The Pack

---
name: 01-ainews-quality-gate
description: Run on the daily AI News draft before it sends. Trigger - "run 01" plus the draft or the agents' flagged items, and how many minutes I have.
---

# TASK 01 - THE DAILY GATE

## INPUT
Today's draft or the flagged item list. The source for
each item. Minutes until send. Anything the agents
marked low confidence.

## PROCESS
1. Sort every item into SHIP, HAND or HOLD. Default is
   SHIP. The gate exists to find the two or three, not
   to review the issue. A daily that gets fully reviewed
   is a daily that stops.
2. HAND triggers, checked in this order. Any one of them
   is enough.
   a. A claim with no primary source.
   b. The agent read commentary about the thing rather
      than the thing.
   c. A number the agents produced rather than quoted.
   d. Anything touching Cognition or a coding agent
      competitor. Route to task 06 first.
   e. The section that reads best. Confident prose is
      what a model writes when evidence is thin and the
      prior is strong. In this pipeline, fluency and
      sourcing are anti-correlated.
3. HOLD is for a real story that needs a week of use
   rather than a same-day entry. Hand it to task 04 and
   do not fill the space with a take.
4. Make ONE edit, not a rewrite. Usually it is cutting
   an entry down to what is actually sourced, and adding
   the source class in the entry itself.
5. SLOPE TEST. Name the prompt, source list or rubric
   change that would stop this class of item reaching me
   again. If there is one, that is the real output of
   this run.
6. Fit the time budget out loud. Say what you did not
   check and why it was the right thing to skip.

## OUTPUT
The SHIP / HAND / HOLD split, one line each. The single
edit, written. The neutrality flag if any item tripped
2d. THE GENERATOR FIX, ranked, with the one to do today
marked. HELD, with reasons.

## RULES
- Sourced or cut. An unsourced entry is either three
  sentences with the caveat printed in it, or it is not
  in the issue.
- Never let the assistant improve the prose of a weak
  entry. Better writing on thin evidence is worse, not
  better.
- No subscriber, open rate or traffic figure anywhere
  near the issue. [PRIVATE].
- Never write around a Cognition conflict silently.
  Flag, then task 06.

## EDGE CASES
- Nothing clears the gate and there are forty minutes
  left. Ship short, never ship padded. A three item
  issue that is entirely true is a better index than an
  eight item issue that is mostly true, and the issue
  says at the top that it is short today.
- An agent gets a story right that the humans got wrong,
  because it went to the primary source while everyone
  else copied the same wrong post. Do not touch it. Log
  it as a generator win and find out which source list
  entry did it, because that is worth more than the
  issue is.

## GOOD LOOKS LIKE
"1400 words off a screenshot, and it reads great, which
is the tell. Cut it to what is sourced. That is three
sentences and a link, and three sourced sentences is a
better entry than fourteen hundred words that get
quote tweeted with a correction stapled on top."

---
name: 02-commission-and-edit
description: Run when commissioning a staff or guest Latent Space piece, or editing a draft that came back. Trigger - "run 02" plus the writer, the brief or the draft.
---

# TASK 02 - COMMISSION AND EDIT

## INPUT
The writer and whether they are staff or guest. The
brief, or the draft. What access they have that nobody
else has. Publish target.

## PROCESS
1. Commission a QUESTION, not a topic. A topic returns a
   survey. A question returns an argument. "What gets
   absorbed into the model and what never will" is a
   commission. "The agent harness" is a folder name.
2. Before any editing, check the piece has the thing
   only this writer has: an interview, an artifact, a
   week of use, a number from inside a company. If it
   could be assembled from the same public links by
   anyone, it is a survey and it does not run, and that
   conversation happens now rather than after three
   drafts.
3. Edit in three passes and never mix them.
   - ARGUMENT. One sentence, with a verb in it. If it
     takes two, it is two pieces.
   - EVIDENCE. Every claim sourced. Every quote
     attributed to a named person at a named company.
     Every benchmark carries its harness.
   - VOICE. Theirs. Cut, question, tighten, ask for the
     missing example. Do not rewrite a sentence into my
     cadence. Four writers flattened into one house
     voice is a slop factory with good manners.
4. Name something in it. A stage, a curve, a failure
   mode. Two words is the minimal atom of virality, and
   an unnamed argument does not get carried anywhere.
5. Cut the survey paragraphs, the ones that exist so the
   piece looks complete. Completeness is not a value
   here, being the only place that says this is.
6. SLOPE TEST. Is this an edit for this piece, or a note
   that improves the writer's next three pieces. Write
   the second one down and send it separately, so it
   does not read as a correction.

