Part 8 Plan: How I Wrote This Series With AI

Working title options

  1. How I wrote this series with AI
  2. The postmortem: writing a series about agentic systems with agentic tools
  3. This series was mostly AI-assisted. That is not the interesting part.
  4. The writing process proved the thesis

What this article should do

This piece should not read like a novelty disclosure, a tool review, or a defensive justification.

It should read like a postmortem.

The article has four jobs:

  1. Acknowledge plainly that the series was heavily AI-assisted.
  2. Explain the actual workflow in enough detail to be useful.
  3. Show that the process only worked once it became structured, staged, and artifact-driven.
  4. Reflect on what I would change next time to reach the same quality faster.

The key move is to make the process itself an example of the series thesis.

Core thesis

The central claim for Part 8 should be:

The interesting part is not that AI helped write the series. The interesting part is that the writing only became good once the work was turned into a system: plans, artifacts, review loops, explicit roles, and human judgment at the irreversible edges.

Supporting claims:

  1. AI accelerated exploration, drafting, and revision.
  2. AI did not automatically produce coherence, quality, or taste.
  3. The biggest gains came from structure, not model cleverness.
  4. The final quality came from repeated human editorial correction.
  5. The writing workflow ended up mirroring the argument of the series itself.

Narrative spine

The article should move through this sequence:

  1. Direct disclosure
  2. Reframe the real lesson
  3. Walk through the pipeline that emerged
  4. Show what worked and why
  5. Show what failed and why
  6. Explain what would have reached the outcome faster
  7. Close by linking the process back to the thesis of the series

The article should feel less like “here are the tools I used” and more like “here is the system I accidentally built.”

Detailed section plan

1. Opening: say it plainly

Goal

Establish trust quickly. Do not bury the disclosure.

What to say

Desired effect

Readers should feel that the article is candid and precise, not evasive or theatrical.

Opening angle

Use a sentence like:

This series was largely AI-assisted, but the surprising part is not that the tools could generate prose. It is that the prose only became worth keeping once the work was turned into a structured system.

2. Reframe: this was not one-shot generation

Goal

Kill the naive picture early.

What to say

Reflection to merge from both notes

Key line to build toward

I was not using AI as a ghostwriter. I was building a pipeline that let different models do different kinds of work under constraint.

3. Explain the actual workflow in order

Goal

Give the reader a concrete process they can inspect.

Section structure

Break this into five short subsections.

3.1 Discovery

3.2 Architecture

3.3 Evidence

3.4 Artifacts and persistence

3.5 Drafting and review

Critical interpretation

This is where the article should say explicitly:

The important shift was not “AI wrote the article.” The important shift was that the writing process became decomposed into stages with stable inputs and outputs.

4. What actually made the process work

Goal

Extract the principles, not just the chronology.

Subsection A: Role separation

Make the point that role separation worked better than using one undifferentiated assistant for everything.

Subsection B: Artifacts

Call them artifacts explicitly.

Subsection C: Iteration over prompting

Subsection D: Governance

Best synthesis sentence

The system worked because I kept turning fuzzy conversation into explicit artifacts, then forcing each later stage to operate on those artifacts instead of improvising from scratch.

5. Where AI failed

Goal

Show the reader that the real problems were deeper than cosmetic glitches.

Subsection A: Structural errors

Use the line:

The hardest errors were structural, not stylistic.

Subsection B: Source quality failures

Subsection C: Generic prose

Subsection D: Operator dependence

Best synthesis sentence

The models were good at producing text that looked finished. They were much less reliable at producing argument that actually was finished.

6. What I would do differently to reach the outcome faster

Goal

This is the most useful section. It should be concrete and operational.

6.1 Lock constraints earlier

6.2 Add explicit planning artifacts

Create these before drafting begins:

6.3 Separate review types earlier

Do not mix everything into one revision prompt.

Use separate passes for:

6.4 Measure the workflow

Track:

6.5 Write reusable AI skills for the roles

This is the most important addition beyond the current two notes.

The faster path next time would be to codify the recurring roles into reusable skills instead of rediscovering the same instructions during each pass.

