Day 2 Michael Stewart SPS Live 2026

Self Publishing Show Live 2026

How to Work
With an (AI) Agent

...and why you probably want to.

Michael Stewart, Sudowrite

Full disclosure

I built this deck with my AI agent.

The ideas are mine. The words are mine. The HTML is his. The workflow was collaborative.

But first

Unpacking "Author"
(Your Actual Job Description)

Maker

IdeasPremiseGenre researchTropesComp titlesOutlining DraftingPacingCharacter arcsWorldbuildingResearchVoice DialogueScene structureRevisionSelf-editingDevelopmental editsLine edits CopyeditsProofreadingContinuitySeries bibleSensitivity readsBeta feedback Critique notesManuscript formattingAudiobook script prepCreative confidenceBurnout management

Marketer

PositioningReader promiseBook descriptionHookTaglineMetadata KeywordsCategoriesCover briefA/B testingNewsletterWelcome sequence Reader magnetARC teamLaunch teamSocial postsBookTok hooksAds Amazon AdsMeta AdsBookBub AdsPromo stackingNewsletter swapsGoodreads ReviewsWebsiteSEODirect sales copyKickstarter pageReader community Backlist promosBrand voicePress kit

Manager

Publishing scheduleProduction calendarISBNsImprint setupKDPIngramSpark Draft2DigitalKoboApple BooksGoogle PlayACX and audio platformsFiles and versions Contractor hiringCover designer briefEditor managementNarrator managementBudgetingRoyalties BookkeepingTaxesLLC or DBARights trackingContractsCopyright AI disclosureAccessibilityAlt textEPUB compliancePrint specsDistribution choices KU vs. wideDirect sales stackShopify or PayhipBookFunnelCustomer supportBacklist audits Platform changesContinuing education

But why?

It's not a question of talent.

Context switching Hard to achieve flow state switching from craft to spreadsheets.
Decision fatigue Every bit of surface area forces fresh decisions, each decision a job.
Domain expertise It takes time to become proficient, never mind master a discipline.

No one can do it all.

Not perfectly, at least. And not without help. But you will have to make decisions about all of these things, even when the decision is "This is not the best use of my time."

Consider

What to hand off?

Stuff to Keep

Can do and want to

Your creative decision making. Things like taste, voice, canon, ethics, emotional truth, reader promise, and final approval. But also, anything you want to do—things that are precious to you!

Stuff to Pass

Can't do or don't want to

Research, summaries, checklists, comparisons, edits, cleanup passes, asset drafts, calendar math. SEO? Marketing? Admin tasks (or anything you'd consider paying someone else to do).

Exercise

Make your own list.

You're doing everything on this list already, so start from a place of assuming it's all yours. Circle the things you'd be willing to hand off to a trusted assistant.

Basically

What would you be willing to hand off to a human assistant?

Important caveat

"Agent" is a squishy term.
(But then, so is AI.)

Every product you use right now is adding an "agent" of its own, but some of these are really just chatbots.

The important question is not what the company calls it. The important question is: Can it do the things you need fairly autonomously?

Let's explain.

A Very Brief History

From autocomplete to agents

01

Autocomplete

Suggests the next word or words in a string of words.

02

Chatbot

Trained to respond probabilistically to what was said to it.

03

Assistant

Trained to make decisions that help it complete tasks.

04

Agent

Understands a goal, plans steps, uses context, and helps move work forward.

Useful differentiators

What does an agent do differently?

Multi-step planning and execution
Initiates work without prompting
Improves from feedback
Creates memories
Uses custom skills
Uses tools or files
Interface options (SMS, Email, Slack, Discord, etc.)
Can interact with the world (via browser, CLI, or otherwise)

An assistant responds to requests.
An agent proactively pursues goals.

Six Things

Anatomy of an agent.

🧠 Brain

The AI model(s)

One or more large language models doing the actual thinking. This is the engine.

📋 Instructions

The job description

A prompt framework that tells the agent who it is, how it works, and what matters.

📚 Knowledge

Everything it knows

Your notes, files, prior conversations, and anything you've told it to remember.

🔧 Tools

What it can use

What you've said it's allowed to use—apps, services, the web, up to and including your whole computer.

💬 Interface

How you talk to it

Could be terminal, a browser page, a dedicated application, or even a text message. Your call.

🖥️ Runtime

What keeps it alive

The server hosting the agent—might be your laptop, but if you want it always-on 24/7, you should consider a $5/mo VPS, or a Mac Mini/NAS in your attic.