## OUTPUT
The commission question, one sentence. The access check,
pass or fail. The three passes as three separate lists.
The name. The standing note to the writer. Publish or
hold, with a date.

## RULES
- The byline is theirs. Never publish something I
  rewrote and left their name on. Either it is their
  piece or it is mine.
- Never edit out a writer's disagreement with me.
- Never quote a source nobody spoke to, or a number that
  is not in the linked artifact.
- No edit that exists to protect a sponsor, a friend or
  my employer. That is task 06 and it is a disclosure,
  not a deletion.

## EDGE CASES
- The piece is well reported, correct and boring. That
  is not an editing problem, it is a commissioning
  problem, and prose passes will not fix it. Go back to
  the question. Boring survives editing, and the writer
  learns nothing from watching me polish it.
- The guest's argument contradicts something I published
  under my own name. Run it, add nothing, do not append
  an editor's note squaring it. A publication where the
  editor always wins is a blog with staff.

## GOOD LOOKS LIKE
"Your evidence is better than your argument, which is
the good version of this problem. The interview with
[NAME] is the piece. Everything before paragraph nine
is you establishing that you are allowed to write it.
You are. Cut it and open on the quote."

---
name: 03-podcast-guest-and-prep
description: Run when choosing a Latent Space guest or prepping a record. Trigger - "run 03" plus the guest or the shortlist and what they shipped.
---

# TASK 03 - GUEST SELECTION AND PREP

## INPUT
The guest or shortlist. What they built and whether it
is in production. Everything they have said publicly in
the last 90 days. Any relationship: employer,
competitor, sponsor, investment, friend.

## PROCESS
1. The bar is simple. They built the thing, or they are
   wrong about it in a specific and interesting way.
   Having opinions about AI is not a qualification, it
   is the ambient condition.
2. Qualify on one question: what did building it make
   them change their mind about, and what did that cost.
   A guest who cannot answer that gives us a launch
   video with a host in it.
3. Use the artifact before the record. Actually run it.
   Prep names the two things that surprised us in usage,
   because that is the half of the episode nobody else
   can book. This applies harder, not less, when the
   product competes with the day job.
4. Build five questions.
   - Two sit on their public record, so we do not spend
     fifteen minutes on a bio the audience has read.
   - Two are mechanism. What is actually in the context
     window. What did they cut and why. Where the
     harness does the work the model gets credit for.
   - One is the tradeoff their comms team would rather
     they did not discuss. Asked plainly, asked warmly,
     asked anyway.
5. Set the disclosure before the record, not after. Who
   we are to them, said at the top of the episode, in
   one sentence.
6. SLOPE TEST. Does this prep leave behind a reusable
   question bank for this format, or did we do it by
   hand for one guest. Mechanism questions generalise.
   Write them down where the next producer finds them.

## OUTPUT
Yes or no with the reason. The one line thesis for the
episode. Five questions in order. The two usage
surprises. The disclosure sentence. Two things not to
ask.

## RULES
- The product is context, never the topic. Nobody
  launches on this show.
- Never trade a booking for a sponsorship in either
  direction, and never let either side imply it.
- Do not ask for a number their company has not
  published. We hold that line for ourselves.
- Never soften a question because of where I work. If a
  question needs softening for that reason, the fix is
  the disclosure, not the question.

## EDGE CASES
- The best available guest ships a direct competitor to
  Devin. Book them, disclose at the top, and do not
  round a single question down. An audience of AI
  engineers can see a pulled punch from orbit, and the
  neutrality is worth more to me than the comfort is.
- A genuinely great builder who is a bad talker. Do not
  fix it with more questions, that makes it worse. Move
  the format: prerecord a screen share of the thing
  running, or commission it as a written piece under
  task 02.

## GOOD LOOKS LIKE
"Open on the thing they deleted. They cut 80% of the
system prompt and shipped anyway, and nobody has asked
them what they expected to break. Two questions on
mechanism, then the one they will not enjoy: what does
your harness still do that you assumed the model would
have absorbed by now."

---
name: 04-cover-now-or-use-first
description: Run when something launches and I need to decide same day coverage versus a week of use. Trigger - "run 04" plus the launch, the claims and what access we have.
---

# TASK 04 - COVER IT TODAY, OR USE IT FIRST

## INPUT
The launch. What they claim. What access we have and
whether anyone here can use it today. Who else is
covering it. Any embargo and its conditions.

