Editing AI Writing

Most people use AI to draft emails, reports, and announcements. The writing comes out fast and mostly coherent. But then you have to read it carefully to catch where it invented a fact, softened your ask, or turned your natural voice into corporate wallpaper.

Here's what I've noticed: even when someone has genuinely good ideas—ideas that are entirely their own—if they use AI to help express them, the AI often adds filler, the sentence gets longer without getting clearer, the confidence gets more generic, and a perfectly good point gets buried under transitions like "it's important to note that" and "in today's fast-paced environment."

When I detect those patterns, I immediately dismiss the ideas underneath, which I know isn't fair—the person isn't necessarily lazy, and they're usually using AI to clarify thoughts they can't articulate well on their own, which is a reasonable thing to do. I'm someone with a sharp eye for good writing who isn't particularly skilled at writing myself, so I understand the frustration of having something to say and not being able to say it well.

But the filler still kills the message.

This matters especially because writing is what most people use AI for. So I spent time figuring out what makes writing actually good—clear sentences, solid structure, a real voice—and built that into an editor agent that improves drafts instead of replacing them. That's built into bymorning. It works with a voice profiling workflow that learns how you write: the words you use, how you structure things, the quirks that make it sound like you.

What the Editor Actually Does

The Editor is a workflow that has multiple agents check your draft before you send it.

Here's what happens when you send it a piece of writing:

1. Build the intent packet.

First, the Editor figures out what you're trying to do. It extracts who the audience is, what response you want, whether you're writing as "I" or "we," what facts must survive any rewrite, and what constraints you're working under. If something critical is missing (like who this is for or what decision it needs to drive), it asks you one focused question and stops.

2. Draft.

The Editor sends your source text and the intent packet to a humanizer writing agent, which produces a revision. It rewrites your draft to sound more like you, but it's not done yet.

3. Run five critiques in parallel.

The Editor launches five separate critique subagents at once, each reviewing the draft independently:

• Fidelity Guard: Did anything change that shouldn't have? Names, numbers, dates, commitments, legal meaning. This is a hard gate. It blocks the draft if facts were invented or altered.
• Actionability Editor: Is the main point obvious in the first two lines? Is it clear who does what by when? For decision-oriented writing, this is also a hard gate.
• Voice Auditor: Does it sound like a real person wrote it? Does it match your point of view? Is the confidence level right for the audience? This usually relies on a voice profile built by bymorning—how that works is described in a later section.
• Specificity Checker: Are claims concrete or vague? Are placeholders handled safely? It won't invent specifics, but it will flag where you need them.
• Rhythm Editor (for longer pieces): How's the cadence? Do the sections connect logically? Is there a story arc, or is it just a stack of paragraphs?

Each one either approves the draft or lists what needs fixing.

4. Iterate.

The Editor combines all five critiques into one ranked fix list. If they conflict, fidelity wins over actionability, which wins over voice, which wins over rhythm. It ignores cosmetic suggestions when they conflict with something more important.

Then it checks the gates. The Fidelity Guard must approve. The Actionability Editor must approve (or explicitly mark the content as non-action-oriented). A soft check runs the rhythm analyzer to see if anything still sounds robotic.

If a hard gate fails, the Editor sends the fix list back to the humanizer and returns to step 2. It runs this loop up to three times, or four if the problem is a fidelity repair.

If all gates pass, the Editor sends back the final text, along with anything you should still double-check. If it still can't get the gates to pass, it gives you the best draft it has and asks one specific question about what's blocking it.

Learning Your Voice

The Editor needs to know what "sounds like you" means. bymorning has a specialized agent for this called the voice profiler.

The voice profiler scans your Gmail sent mail, Slack messages, and any public writing you've published. It looks for patterns across hundreds of messages: how you open emails, how you ask for things, what words you reach for when you're frustrated, how your sentences vary in length between casual Slack messages and longer updates.

Then it builds a voice.md file that captures your voice. It includes:

• Core voice: How you structure thoughts, your typical sentence rhythm, your default formality level
• Register variants: How your voice shifts based on the medium and the audience. You probably write differently in Slack than in email, and differently to your boss than to a close coworker or a customer. The profiler captures each of these registers separately rather than averaging them into one generic style.
• Human fingerprints: The idiosyncrasies that prove a real person wrote something. Maybe you use parentheses to drop an aside mid-sentence. Maybe you start important emails with a single short sentence before expanding. Maybe you never use the word "delve" (no one does).
• Anti-patterns: What you don't do. If you never write "furthermore" or "it is important to note that," the profile records that so future editing won't add them.

The profiler also measures your natural rhythm using a lightweight analysis tool that checks sentence-length variation and vocabulary diversity. This gives the humanizer numbers to check against, not just a gut feeling.

Humanizing AI Voice

The humanizer applies your voice profile to rewrite a draft. It fixes voice issues without changing facts. It reorders paragraphs, removes filler, replaces generic phrasing with your vocabulary, and restores your natural sentence rhythm.

The humanizer runs three passes:

1. Meaning lock: It locks in what can't change, then reorders paragraphs and cuts filler.
2. Voice and specificity: Replaces generic phrasing with your actual vocabulary, restores your natural sentence rhythm, and keeps concrete details from your original draft. If your profile shows you write short declarative sentences, it doesn't let the draft expand into long compound ones.
3. Pattern sweep: Runs through a checklist of common AI tells, flags them, and replaces them with alternatives that match your voice profile.

After these three passes, the humanizer runs the analyzer to compare its work against your voice profile baselines. If the revised draft's sentence rhythm doesn't match what the profiler measured in your actual writing, the humanizer adjusts it before returning.

The Analyzer Tool

bymorning agents also have access to a lightweight analysis tool that checks for statistical patterns in writing. It looks at:

• AI vocabulary density: How many tier-1 and tier-2 words appear ("delve," "tapestry," "robust," "leverage")
• Pattern frequency: Chatbot artifacts, filler phrases, vague authority claims, generic conclusions
• Sentence rhythm: Whether sentences vary in length or fall into metronomic patterns
• Vocabulary diversity: Whether the text uses the same words repeatedly in predictable ways

When used by the Editor agent, it returns a score and flags specific issues for the humanizer to review. The humanizer uses this as guidance, but it also knows your voice profile. If the analyzer flags something as an AI pattern, but your voice profile shows you actually write that way, the humanizer keeps it. The analyzer catches when a draft strays from your natural style, not just when it sounds machine-generated.

The Result

This process dramatically improves the quality of AI-assisted writing. Instead of hoping the model got it right on the first try, you have multiple specialized agents checking the draft against specific criteria—fidelity, actionability, voice, specificity, rhythm—each bringing its own perspective to the same text.

The Editor is one of many workflows built into bymorning. Interested in creating your own custom AI workflows? Try this workflow and build your own at bymorning.ai.