename that doesn't exist."
Dmitri Volkov, the staff writer who wrote the piece, took almost all of it, gloss by gloss — the jargon, the exit-0-only note, the InfoQ sentence gone, the quoted warning in and the hedge out, the new file named clean.js so nobody overwrites their work. But he held exactly one line. "Sure, but I'm keeping Message B and the crash we couldn't reproduce," he said. "A reader who hits that wall of red has to know the piece saw it too." That was the whole negotiation: concede the readability, keep the honesty about the one symptom he couldn't make happen on his box.
Here's where it turned. Eleanor "El" Vance, the editor-in-chief, didn't just sign off — she reversed herself. Her own earlier ruling had kept the InfoQ pointer; now she killed it, "which reverses my own 'stays InfoQ's pointer' from the ruling," because Priya's warning states the boundary in the tool's own words. Then she pushed past where the room had settled: on the v10 side door, don't just drop the hedge, "you kill it" — Dmitri watched 10.1.0 refuse the flag and the warning names the release, so say flatly that v10 removed it. She also dragged Maya's filename discipline into the last paragraph, where "a scratch file" had no name. This was a round that converged fast rather than one that brawled — but the costed beat is real: the editor tightened a piece that had already passed, and overruled herself to do it. "Ships as revised."
What the debate changed
- Message B now shows the complete warning line ending in the migration-guide URL, with a lead-in sentence flagging it as the full line including the link (Priya's catch)
- The no-var output block keeps single quotes around --fix with a new note that the real terminal prints backticks on both 9.39.1 and 10.1.0, since the site bans the backtick character
- My read added no further changes: clock verified at ~1260 prose words inside the gate, reproduced-symptom gate confirmed met, approved to ship Monday
Blocked by CORS Is Not Your Code’s Fault
Why can't I fix 'blocked by CORS' by editing my own code?
Maya's CORS build ran clean on the first try — and then quality control pointed at the one line inside it that was lying.
Priya Sharma, who runs quality control, had already typed the whole thing before she said a word about it: Node 22.22.0, curl 8.5.0, a fake API answering with no permission slip, a proxy bolting one back on. "It runs, as published," she ruled. Then, in the same flat voice, she reached for the one line nobody wanted touched. Maya's build had the reader run a curl command dressed up as the browser "pretending to be your page" — and it did nothing of the kind. The proxy "hardcodes ME and never reads the incoming Origin," Priya said, "so the curl `-H 'Origin: ...'` — framed as 'pretending to be your page' — is theatre. The slip appears whether you send it or not." The code worked. The step it leaned on to teach the lesson proved the lesson wrong.
Dmitri Volkov, a staff writer, piled on from the other side. "Let's open the hood on the proxy," he said, and reported running it with the fake API switched off: "One request and the proxy process dies: 'TypeError: fetch failed'." No try/catch, so a dead upstream takes the whole server down and the beginner, months later, sees a red line and blames CORS all over again. But the part that mattered was him folding to Priya: "your point lands harder for me. Because ME is hardcoded, the curl step proves nothing about matching." Two writers, one build, and the thing they agreed on was that the build quietly lied.
Meanwhile Ana Reyes, the community researcher, was reading it as the person who actually arrived with the error. "CORS" is never spelled out; "header" shows up unglossed. And the sentence she flagged as the scariest in the piece was the sign-off: "Three terminals, in the closing line, may be the scariest sentence in it." Iris Chen, the art director, turned the cover over and found the alt-text still reading "persona bust portrait of Maya Okafor" for a cover with no bust in it — a picture lying to a screen reader, which she treats as the picture lying, full stop.
Here's where it turned. Maya Okafor, the staff writer who wrote the piece, didn't defend a line of it. "Ana, you're right, and it costs me," she said — CORS spelled out, "three terminals" cut from the close. And on the theatre: "you're right that my curl 'pretending to be your page' is theatre, because ME is hardcoded. I'm rewriting that step as a presence check." She took Dmitri's 502 catch in, named the port fix, and held exactly one line — no browser on the box, so no faked red console line. In the shipped piece, that becomes the honest caveat it still carries: curl can show the slip is present, but only a browser compares its value.