An agent lives and works somewhere, even while you sleep.

Right now

The combined tool landscape

Maker

Writing and publishing tools

Tools built to support the creative process in particular. Fiction-aware workflows, manuscript context, story bibles, drafting, revision, and writing-adjacent support.

Marketer

General AI assistants

Your chatbots are great for research and planning. Brainstorming, summarizing, drafting copy, analyzing text, making plans, and working with uploaded files.

Manager

Automation tools

Repetitive admin, notifications, publishing operations, and moving data between the tools you already use.

Multidisciplinary

Agents

These multi-tools are super capable AI-powered blank slates, intended to be personalized. They work across Maker, Marketer, and Manager domains. They may or may not have an interface, but can typically even use the tools mentioned above on their own.

Right now

The combined tool landscape

Maker

Writing and publishing tools

Tools built to support the creative process in particular. Fiction-aware workflows, manuscript context, story bibles, drafting, revision, and writing-adjacent support.

Sudowrite NovelCrafter ProWritingAid
Marketer

General AI assistants

Your chatbots are great for research and planning. Brainstorming, summarizing, drafting copy, analyzing text, making plans, and working with uploaded files.

ChatGPT Claude Gemini Perplexity
Manager

Automation tools

Repetitive admin, notifications, publishing operations, and moving data between the tools you already use.

Zapier n8n Make Airtable
Multidisciplinary

Agents

These multi-tools are super capable AI-powered blank slates, intended to be personalized. They work across Maker, Marketer, and Manager domains. They may or may not have an interface, but can typically even use the tools mentioned above on their own.

Hermes OpenClaw Claude Code Codex Antigravity

Durable workflows

Use (or create) repeatable flows.

/goal

Complete any task

Tell the agent to do something finite and it will keep at it till it's done.

/loop

Put it on repeat

Re-run a task on a cadence — a daily review, polling until something's ready, or repeating a workflow without re-asking each time.

/grill-me

Pressure-test an idea

Challenge assumptions, ask sharper questions, and find weak spots before readers do.

/canon-check

Protect continuity

Compare a scene or plan against the story facts, series promises, and open threads.

A skill is a reusable way to tell the agent how you like to work.

get the most from ai agents

No matter which agent you'd like to work with, in order to do its best work an agent needs context.

Telling the agent who you are, what you write, and how you work is key.

You need this

The Author Context File

A whole bunch of context you have in your brain (and on your hard drive) that can help the agent help you.

author-context.md
Identity
Creative preferences
Books and series
Business context
Agent rules

Context turns agents into informed & capable collaborators.

Minimum viable memory

What goes inside?

Author identity

What you write, who you write for, genres, tropes, promises, and what your work should never feel like.

Creative preferences

Drafting style, feedback style, sacred boundaries, and what kind of help you usually reject.

Books and series

Pen names, current projects, character/world details, timelines, comp titles, and reader expectations.

Voice and brand

Sample newsletter, sample blurb, sample scene, tone preferences, words you use, words you hate.

Business context

KU or wide, publishing cadence, launch process, newsletter cadence, ad comfort, revenue goals.

Agent rules

What it may draft, what it must ask before doing, what it should never do, and what good looks like.

Then what

Here's what that lets you do.

The agent can be pretty helpful once it's primed with some vitals.

Maker support Examples

Need an assist on creative admin?

Continuity & Series Memory

"I'm writing book four and am losing the thread. I need a better way to track the details."

  • Summarize prior books
  • Extract character and worldbuilding details
  • Audit your work and flag contradictions
  • Maintain a series bible
  • Create "remember before drafting" briefs

Beta Feedback Analysis

"I got some conflicting beta reader feedback, and I'm not sure how/whether to action it."

  • Analyze and group feedback by theme
  • Evaluate substance, separating taste from repeated issues
  • Identify high-priority revision areas
  • Turn messy notes into a revision plan
  • Generate scene-level tasks

Agent as a writing assistant, facilitating the creative process and surfacing possible issues for the author.

Marketer support Examples

Need help with that whole
"write to market" thing?

Positioning & Blurb Assistant

"I know what my book is, but not how to explain it to readers... at least, in a way that'll make them buy it."

  • Identify reader promise
  • Evaluate the existing market comps
  • Draft retailer description, newsletter pitch, ad hooks, and social captions
  • Adapt tone by target audience

Launch Campaign Assistant

"My book launches in six weeks and I am already behind."

