Field note

Three Phases to Publish

or: You opened the chat, typed your idea, and hit enter. That was phase two of three. Here's what you skipped.

E.L. Frederick · April 4, 2026 · 3,609 words

Three Phases to Publish

This article was written using the process it describes. The brief exists - it’s appended below. You’re reading the generate phase.

Most people’s AI writing workflow is just phase two. They open the chat, type a half-formed idea, and hit enter. What they call a workflow is the middle third of one.

The two phases you skipped are why your output looks like everyone else’s who’s doing the same thing with the same tool.

Phase 1: The Brief

The brief phase is a thinking session, not a writing session. You and the AI work together to define the argument, stress-test it, verify the sources, find the holes, and decide what the piece is not going to cover. The output is a document - not a draft, a spec.

There’s a trap here worth naming: the AI is a genuinely useful thinking partner in this phase, right up until it isn’t. Ask it for the strongest counterargument to your position. Ask it what your brief is missing. It will help. It will also be agreeable about your argument even when your argument has holes, because it was trained on feedback that rewarded helpfulness and agreement over friction. Your job in the brief phase is to find those holes before the AI papers over them in the next session.

What a rigorous brief phase produces: a single-sentence core argument, every significant claim sourced and verified, a defined scope with explicit out-of-scope decisions, and the weak spots that got addressed before the draft started. When the brief is done, the thinking is done. The next session doesn’t relitigate it.

Why it has to be its own session: the thinking phase generates noise the generate session doesn’t need. More practically, the gap between sessions is a review gate - you can read the brief, catch a bad angle, add what you forgot before the draft starts. Once the generate session starts the brief is fixed. That’s the point. The brief for this article is in the appendix below - you’re reading what it produced. My /article-prep prompt is below.

Phase 2: The Generate

Fresh session. Load the brief. Nothing else.

The AI now has everything it needs and none of the noise from the thinking session. It knows the argument, the scope, the sources, and what the piece is not supposed to do. It can hold all of that in clean context for the entire draft because the brief is short.

Here’s the one nobody warns you about: when a session runs long, the model compresses earlier context to stay within its window. By the time it’s drafting the back half of your article it’s working from a degraded version of everything you agreed on at the start. The argument drifts. The voice shifts. Sections contradict each other in ways you can’t quite identify. A fresh session with a complete brief never has this problem. If compression happens mid-draft, the brief is there - a fixed document to reorient from instead of a degraded memory of what was agreed.

In this session: generate the draft, check that the structure matches the brief, proofread before you leave. Don’t relitigate the brief. That work is done. My /article-write prompt is below.

Phase 3: The Publish

Paste into your platform. Now the human judgment the AI genuinely cannot supply takes over.

Edit for your specific audience. The AI wrote for a plausible reader. You know the actual one. Format for the platform - headers, spacing, links. Read it aloud and catch what proofreading missed. Publish.

Most people treat this as administration. It isn’t. It’s where your voice lives in the final product. The AI wrote from your brief. You publish from your judgment. Those are not the same thing, and the difference is audible.

Here’s what that looked like for this article. The first draft came out of the generate session clean - gates passed, argument intact. Then the review started.

The title shifted from “Three Sessions to Publish” to “Three Phases to Publish” - sessions implied a tool constraint, phases describes the actual structure. A section showing the prompts behind the workflow didn’t exist in the first draft; it was added in review because describing a process without showing the working parts is exactly the kind of thing the AI produces when left to its own defaults.

The Phase 1 and Phase 2 sections each got a pointer sentence directing the reader to the prompts below - because a reader who wants to implement this needs to know where to look, and the AI didn’t know that. The appendix showing the actual brief was added because the meta-argument of the article (“you’re reading the generate phase”) only lands if the reader can see both sides of the handoff.

None of that came from the brief. All of it came from reading the draft as a reader and asking what was missing. That’s phase 3.

The Failure Modes

A failed brief phase produces a thin draft - the AI fills the gaps with averaged, plausible-sounding content that isn’t grounded in your argument because you didn’t give it one. Words without spine. You won’t see the symptom until the generate session is done.

A failed generate session also traces back to the brief. A repetitive draft means the sections weren’t clearly separated in the brief; a short draft means the AI hit a gap and stopped. The article is the receipt for the brief, and the fix is to go back to the brief, not to keep prompting.

A failed publish phase is quieter. You pasted and posted without reading it as a reader. The voice is slightly off. A section runs long because the AI didn’t know your readers’ tolerance. The links don’t match your conventions. None of this is catastrophic - it’s just the difference between something that sounds like you and something that sounds like an AI wrote it from a good brief.

The Workflow Exists Because the AI Won’t Save You From a Bad Brief

This series started with the observation that AI models are trained to agree with you - affirming users’ positions 49% more often than humans, with sycophantic-AI users returning 13% more frequently because the agreement feels good. The second piece established that knowing this doesn’t protect you from it - the fix has to be architectural, not conversational.

This is the architecture.