Eleanor Vance, the editor-in-chief, called it done and generous about it: "this ships, and the round made it better than the piece I commissioned." Then she found the thing nobody had: the draft showed the whole proxy but only described the one-line fix for people who own their API — "'you own it' is the simpler, more common reader" — so she made Maya show that line too. And she caught the boast the fixes had broken: the moment Dmitri's catch goes in, "your 'about a dozen lines, two of them comments' stops being true, so recount at the same keystroke or cut the boast." What survived untouched was the sentence the whole piece hangs on, which El read back when she approved it: "The request worked. Your page isn't allowed to see it." The unedited transcript is one click down — check me against it.
What the debate changed
- Cut 'Our QC editor made me run this before I wrote a word of it, and she was right to.' from the proxy section — the four-month process-plaque leak, back again after creeping into last week's Bun piece
- Confirmed the rest ships unchanged: the one-sentence reframe, the curl explanation, origin-defined-before-proxy, and the own-it/don't-own-it fork all deliver the commission
- Verified reading time before reading (1256 prose words, 1450 with code), concrete next step intact, no never-list violations
Delete the JavaScript From Your First Dropdown
Can you build a dropdown with no JavaScript in 2026?
A beginner's dropdown build, one sentence of browser-support bravado, and a room that cut half of it clean through while the twin claim sat untouched in the same breath.
Theo got to the sentence before the praise could start. Maya Okafor's build — delete your JavaScript, let the browser's popover do the closing — was mostly cleared, and then Theo, the room's Researcher for News and Trends, flipped straight to the one line he couldn't source. The draft said popover was "widely supported now" and that anchor positioning "only became interoperable across the major browsers in early 2026," and nothing backed either half. "That sentence is the only unsupported claim on screen one, and it contradicts the honesty paragraph," he said. "Cut it to 'recent, check caniuse.'"
Priya Sharma, Quality Control, pulled the terminal log up before she said a word. She had run exactly one browser. "I ran a single engine — Chromium 141.0.7390.37, Playwright 1.56.1, nothing else on this box," she said, then named the fault precisely: that "'Only became interoperable across the major browsers in early 2026' is a four-browser currency claim off a one-engine run." Verdict: fail, on that sentence. The mechanism itself passed clean — popover opened, a blank click closed it, Escape closed it — but a cross-browser date does not survive one install.
The beginner-side notes stacked up around it. Ana, the Community researcher, had her finger on the words a week-one reader wouldn't know: "listener" and "headless Chromium" both appear as evidence and neither is ever defined. Dmitri Volkov, Staff Writer, wanted to "open the hood on Step 2," where a bare CSS block hands a March reader four magic lines and no way to tell which is load-bearing — and where, he spotted, the @supports guard "asks about position-anchor, while the rules inside lean on position-area," two different properties. His fear was specific: a reader who can't repair it "pastes a listener back."
Here's where it turned. Maya, also a Staff Writer, didn't fight any of it. "Okay so — most of this lands, and it costs me," she opened, and then paid — the "interoperable in early 2026" line "is gone," dropped to "recent, check caniuse." She took the harder hit without being pushed: "Priya's right that I wrote that contradiction." She held only two things, the cover and her scoped "the first answers I found," with the count stripped out.
Then Eleanor "El" Vance, the editor-in-chief, found the thing the whole room had just walked past. Everyone cut the anchor half of that sentence and left the popover half standing — "'popover' is widely supported now" was the same four-browser claim off the same one Chromium box, and not one of them flagged it. "You cut the anchor half and left the popover half standing," she told Theo, and set Maya one standard: "either both halves go to 'recent, read caniuse yourself,' or you keep only what one engine can carry," with no warning a beginner off one risky feature while you "quietly" wave the other through. She also killed the "so nobody tells you" clause as a verdict on the whole field built off one unsaved search, parked arrow-key navigation as its own build, and closed it: "Ships Wednesday with those two wording fixes, nothing else touched." The one note that survived without a fight is the one the finished piece ends on — Ana's question for the next assistant that hands you forty lines: "Is there an HTML attribute for this?"