  • Build launch calendar
  • Identify missing assets
  • Draft emails and ARC reminders
  • Repurpose copy across platforms
  • Flag missing decisions

Agent as a marketing assistant, using (simulated) domain expertise and doing vital research work.

Manager support examples

The worst part of being your own boss? Your employee. 

Operations Assistant

"I have no idea what I'm doing with this."

  • Create production checklist
  • Track down platform requirements
  • Build a backward calendar
  • Scheduling or actioning vendor emails
  • Prepare handoff notes
  • Collate receipts, tax research/prep

Backlist Audit Work

"I have old books underperforming, and I do not know where to start."

  • Review metadata
  • Suggest refresh priorities
  • Create a maintenance calendar
  • Compare books by opportunity
  • Surface stale assets
  • Google Ads campaign prep

Agent as an operator: handling (or automating) structured, repetitive, and easy to postpone tasks for you.

Audience moment

Agents are powerful. You don't have to choose. But it helps to have a /goal in mind.

1

Maker

"I am stuck in the work itself."

Drafting, revision, feedback, continuity, creative decisions.
2

Marketer

"I struggle to connect the work to readers."

Blurbs, ads, newsletter, launch, positioning, visibility.
3

Manager

"I am drowning in the business and logistics."

Deadlines, platforms, contractors, files, money, operations.

Your agent should not feel like "yet another" tool.
It should allow you to personally drive fewer tools yourself.

Why this matters now

Advancements will make working with agents scary easy (and soon).

Setup friction is dropping. Custom agents will become easier to install, connect, and personalize.
Context stays durable. Your author context, workflows, boundaries, and repeatable processes remain valuable across tools. Document them!
What if readers do it? When someone asks their agent to purchase a book they'd like every week, you want it to be yours. Learn the tools, endeavor to get included in agentic recs. Knowledge is power.

Worst case you're preparing your author business so that any good human assistant could understand it and help!

The messy middle

A few agent gotchas.

Increasingly powerful models, robust harnesses… but a web that wasn't built for them.

Reliability Agents are finnicky. Hosting setups vary, sessions drop, and automations break in quiet, frustrating ways.
Autonomy risk Not sci-fi runaway AI—but agents are nondeterministic, and they usually have the password to your inbox.
Cost Top-tier models aren't cheap. An always-on agent can rack up bills the same way a new hire does.

Demo

Creating a Story (or Series) Bible

1

Give context

Upload or paste series notes, prior summaries, and the new scene.

2

Ask for a brief

"Before drafting this scene, what should I remember?"

3

Evaluate the scene

Ask it to flag contradictions, unresolved threads, and canon questions.

Working with agents

My top tips.

Invest in yourself Spend a few bucks getting the setup right. It's dramatically cheaper once the scaffolding is in place — assuming you're not going bananas.
Be permissive The more you let an agent use—your calendar, your inbox, your files—the more it can actually do. Grant access deliberately, with guardrails, but grant it.
Be specific We're in the monkey's-paw era of AI: you'll get exactly what you asked for. Make sure that's what you actually wanted.
Get ambitious Ask for more than you think it can handle. Trust the agent to push back — it will impress you if you let it.

When you get home

Run one small agent experiment.

  1. Pick an agent to experiment with.Choose the tool that best matches the bottleneck you named.
  2. Prime it with your author context.Give it your project, voice, goals, constraints, and rules.
  3. Pick one tedious task.Not "fix my career." Something specific enough to delegate.
  4. Evaluate the results.Judge usefulness, accuracy, voice fit, and whether it reduced effort.
  5. Close the loop.Tell it what worked, what missed, and how it should improve next time.

Practical advice

Use an agent to set up your agent.

Always-on assistants are powerful—but complex to configure. Don't go it alone.

First

Managed agentic coding tools

Tools like Codex and Claude Code can already handle the project work. Start there, and if you still want an always-on assistant, get help from one of these to set that up. (They can run your computer's terminal, edit files, and use managed services' CLI tools to set things up, so you don't have to.)

Then

Always-on personal agents

Systems like Hermes or OpenClaw run continuously and manage tasks on your behalf—but configuring and even maintaining one can be a serious project. Let a coding agent do all the heavy lifting.

The easiest way to get an agent set up is to ask an agent to do it.

Closing thought

Don't overthink it.

Prime the pump, point it at a problem, and iterate.

Thanks!

Questions?

Thanks!

Notes & Links

Scan to grab the slides, notes, and links from this talk.

QR code for notes and links