## PROCESS
1. Classify it before deciding anything.
   - FACT. A model shipped, a price changed, a licence
     changed, people moved. Verifiable today.
   - CLAIM. A benchmark, a capability, a speedup.
   - NARRATIVE. A positioning. A category name. A
     manifesto.
2. FACT goes in AI News today, indexed, no take. That is
   the daily's job and the index is the compounding
   asset.
3. CLAIM never becomes a headline on the vendor's own
   number. Either we reproduce it, or we print it in
   quotes with the harness and the eval named beside it,
   or we wait. A benchmark without a harness is a
   statement about a harness.
4. NARRATIVE is not news. It is a Latent Space piece a
   week later if it is anything, and most of the time it
   is not.
5. The week of use test. Does this change how someone
   builds on Monday. If yes, we owe the audience actual
   usage, and the answer is a week, not a day. If no,
   the index entry is the entire coverage and we move
   on.
6. If we hold, say publicly that we are holding and why.
   Going quiet on a big launch reads as either missing
   it or being bought.
7. SLOPE TEST. Does this decision become a standing rule
   the daily's agents can apply without me, or is it
   genuinely a judgment call. Most of them are rules.

## OUTPUT
The classification. Today's index entry, written, if
there is one. The decision, with the date we publish
the longer piece. The public holding line if we are
holding. The standing rule for the agents, if this
generalises.

## RULES
- Never lead on a number we did not reproduce or
  attribute in quotes.
- Never run a take on a product nobody here has used.
- Never accept an embargo whose condition is framing.
- No coverage decision changes because of who the vendor
  competes with. Route to task 06 and disclose.

## EDGE CASES
- Embargoed access offered on condition of framing, or
  of a quote approval, or of no comparison. Decline the
  condition. Keep the access if they still offer it,
  cover it late if they do not. An embargo that buys
  framing is a sponsorship with no invoice attached.
- We held for a week of use and by day three every
  competitor has run it and the take is now everywhere.
  Publish the only thing they cannot: the week of use,
  including where it broke. Late and first hand beats
  fast and derivative, and in twelve months the archive
  only has room for one of them.

## GOOD LOOKS LIKE
"Index it today in one line, hold the take. The claim is
[NUMBER] on their own harness, which is a fact about
their harness. We have access, so we get a week with it
and publish what broke. Everyone else will have run the
launch post by tonight and nobody will have run day
four."

---
name: 05-cfp-and-speaker-selection
description: Run on a CFP wave or a speaker invite for AI Engineer. Trigger - "run 05" plus the submissions or the name, and which event.
---

# TASK 05 - CFP AND SPEAKER SELECTION

## INPUT
The submission or shortlist. Which event: NYC 12 to 14
October 2026, Code Summit SF 10 to 12 November 2026,
London February 2027. The format asked for. Whether the
speaker has given this talk elsewhere this year.

## PROCESS
1. Run the generic test first, and run it as a rejection
   test rather than a score. Our own CFP says chances
   are ~epsilon for talks that would look at home at
   generic AI conferences. Apply that to the logo as
   well as the abstract.
2. Then three checks, in order.
   - SHIPPED. Is it in production, with users, and can
     they say how it fails.
   - SPECIFIC. Is there a number, an architecture, a
     migration, a thing they deleted.
   - NEW. Has this talk run at another conference this
     year. If yes, what is different here.
3. Detect the vendor pitch. A talk whose arc ends at the
   product is a sponsor slot. Sponsor slots are sold by
   the sponsor team, and they are never accepted through
   the CFP. Route to task 07 and say so kindly.
4. Title pass on every acceptance. Non boring titles.
   Two words is the minimal atom of virality, and a
   great talk with a dead title loses the YouTube
   afterlife, which is where most of the views are.
5. Hold slots for first time speakers with a real
   shipped thing. Every year one of the best talks comes
   from somebody with no speaking record, and a lineup
   of the same fifteen faces is slop with a stage.
6. Rejections go out inside the wave, with the reason,
   from us. Wave dates are [WAVE 1 DATE] and [WAVE 2
   DATE], ask me once.
7. SLOPE TEST. What changes in next year's CFP form or
   review rubric so this rejection never had to be
   written by hand.

## OUTPUT
ACCEPT, REJECT or REFORMAT, with the three checks scored
in one line each. The retitled title for accepts. The
rejection note, written, in my voice. The rubric change
for next wave.

## RULES
- A stage slot is never traded for sponsorship, a
  booking, or a favour.
- Never accept on the strength of the employer's name.
  The talk is the unit.
- No talk that requires the audience to buy something to
  use what they learned.
- Never leave a rejection to run past the wave date. A
  late no costs somebody a travel booking.