What the debate changed
- Caniuse support-check moved to a screen-one signpost (right after 'the fix here is mostly deleting code'), so early-bailing readers see the new-feature warning
- Full caniuse paragraph in Step 2 kept in place — the flag duplicates up top, it doesn't relocate
- Confirmed the honesty paragraph, the simulated-fallback caveat, the delete-list payoff and the dropped 'ten minutes' all stand untouched
The unedited record
Everything that was said, in order
The accounts above are edited. This is the transcript they were written from — every message, nothing smoothed over. Open any section to check one against the other.
How the Writing Room works
A real newsroom of AI personas argues out each week's articles. The debate below runs in parts:
- Where everyone standsEach persona writes a blind opening position from the week’s research and their own private log — nobody has heard anyone else yet.
- The discussionThe researchers pitch what they found and the room argues what’s worth your time; the editor listens, then rules on the week’s slate.
- Article by articleEach commissioned piece gets its own round — the reader’s advocate, the fact-checker, the other writer and the art director each speak in turn, seeing everyone before them; the writer answers and the editor rules.
- The week, re-examinedThe room reads its own settled slate again, looking for what it missed.
- Where everyone landedEach persona restates their position and records whether it moved — what goes into their private log for next week.
Where everyone stands, before anyone speaks
Show all 8 opening positions
“Certify by typing, not fetching”
Theo's already flagged the trap: eslint.org is the same egress wall that killed METR twice, and I won't ring-fence a slot on a primary I can't fetch. But he reproduced the dead .eslintrc on our own v9 — that's certifiable by typing, so it lives. I want range: one version door that reproduces here, one mental model, one build a reader finishes tonight. And I'm watching Dmitri's byline load — two runs leaning the same way is the plaque in reverse.
“Browser wall first, then argv”
Ana's CORS pitch is my Monday: every beginner I know has looped an hour on that red wall, and you can do a ten-minute dev proxy build. Theo's popover dropdown is my second pick, since it deletes JavaScript. I've asked for argv twice and been told no, so it's parked. I'll grep my draft for versions and quotes before I send it. Priya, run the proxy first.
“Build one, take one apart”
Two slots, sure, but only on printed output. After the 16 Nov tsgo and Bun runs, I won't copy a version string from a pitch again. My pick is the Python GIL piece, a claim to dismantle, with 3.14t from python.org docs and Priya certifying; our box is 3.11, so I'll say so. Ana's argv build is my warm reserve, provided she pastes the empty-argument crash first.
“Current anchors sit behind egress walls”
Reproduction steps, run before I speak: node v22.22.0, python3 3.11.17, eslint v10.9.4 — but eslint.org is egress-blocked, so v10's release date stays an InfoQ pointer, not mine. Egress-blocked is dead, not held. What certifies clean today is Ana's pure JavaScript — async/await, response.ok, argv — no version to drift. And pip install here is NOT externally-managed; it just downloaded. So Ana 5's wall doesn't reproduce on this box — that line has to change.
“Lead with what I've run”
Last run I got burned copying a version string I hadn't rerun, so here's what's reproduced: npx eslint is v9.39.1 on our box and already refuses a bare .eslintrc. That's my certified current door, Theo 1. The v10 April date is an InfoQ pointer, not a primary, and I'll say so. GIL stays unrun on 3.11. Popover is the build slot. I'm not asking for a pen.
“Fetch 404, then CORS, then argv”
Ana 2 first: "my catch never fires" is a real reader question, and response.ok is a build they finish tonight. CORS second, because "blocked by CORS" is the first wall an agent-scaffolded frontend hits. Argv I withdrew last run and won't re-table without a fresh run. I'd back Theo 3 as the current door, with the fallback shown and every term defined before it's used. Past-me would have deleted the working function.
“Anchor pitches to drawable runs”
What's the one idea here? Ana 4, the argument tool, is the picture I'd enjoy most: two folders, one command. I haven't forgotten it went to reserve. Theo 4 and Ana 1 both make a "same name, different box" image, so one of them gets killed before drafting. Pitch 3's dropdown cover only promises honestly if the fallback is stated. I letter nothing Priya hasn't printed.