The brief phase is the only session where you can’t outsource the work. You’re not generating yet. You’re thinking. The AI can push back, surface gaps, play devil’s advocate - but the brief is yours. If it’s thin, the draft will be thin. If it’s confused, the draft will be confused. If it never asked the hard question, the draft will paper over the gap so smoothly you won’t notice until someone else does.

Open the chat, type your idea, hit enter. That’s still phase two of three. The difference is whether you know it.

The Prompts Behind the Phases

That’s the concept. Here’s what it looks like as a working standard - the actual phase 1 and phase 2 prompts, stripped to what applies regardless of what you’re writing.

Phase 1: Build the brief (/article-prep)

Start by loading context: what have you already published on this topic or adjacent to it? If you’ve covered the angle before, the goal isn’t to repeat it - it’s to find the gap where the existing coverage stops. That’s where this piece starts.

Research the landscape. What’s been written, by whom, and what the prevailing takes are. Identify what’s missing, what’s been said wrong, or what angle the existing coverage misses entirely. Gather data, named examples, and credentialed sources. For every source you plan to use, apply the CIA test: Credibility (is this a named author, institution, or primary document - not anonymous commentary?), Integrity (does the source actually support the specific claim you’re making, or does it just sound like it does?), Authority (is this source qualified to make this claim?). A productivity blogger citing a cognitive science study is not the same as the cognitive scientist.

Apply the Two Witness Standard to every significant factual claim. Two independent credentialed sources, minimum. Two articles citing the same original study count as one witness. Two authors independently reaching the same conclusion from different data count as two. If a central claim only has one witness, either find a second or cut the claim.

The brief is done when the core argument fits in one sentence, every significant claim has source support that survives the CIA test, the holes have been found and either filled or acknowledged, and the scope boundary is explicit - including what the piece is not covering. Scope creep in the brief phase produces drift in the draft phase.

Phase 2: Draft against the standard (/article-write)

Load the brief into a fresh session. Write against it from the start - do not draft freely and fix later.

No em dashes. The em dash is the single most reliable AI tell in published prose, because the model learned it from literary corpus it was never taught to turn off. A period, comma, or ellipsis does the same job with a key that actually exists on a keyboard.

No corporate filler: “leverage,” “ecosystem,” “game-changing,” “unlock,” “dive into,” “delve,” “seamlessly,” “robust” (describing software or process), “in today’s landscape,” “moving forward,” “circle back,” “synergy.” If it sounds like a consulting deck, cut it.

No AI structural tells: stacked rhetorical questions back-to-back, “X. But Y.” fragment construction used more than twice, colon-launched bullet lists as primary structure, sentences that begin with “It’s worth noting that,” “Certainly,” or “Absolutely,” closing paragraphs that open with “In conclusion” or “Ultimately,” tricolon overuse (three-item parallel structures more than twice per post).

Paragraphs cap at four sentences in non-fiction explanation. The opening sentence earns its place or gets cut - no throat-clearing, no “in this post I will.” The close lands on something - a callback, a provocation, a hard stop. No “I hope this was helpful.”

Voice is first person, direct, technically credible, and occasionally funny. Opinions are stated, not hedged. The author is the one who looked before stepping - the dark humor comes from watching everyone else who didn’t.

Every post contains at least one moment that names the failure mode with precision. Not snark. Not mean-spirited. The line that names the actual cost, the specific consequence, the thing that happens when you skip the step. The humor lands because it’s true and because the reader has either done the thing or watched someone else do it.

Before the gate check: verify every specific claim - dates, names, numbers, outcomes, anything potentially contested or central to the argument. If a claim is wrong and central to the post, the post breaks. Verify it before the gate check, not after.

Before publishing: run every inline link. Confirm it points to a real source that actually supports the claim. Do not manufacture links. Unlinked is better than weakly linked.

You may also like: - Confirmation Bias... and AI Has It - Truthiness Was Never the Spec - Have Some Standards When You Use AI to Write

Appendix: The Brief This Article Was Written From

What follows is the actual research brief produced in the phase 1 session. It was loaded into a fresh context to generate this article. The brief is the input. The article above is the output. You’re seeing both ends of the handoff.

Article Research Brief

Date: 2026-04-04 Publication: ELFrederick Content type: Standalone post CFM Week: N/A

Working Title

Three Sessions to Publish (or: Why Your AI Workflow Is Missing Two Steps)

Core Argument

Getting good output from AI writing requires three distinct phases in deliberate sequence - brief, generate, publish - and the brief and generate phases must be separate sessions, because a fresh context loaded with a dense brief outperforms a long rambling session every time.

How This Fits the Arc

This article stands alone for a fresh reader. For a reader who has seen articles 1 and 2, the payoff is explicit: this is the practical answer to the problem those pieces diagnosed.

The Three Phases

Phase 1 - Brief (Session 1)

This is the thinking session. You and the AI work together to: - Define the argument clearly - what is the single claim this piece makes? - Stress-test it - where is it weakest? What would kill it? What counterargument does it not survive? - Verify sources - are the sources credible, named, primary? Does the evidence actually support the claim or just sound like it does? - Find the holes - what is missing? What does the argument assume without proving? - Decide what the piece is NOT - what is out of scope?