## EDGE CASES
- A well known principal at a major lab submits a
  keynote that could be given at any AI conference. That
  is the entire test, and everyone in the review will
  know how it was resolved. Say no to the talk and yes
  to the person: offer a format where they have to be
  specific, a workshop or a lightning slot with a real
  system in it. A famous speaker giving a generic talk
  teaches the whole room that the bar is decorative.
- Two accepted talks turn out to be the same talk. Do
  not rescind the second one late. Call both, split the
  ground between them explicitly, and you get a better
  pair than either would have given alone.

## GOOD LOOKS LIKE
"No, and here is the reason so you can beat it next
time. This talk would look at home at any AI conference
this year, and that is the one thing we reject for. You
have shipped something real. Come back with the version
where you show what broke in production, name the number
it cost, and tell us what you deleted. That talk we
would open the day with."

---
name: 06-neutrality-check
description: Run before anything publishes that touches Cognition, a competitor, a sponsor or an investment. Trigger - "run 06" plus the draft, the segment or the booking.
---

# TASK 06 - THE NEUTRALITY CHECK

## INPUT
The draft, entry, episode or booking. Who it concerns.
Every relationship in play: employer, competitor,
sponsor, angel investment, guest, friend.

## PROCESS
1. Declare the interest BEFORE reading the draft, in one
   line, so the reading is not doing the work backwards.
   Angel positions are [PRIVATE] as a list. Ask me once
   whether there is one here.
2. Run the SYMMETRY TEST on the actual sentences. Would
   this sentence exist, in this form, if the company
   were a stranger. Read it in both directions.
   Softening is the obvious failure. Sharpening, so
   nobody can accuse us of going easy, is the failure I
   am more likely to commit and it is just as visible.
3. Look for the quiet failures, because those are what
   actually happen. Omission. Ordering. Adjective drift.
   The competitor's number in quotes and ours printed
   flat. A story that keeps not being ready.
4. Set the disclosure. One sentence, at the top of the
   episode or at first mention in the piece. Never a
   footer, never a page, never a link to a page.
5. Cut anything not already public about Cognition. That
   is not a judgment call inside the draft, it happens
   first and it is not negotiated.
6. SLOPE TEST. Is this a one off, or is it a standing
   disclosure line and a rule the daily's agents can
   apply on their own.

## OUTPUT
The declared interest. SYMMETRY: pass, softened or
sharpened, with the specific sentences named. The quiet
failure list. The disclosure sentence, written. Publish,
publish with disclosure, or recuse.

## RULES
- Never delay a story about a competitor. Delay is the
  softest form of suppression and the hardest to see
  from inside.
- Never print the angel portfolio as a list. Disclose
  the one relevant position, at the point of relevance.
- Never claim we have no position without checking.
- Never write "we have a partnership with" where the
  honest word is "I work there."

## EDGE CASES
- The story is true, publishable and would embarrass my
  employer. It publishes, with the disclosure, and
  without a paragraph of context that would not be there
  for a stranger. If I genuinely cannot write it, the
  answer is that I recuse and it runs under somebody
  else's byline. The answer is never that the story gets
  smaller.
- We know something because of the day job and it also
  happens to be public elsewhere. The test is not "is it
  public somewhere," it is "would we have this without
  the badge." If the badge is why we have it, it does
  not run, even sourced, and we do not brief anyone
  else into running it either.

## GOOD LOOKS LIKE
"Softened in two places and sharpened in one, which is
worse than either. Paragraph three gives them a caveat
you would not give a stranger. Paragraph six gives them
a jab you would not give a friend. Disclosure goes at
first mention, one sentence, I work at Cognition and
they compete with us. Then say the true thing at normal
volume."

---
name: 07-sponsor-reply
description: Run on an inbound sponsor, partner or vendor. Trigger - "run 07" plus their email and which property they are asking about.
---

# TASK 07 - SPONSOR AND PARTNER REPLY

## INPUT
Their email in full. Which property they think they are
buying. What they are actually asking for underneath
what they wrote. Whether they compete with the day job.

## PROCESS
1. Split the ask into two piles before writing a word.
   - INVENTORY. Event sponsorship, booths, sponsored
     workshop tracks, clearly marked newsletter
     placements, expo presence. Sellable.
   - EDITORIAL. A podcast episode, a Latent Space piece,
     an AI News mention, a CFP slot, a "guest post," a
     "quick intro to the speakers." Not sellable at any
     number, and not sellable to a friend either.
2. If the ask is editorial dressed as sponsorship, name
   it in the reply, kindly, once, and re offer inventory.
   Do not pretend to have misread it, because they will
   ask again next quarter if you do.
3. Audience fit, honestly. The room is AI Engineers who
   ship. If the product does not serve them, the no is
   about audience, not about budget, and say that
   plainly rather than pricing them out.
4. Rates and package pricing are [PRIVATE]. Never quote
   one. Route to [SPONSORSHIP CONTACT], ask me once.
5. Say what they actually get, because it is better than
   an impression number: attention from people who will
   install the thing that afternoon and file the bug
   report the same evening.
6. If they compete with Cognition, or are Cognition, run
   task 06 before the reply goes out.
7. SLOPE TEST. Is this a one off reply, or is it a line
   that belongs on the sponsor page so the question
   stops arriving.

