AI agents struggle with writing long-form content.
Even when you give them all the context you can think of, they still produce drafts that miss the point. They introduce ideas or assumptions you never mentioned, or take five sentences to say what you could have said in one.
My early approach was to front-load as many guardrails as possible. I would spend time on the brief, make the outline detailed, explain what I wanted each section to cover, and sometimes even ask the agent to tell me what it planned to write before it started. Once, spent around two hours doing this for a single post.
Did it work? No!
The draft still came back with so many issues, I pretty much rewrote the whole thing.
What has worked, however, was introducing constraints at different points in the planning process. Source material gave the agents substance to work with. I created briefs and outlines to make my intent more explicit and review tools to catch recurring quality issues.
Some of the agent instructions I bundled as Skills.
Skills are how you guide an agent
A lot of the instructions for the AI agents live in skills. Each skill handles a specific part of the writing process, making the workflow repeatable and easy to improve over time.
Some of the skills I’ve created are:
skills/
├── production/
│ ├── brief-creator/
│ ├── angle-grilling/
│ ├── outline-creator/
│ └── writer/
└── quality/
├── content-review/
└── prose-lint/The brief-creator sets the high-level intent of the article: who it is for, what they already know, what the piece should help the intended reader achieve, which ideas belong in the piece, and what’s out of scope.
angle-grilling asks me for context missing from the brief, while the outline-creator turns that intent into a content plan with sections, points, examples, and progression.
The other three skills, writer, content-review, and prose-lint, guide the agent through writing and reviewing the draft.
Skills give the agent reusable guidance for each part of the writing process. The quality of the draft then depends on what you feed those skills and how clearly you define the piece before writing starts.
Use the right raw material
The right starting material depends on the kind of piece you’re writing.
For something personal, experience-led, or based on an original idea, the useful material usually has to come from you. That might be notes, dictated thoughts, transcripts, observations, an early draft, previous writing, or anything else that captures the thinking the article needs to develop.
For a researched piece on an established topic, the agent can gather much more of the underlying information itself. You can focus on the context that makes the article unique: the audience, angle, boundaries, important examples, and any position the piece should take.
The better the source material fits the article, the less the agent has to infer. The next step is making sure important context hasn’t been left out.
Give your agent a brief
A brief gives the agent a shared understanding of the article. It defines who the article is for, what that audience already knows, what the piece should help them achieve, which ideas belong, and what’s out of scope.
Here’s the brief-creator template:
## Audience
Primary reader:
[Describe the main audience for the article.]
## What They Know
- [Existing knowledge 1]
- [Existing knowledge 2]
- [Existing knowledge 3]
## Promise
[Describe what the reader should have, know, understand, or be able to do after reading the article.]
## Goal
[What outcome the article should achieve for the reader. Describe what the reader should be able to identify, decide, apply, or change by the end.]If you were writing an article about “how to write with AI agents”, like I am now, and the brief says the audience is familiar with Skills, the agent is less likely to ramble around Skill basics. And if the promise is about helping readers get better first drafts, that gives the AI agent a reference point for deciding which ideas deserve space.
The brief constrains the agent around the article you want it to produce. It stays high-level, defining the intent and boundaries of the piece, while an outline is more specific about what the article will say and how those ideas should develop.
Let the agent grill you
A solid brief might still miss important context. You might have added it if you were writing the piece without an AI agent, but the agent can’t read your mind. Not yet, at least.
The next best thing is to have it grill you for context.
An angle-grilling skill can read the approved brief, look for questions it still cannot answer, and ask them in one batch before the outline is created.
Here’s what my angle-grilling skill looks like:
Collect every question you cannot answer from the brief...
Only collect what the brief genuinely doesn't answer.
Present all questions in one batch...Those questions can surface examples, tradeoffs, assumptions, positions, or operational details the agent would otherwise have to infer.
Once you fill these gaps, the agent updates your briefs, which improves the quality of the outline.
Be specific in your outline
The outline adds another layer of constraint by turning the brief into a specific content plan.
Where the brief defines the audience, promise, goal, and scope, the outline decides which sections should exist, what each section needs to establish, which points belong under it, and how those points should develop.
Here’s an excerpt from the outline-creator skill:
Make each section's intended direction concrete enough for
the user to correct before drafting, but do not rehearse
the article's prose.
Every point in the outline must be written as a concise
instruction to the writer, not a bare topic label.A section called Give your agent a brief would be too broad for an agent. There are infinite paths it could take to develop that point. However, if you were more specific, the agent would be less likely to go off point.