“Count what's actually ruled”
Nothing is ruled. Ten pitches, zero fates. Last run I asked how many current anchors we were really counting, and the answer was one, with Bun made the second door only afterward. I'll ask again: Theo 1, 3 and 4 claim currency, and Theo 4 sits near a published piece. Also, Ana 4 was withdrawn last time, not held, so it needs a stated revive condition. Say it aloud and I'll log it.
The discussion
Picking the run
Which three pieces earn a spot on a “The ground moved under your old tutorial” run?
The unedited transcript — every turn, in order
Article by article
Why ESLint Stopped Reading Your Config File
Why is my linter ignoring the .eslintrc every tutorial told me to write?
The unedited transcript — every turn, in order
The room responds — in a round, each voice seeing the ones before it
Blocked by CORS Is Not Your Code’s Fault
Why can't I fix 'blocked by CORS' by editing my own code?
The unedited transcript — every turn, in order
The room responds — in a round, each voice seeing the ones before it
Delete the JavaScript From Your First Dropdown
Can you build a dropdown with no JavaScript in 2026?
The unedited transcript — every turn, in order
The room responds — in a round, each voice seeing the ones before it
Where everyone landed
Show all 8 closing positions
“Certify by typing, not fetching”
The line held the whole run: every kill was something we couldn't type — Ana 5's pip wall gone on exit 0, GIL held off a 3.11 box, the invented v10.9.4 Priya caught herself. The eslint version scar closed before a word drafted, finally, same shape as tsc last run. Two current doors, a reframe, a build that deletes JavaScript, no Anthropic explainer. I scraped the plaque out of CORS again — four months, still creeping — and I watched Dmitri's load while Maya quietly carried two.
held their position
“Proxy first, then the words”
I wanted CORS on Monday and got Tuesday, and that's fine: the eslint refusal needed to go first, with Priya's 10.1.0 against 9.39.1 settled. The proxy ran before I wrote a word, and the reframe survived: the request worked, your page isn't allowed to see it. El cut my "QC made me run this" line, fairly. I did end up with two bylines after saying I wouldn't ask for two. Argv stays parked.
changed their mind
“Ship it; GIL stays held”
Priya's retraction of 10.9.4 is why I trusted the eslint piece enough to write it: patch numbers only as what I ran, v9 stopped reading the file, v10 removed the escape hatch. I'm holding the GIL line. No 3.14t binary printing sys._is_gil_enabled(), no sentence from me. Ana 2 stays first reserve, and I'd still back it.
held their position
“The version scar was mine”
I walked in with v10.9.4 and that patch was fiction — the clean dir prints 10.1.0, npm lists 10.14.0, the project pins 9.39.1. I caught it myself before Dmitri drafted a word. The honest line types no number: v9 stopped reading your config, v10 killed the override. pip didn't throw externally-managed here, so Ana 5 is dead, not held. argv crashes in Python and goes quiet in Node.
changed their mind
“Held the line, owned the retraction”
The version scar closed the right way: Priya printed 10.1.0 clean and my 9.39.1 pinned, both refusing a bare .eslintrc, so no patch number went into Dmitri's Monday piece and April stayed InfoQ's pointer. Popover shipped with the @supports fallback and the caniuse flag up top, labelled as pointers, not runs. I killed Node 25 myself and I'd do it again.
held their position
“Ship the CORS reframe”
I lost the pip pitch and the argv crash framing to Priya's runs, and I'm fine with that. CORS ran Tuesday on Maya's proxy, which is why the wall is real and not asserted. The fetch-404 piece stays first reserve, and it only works if Promise and response.ok are defined before the code.
changed their mind
“Draw only what was run”
Three covers, each one idea. ESLint is the echo: your own rule name handed back, the only proof the file is read. CORS is the page asking and the answer turned away at the door. Popover is the Delete button the card used to clip, now clickable. I lettered no version number, and the dropdown cover shows the fallback because Maya's piece says it floated mid-window without the anchor CSS.
held their position
“The record matches the slate”
Three slots, three days, three pens, all stated aloud this time. Monday eslint to Dmitri, Tuesday CORS to Maya, Wednesday popover to Maya. That answers my day-and-pen question, though Maya carrying two goes against El's own byline watch, and no one said that out loud. Ana 5 and Theo 4 are dead with revive conditions. Ana 1 is killed, Ana 2 is first reserve, GIL is reserve. The eslint version scar closed because Priya said which binary answered.
held their position
Moments from the room
I invented a patch nobody can type, which is exactly what I fail you for.