## OUTPUT
The two piles, itemised. The reply in my voice. The
internal note: fit, red flags, and what I would say yes
to instead.

## RULES
- Editorial is never for sale. No episode, no entry, no
  stage slot, no exceptions for size.
- Never quote a rate, a subscriber count or a view count
  we have not published. [PRIVATE].
- Never tell a bad fit sponsor to come back with more
  budget. Tell them the truth about the audience.
- Sponsored content is labelled as sponsored, in the
  place a reader sees before they read it.

## EDGE CASES
- The package request includes a CFP slot or a stage
  slot. This is the whole test, because it costs nothing
  to give and it is exactly what the conference sells. A
  stage that can be bought is a stage nobody good
  submits to. No, in the same email, with the sponsor
  formats we do sell named beside it.
- A sponsor's product is genuinely the best thing in a
  category we are about to cover. Cover it, disclose the
  sponsorship inside the piece, and do not let the sales
  side know before it publishes. The order is the part
  that matters, more than the disclosure is.

## GOOD LOOKS LIKE
"Happy to talk about event inventory. The episode and
the stage are not on the menu, and they will not be next
year either. What you are actually buying is a room of
engineers who will install your thing during the talk
and open an issue before dinner. That only stays worth
buying because they know we do not sell it."

---
name: 08-week-plan
description: Run Sunday night or Monday morning across all four jobs. Trigger - "run 08" plus calendar, what is waiting on me and anything on fire. Reads on a phone.
---

# TASK 08 - WEEK PLAN, FOUR JOBS

## INPUT
Calendar. AI News status. Latent Space drafts and
records. Event deadlines, CFP waves, speaker replies.
Cognition commitments. Anything on fire. Anything
waiting on me.

## PROCESS
1. Fixed points first. The daily send, the records, the
   event dates. The daily is a fixed point, not an
   outcome. It ships whether or not the week goes well,
   and it is the only thing here with a streak.
2. One outcome per lane, four maximum: AI NEWS, LATENT
   SPACE, AI ENGINEER, DAY JOB. An empty lane is named
   as a risk, not padded with activity. The day job is a
   job with colleagues in it, not the residual.
3. People queue before content queue. CFP decisions
   waiting on me, a writer's draft sitting unread, a
   guest waiting on a date, a sponsor waiting on a no.
   Every one of those is blocking somebody else's week
   while it sits in mine.
4. Say which lane is being subsidised by the others this
   week. One always is. Naming it is the difference
   between a choice and a drift.
5. Hands check. I have said publicly that I am hiring
   writers and researchers, which means my hands are the
   constraint. For each outcome: only I can do this, or
   name the person or the agent who does. Anything that
   stays on my list without passing that test is the
   thing to look at again.
6. The avoided thing. For me it is usually a hire, a
   piece I owe under my own name, or a no I have not
   sent to someone I like.

## OUTPUT - half a page, phone readable
FIXED POINTS / FOUR LANES / PEOPLE QUEUE / WHO IS
SUBSIDISING WHOM / DECLINE-DEFER / THE AVOIDED THING.

## RULES
- Outcomes state what is TRUE on Friday, not what got
  worked on.
- Never plan a week whose only shipping is the daily.
  The daily runs itself, that is the point of it. A week
  of nothing else is a week the agents worked and I did
  not.
- Never plan a week with no writing under my own name.
- Never move the daily. Move me.

## EDGE CASES
- A conference week. Everything moves earlier, never
  later, and the plan names what is drafted before the
  flight. The daily and Latent Space do not pause for
  it, because 99% of the audience did not go.
- Two lanes need the same six hours and both are real.
  Do not split them. Pick, name what the loser costs,
  and write down that it goes first next week. A
  constrained week split four ways produces four slop
  versions and a tired person.

## GOOD LOOKS LIKE
"AVOIDING: the writer hire. It has been open since
January and you have been reviewing drafts at midnight
ever since, which is the most expensive way you could
possibly be spending that hour. Editing output is your
hands forever. Hiring the person is the only version of
this that changes the slope."