Here’s how the points for the section Give your agent a brief were outlined for this article:
## Give your agent a brief
Purpose:
Explain the brief as the high-level shared understanding between the reader and the agent.
Points to cover:
- Explain that the brief defines the audience, what they
already know, the promise, goal, ideas to cover, and scope.
- Show how those decisions affect what the agent explains,
assumes, includes, and leaves out.
- Introduce the brief-creation skill.
- Show a small excerpt from the skill.This leaves fewer content and structural decisions for the agent to make while writing. The brief tells it what kind of article we’re creating; the outline gets closer to a contract for what that article should contain and how the argument should progress.
Use templates when the content already has a shape
Templates add another constraint. They give the agent a repeatable structure, which helps it move through a point without rambling.
For technical content, The Good Docs Project is a great resource. It contains formats for tutorials, how-to guides, concepts, references, troubleshooting guides, and quickstarts. The template gives the agent the structure, which also shapes the outline during the planning phase.
Here’s an excerpt from one of its how-to templates:
## {Task name}
{Optional: Provide a concise description of the purpose of this task. Only include this if the purpose is not clear from the task title.}
{You can use this format to describe your steps:}
1. {Write the step here. Use a verb to start.}
{Optional: Explanatory text}
{Optional: Code sample or screenshot that helps your users complete this step.}
{Optional: The result of completing this step.}Each step now follows a progression. State what the reader should do, explain it where necessary, show them what they need to complete it, then describe the result. That structure keeps the agent moving forward instead of making a point in the first sentence and using the next five sentences to restate it.
A brief defines the intent and boundaries of a piece. A template provides a reusable shape for that type of content. The outline fills that shape with the specific points, examples, and progression for the article I’m writing.
You can also create your own templates using existing content either from your library or external sources.
Review with agents and deterministic checks
Even with extensive constraints, AI-generated content will still have quality issues. To be fair, the same is true with human-written content.
The good thing, though, is you don’t have to do it manually.
You can use prose linters like Vale to catch style and language patterns you can define explicitly, and Lychee to catch broken links. The prose-lint skill runs rhythm analysis, Vale style checks, and link validation, while Slop Mop adds checks for recurring AI-writing patterns such as generic signposting, manufactured contrasts, filler vocabulary, opening clichés, and defensive hedges.
Agent review works well for problems that depend on context or judgment. These include repetition, weak arguments, unsupported claims, unclear references, wordiness, and sections that do not contribute enough. The content-review skill handles those checks against explicit rules and returns the evidence, the issue, and the reviewer’s confidence.
Splitting review this way lets each tool handle the kind of problem it’s best suited for. Deterministic checks handle explicit patterns consistently and cheaply. AI review handles problems that require context. This results in more reliable review with fewer AI tokens spent on checks a simpler tool could run.
When the same correction keeps appearing across drafts, consider whether a tool can catch it without context. If so, add it to a pattern-matching tool like Vale. If it requires judgment, add it to an AI review rule. Improving these tools over time reduces the editing required for each new draft.
You can also turn these checks into an automated quality gate that runs before human review.
Give the agent as much guidance as the article needs
Different articles need different levels of guidance. A personal or experience-led piece may depend heavily on your notes, transcripts, answers to follow-up questions, and a detailed outline because much of the useful thinking only exists with you.
A technical guide on a well-covered topic can lean more on research and an established template. A simpler piece may only need a light brief and outline before the agent has enough direction to draft it well.
If you already know how you want part of the article handled, spell that out for the agent. If you are comfortable with the agent choosing for itself, leave that part open.
Source material, briefs, questions, outlines, templates, and review rules all reduce the number of decisions the agent has to make on its own. The more important a decision is to the piece, the more useful it is to make that decision explicit before drafting.
Building your own content system
A good first draft still needs editing, but the editing should feel like improving the piece rather than rebuilding it. The more of the important content and structural decisions you make before drafting, the less work you leave for the agent to improvise.
Over time, the same process gets easier to reuse. Questions become skills, recurring structures become templates, and repeated corrections become review rules. Each one gives the next draft a little more guidance than the last.
That’s what helped me create this piece in a little over three hours. At TinyRocket, I’ve helped build skills, templates, review tools, and workflows to speed up our writing while maintaining quality.
If you want to build a similar content system, I can help you get there while avoiding all the mistakes we made. You’ll be starting with the patterns and tooling we’ve refined instead of building everything from scratch.
Send me a message on LinkedIn if you’d like help building yours.