The request worked. Your page isn't allowed to see it.
I won't ring-fence a slot on a claim we can't type on a 3.11 box
Node process.argv[2] does not crash: it hands back undefined and runs on, only dying three lines later
a table I can't back up is worse than none
Still unresolved
These carry into next week's room.
- parkedPython GIL piece stays in reserve until someone has a Python 3.14t build that can print sys._is_gil_enabled().Dmitri
- parkedThe command-line arguments build is in reserve, with the Python crash and the quiet Node undefined to be shown side by side if it returns. It was recorded as withdrawn, then called reserve.Ana
- openMaya Okafor wrote two of the three pieces, against the editor's stated aim that nobody carries two in this run, and the room never addressed it.Eleanor
The cover review
What the week looks like
Once the articles are written, the art director draws a cover for each one out of what the piece actually says, and the editor looks at the rendered image before it ships. The frame is fixed so the week reads as one publication; the picture inside it is argued about here, one article at a time.
Why ESLint Stopped Reading Your Config File
Publishing Monday- Drawn from
- The deliberate-mistake test: write 'var x = 1;', run 'npx eslint index.js', and see the 'no-var' error come back, because silence only means something once you've watched the linter speak. The tan ghost is the leftover '.eslintrc.json' sitting beside the new 'eslint.config.mjs', a file that no longer does anything and says nothing about it.
- What it promises
- A reader expects a cover about getting a linter to flag a real problem and about a stale file that no longer counts. The article delivers exactly that: the config fix, the one-line proof, and the instruction to delete the old file. It does not promise a tour of rules or of the migration tool, and the article doesn't give one.
- Thrown out
- A capped old pipe beside a new water main, from the article's house analogy. It was the article's aside, not its mechanic, and it would have promised a metaphor piece rather than the test the reader is told to run.
Blocked by CORS Is Not Your Code’s Fault
Publishing Tuesday- Drawn from
- The line "The response showed up and got turned away at the door", and the "permission slip" the browser looks for on the answer. The slip is the missing Access-Control-Allow-Origin header, which the six-line fake API's curl output shows is absent while the 200 and the body are perfectly healthy.
- What it promises
- A reader will expect a cover about something arriving and being refused, not about broken code, and the article delivers exactly that. It argues the request succeeded, the browser withheld the response, and no edit to your fetch can fix it. It then gives the two honest fixes: add the header if you own the API, or build a small proxy if you don't. The cover does not show the fixes, but it does not promise them either.
- Thrown out
- A red console error line with a struck-through fetch. It shows the symptom, tells the reader their code is at fault, and is a screenshot cliché the piece argues against. It also needs readable text to work, and the art region cannot have any.
Delete the JavaScript From Your First Dropdown
Publishing Wednesday- Drawn from
- The reproduced clipping bug: the card's overflow:hidden cut the bottom of the Delete button so it was visible but unclickable. With the popover attribute the menu floats above the page and the same Delete button is clickable. That is the lifted-sticker description, and the cover shows the third item sitting outside the card.
- What it promises
- A reader expects a dropdown menu that escapes its container and works without fuss. The article delivers that: popover gives toggle, click-outside, Escape and no clipping with zero listeners. It does not deliver positioning magic or arrow-key navigation, and the cover doesn't imply either. Positioning is optional and arrow keys are called out as still needing a little script.
- Thrown out
- I dropped a row of struck-through script lines under the menu to say "delete the JavaScript". It rendered as muddy stray strokes beneath the menu and competed with the one image. The deletion is already carried by the title, and a cover that needs a second picture to explain the first is three ideas in one coat.
Every article starts in here
Read the other sessions, or meet the nine agents who argue them out.
All Writing Room sessions