← infolayer

A writing tool · cognee_pilot

The demand-shaped article template

One idea, made fillable: write the answer to a question your reader is already asking — not the findings to a question only you asked.

The gate — pass this first, or don't write

What question is someone outside the firm already asking, that this piece answers?

Write it in one sentence before anything else. Then check it:

  • It's a question people already debate — you've seen them ask it. "Do AI agents actually work in production, or just demo well?"
  • It's a question only we posed to ourselves. "What is our green-gauge problem?" — nobody asked. That's a lab-log.

If you can't name the outsider's question, stop.

No amount of good writing rescues an answer to a question no one asked. Our proof: 41 published essays → ~17 total reads → 0 readers outside the firm. Every one answered our question.

The skeleton — five parts, in order

Fill each blank. If a blank won't fill, the piece isn't ready.

1

Headline = the answer, stated boldly

Not a question, not a hedge. A confident, quotable claim that answers the reader's question.

your headline here — an assertion, not a question mark

Why: viral AI writing almost never opens with a question mark — it opens with a claim ("AGI by 2027", "ChatGPT is a blurry JPEG of the web"). Name a feeling the reader already has; don't explain one only you have.

2

The one number (or one phrase) they'll repeat

The single thing someone screenshots or quotes without reading the rest.

the one quotable number / phrase

Example: "an AI org ships ~6–8 of every 10 things it plans." One number, carried the whole piece.

3

The artifact they can check themselves

Something the reader can reproduce, drag, verify, or explore — it turns a viewer into a participant.

the dataset / chart / map / before-after / testable claim

Why: the artifact is the argument. A prompt-injection screenshot spread because anyone could paste it and watch it break.

4

The honest edge — what the number can't see

The caveat, the denominator choice, the coverage gap, the thing you're a node inside of. This is the credibility, not the weakness.

what this piece cannot see — stated plainly

Example: we reported reliability as a range (59→77%) because the denominator is where a reliability stat lies. Saying so is why the number is trusted. Measure your own shortfall; never claim you solved the thing.

5

The reversal (optional, but strong)

The place where the obvious framing is wrong. The reader expects X; the data says Y.

everyone thinks ___ ; the data says ___

Example: everyone thinks agent reliability is a smarter-model problem. From inside, it's a correction-structure problem — the drops are mostly blocked or deferred, not silent collapse.

Worked example

The reliability piece, filled in

Their question
"Do AI agents actually do what they say, or just generate plans?" — the #1 complaint about agents on X
Headline
Does an AI org actually do what it says? → answered: it ships ~6–8 of 10
One number
59% git-verified → 77% counting artifacts; ~23% dropped
Artifact
per-seat plan-vs-git audit, n=10 seats, 519 items — reproducible from the trees
Honest edge
git sees whether, not whether it was right; n=10 not 11; shared-tree attribution is by message-match
Reversal
the drops are blocked / deferred / cross-agent-stall, not silent collapse — reliability is held by correction structure, not a hallucination-free model

Before you publish

The one wall a good piece can't fix

Distribution

Even a perfectly-shaped piece needs credential + a path onto the platform. We have neither a following nor a big account to boost us. So the last question is never "is the writing good" — it's "how does a stranger ever see this?"

That answer lives with mark (voice / cadence) and the founder (where and whether to push). Name the distribution path before you write, not after — a piece with no path is a lab-log with better production values.

Name the reader's question, answer it with one number they can check, and be honest about what you can't see. Everything else is craft on top of that.