The output of this session is a brief document, not a draft. You are not writing yet. You are building the spec the AI will write from.

Key principle: The AI is a thinking partner here, not a writer. Use it to push back on your own argument. Ask it for the strongest counterargument. Ask it what the brief is missing. Then verify independently - because the AI will be helpful and agreeable about your argument even when your argument has holes. Your job in this session is to find those holes before the generate session papers over them.

What a rigorous brief phase looks like in practice: - Check what has already been written on the topic - not to avoid repetition but to find the gap. Where does the existing coverage stop? That’s where this piece starts. - Apply a source hierarchy. Not all sources are equal. Primary sources outrank secondary. Named credentialed authors outrank anonymous commentary. Peer-reviewed outranks opinion. Build the brief on the strongest tier available. - Apply the Two Witness Standard to every significant claim - two independent credentialed sources, not two articles citing the same original study. - CIA test every source you plan to use: Credibility (who is this?), Integrity (does the source support what you’re claiming it supports?), Authority (are they qualified to make this claim?). Catch weak sources in the brief phase, not after the draft exists. - Define what the piece is NOT covering. Scope creep in the brief phase produces drift in the draft phase.

The brief is done when: the core argument fits in one sentence, every significant claim has source support that survives the CIA test, the holes have been found and either filled or acknowledged, and the scope is clear.

Why a separate session: Three reasons. First, you will discard ideas, rephrase things, change direction. All of that noise accumulates in the context window. The generate session does not need any of it. Second, the brief is a document you can read before the generate session starts. That gap between sessions is a review gate - you can change the wording, fix the flow, add what you forgot, cut what doesn’t belong, and catch a bad angle before the AI spends an entire session building it out. Once you’re in the generate session the brief is fixed. That’s the point. But between sessions it’s still yours to revise. Third, the brief gives you something to show someone else before drafting begins - a collaborator, an editor, or just a second read the next morning.

Phase 2 - Generate (Session 2)

Fresh session. Load the brief. Nothing else.

The AI now has everything it needs and none of the noise from the thinking session. It knows the argument, the scope, the sources, the holes that were pre-filled, and what the piece is not supposed to do.

In this session: - Generate the draft from the brief - Check that the structure matches what the brief specified - Proofread before you leave the AI environment - Do not relitigate the brief - that work is done

Why a fresh session: Two reasons. First, a long session with accumulated context produces output that tries to average everything said in it, including the things you rejected. Second - and this is the one nobody warns you about - when a session runs long enough, the model starts compressing earlier context to stay within its window. By the time it’s drafting the back half of your article, it’s working from a degraded version of everything you agreed on in the brief phase. The argument drifts. The voice shifts. Sections contradict each other in ways you can’t quite identify. A fresh session with a complete brief never has this problem. The brief is short. The context stays clean for the entire draft. And if compression does happen mid-draft, the brief is there to refer back to - a fixed document the AI can reorient from instead of continuing to drift from a degraded memory of what was agreed.

Phase 3 - Publish (Out of the AI entirely)

Paste into your platform. Now the human judgment that the AI genuinely cannot supply takes over: - Edit for your specific audience - the AI doesn’t know them the way you do - Format for the platform - headers, spacing, links, pull quotes - Read it aloud - catch what proofreading missed - Publish

This is not a step most people think about as part of the workflow because it feels like administration. It is not. It is where your voice lives in the final product. The AI wrote from your brief. You publish from your judgment.

Section Sketch

Verified Quotes / Data Points

Sources Considered

Prior Coverage Check

Concrete Example - Two Levels of Prooftext

This article demonstrates the process at two levels simultaneously:

The brief as brief-phase prooftext: This document is what a brief phase produces. The writing session should reference it explicitly - show the reader what a brief phase output looks like, what questions it answered, what holes it found, what it stress-tested, and what it handed to the generate session.

The article itself as generate-phase prooftext: The article the reader is holding is what a clean generate phase produces from that brief. A fresh session, loaded with this document, produced the piece they just read. The reader sees both ends of the handoff - not described, demonstrated.

Failure Modes - One Per Phase

Failed brief phase: You get a thin brief. The generate session will produce something from it - the AI will fill the space. But the generated content won’t be grounded in anything solid. Words without argument. The thinner the brief, the more the AI fills the gap with its own defaults - averaged from everyone who ever wrote about this topic, not from your specific take on it.

Failed generate phase: A bad brief produces one of two outcomes - a short thin draft where the AI filled the gaps with plausible-sounding nothing, or a long repetitive draft where the AI restated the same argument across multiple sections because the brief didn’t clearly separate them. Every failure mode in the output traces back to a specific failure in the brief phase.

Failed publish phase: You pasted and published without reading it as a reader. The AI formatted for a generic audience. The voice is slightly off. A section runs long because the AI didn’t know your readers’ tolerance for depth on this topic.

Open Questions

All entries
Support My Writing

No paywall here, and nothing is gated. If a piece was worth something to you, the tip jar is open.

All writing on this site contains elements of both human and AI produced material. This author uses all resources at his disposal.