  1. series-architect
    • when to use: turning a broad topic into a narrative spine and article map
    • inputs: thesis, audience, article count, non-goals
    • outputs: spine, section map, forward progression checks
  2. source-triage
    • when to use: ranking sources by credibility, relevance, rhetorical value, and redundancy
    • outputs: keep / maybe / reject list and reason for each
  3. article-planner
    • when to use: converting a series spine into one article brief with claims, evidence, and transitions
    • outputs: article plan with sections, target examples, and risks
  4. structural-critic
    • when to use: reviewing a draft for duplication, weak logic, inversion, poor transitions, and drift from thesis
    • outputs: findings-first critique, not rewritten prose
  5. style-enforcer
    • when to use: checking for robotic tone, repetitive cadence, over-short sentences, summary-bot endings, and generic phrasing
    • outputs: concrete rewrite guidance with examples
  6. citation-hygiene
    • when to use: removing low-value links, replacing weak sources, and converting trailing anchors into earned inline citations
    • outputs: keep / cut / inline-quote recommendations

Important design note

Do not start with ten skills.

The faster path is probably:

  1. one strong planning skill
  2. one strong review skill
  3. one citation/style cleanup skill

Then split further only if the review prompts become overloaded.

This keeps the workflow aligned with the more defensible guidance from OpenAI: start simple, then divide roles when complexity actually demands it.

6.6 Keep a defect log from the first article onward

Every time a recurring failure appears, add it to a standing list.

Examples:

This turns frustration into system memory.

7. The meta-insight: the process mirrored the thesis

Goal

This is the real closing argument.

What to say

Strong closing line candidate

The way I wrote this series turned out to be the best evidence for the series itself: AI did not replace the system. It became useful only once it was placed inside one.

Tone guidance

The article should sound:

Avoid:

Sources to use in the published article

Use the stronger sources below. The ChatGPT note had useful reflections, but some of its suggested support came from weaker writing-advice sites. For the published article, prefer the sources in this section.

Quote 1: Human-AI collaboration is iterative, not passive

“Rather than passively accepting output, users actively refine, explore, and co-construct text.”

Use for:

Source: Microsoft Research on LLM-assisted writing

Quote 2: Incremental approaches work better than jumping to full autonomy

“Customers typically achieve greater success with an incremental approach.”

Use for:

Source: OpenAI, A practical guide to building agents

Quote 3: Only split roles when complexity actually demands it

“Our general recommendation is to maximize a single agent’s capabilities first.”

Use for:

Source: OpenAI, A practical guide to building agents

Quote 4: Specialized roles are legitimate when the task really differs

“You can create specialized custom agents for different tasks.”

Use for:

Source: GitHub Docs, About Copilot cloud agent

Quote 5: AI output looks finished before it is finished

“LLM outputs always appear at first glance to be well written, well informed and thought out, but closer reading reveals gaps, biases and lack of depth.”

Use for:

Source: Writing literature reviews with AI: principles, hurdles and some lessons learned

Quote 6: One-button automation is the wrong mental model

“A press-button strategy leaving AI to do the work is a recipe for disaster.”

Use for:

Source: Writing literature reviews with AI: principles, hurdles and some lessons learned

Quote 7: Repository knowledge and instructions improve effectiveness

“The more Copilot cloud agent knows about the code in your repository, the tools you use, and your coding standards and practices, the more effective it will become.”

Use for:

Source: GitHub Docs, About Copilot cloud agent

Suggested source handling

For the article itself:

Candidate structure for the final article

Approximate structure:

  1. Opening disclosure
  2. Why that is not the interesting part
  3. The pipeline I ended up building
  4. What actually made it work
  5. Where the models failed
  6. What I would do differently next time
  7. Why this proved the thesis of the series

Estimated length:

That is long enough to be useful and short enough not to feel like an appendix.

One-paragraph summary of the plan

Part 8 should present the series as a case study in human-AI orchestration. It should acknowledge the heavy AI assistance, explain the pipeline that emerged, argue that structure mattered more than raw generation, show the recurring failure modes that required human correction, and end by reflecting that the process itself validated the thesis of the series. The counterfactual section should explain how the same result could have been reached faster by introducing constraints earlier, separating review passes sooner, tracking defects and metrics, and codifying recurring roles into reusable skills.