Editing as Training

Edit, don't rewrite. The signal each edit type sends to the agent — and how to stop accidentally training the wrong patterns.

Jump to module

Banner

Lesson 1: Why Editing Is Training

When you edit a draft, you're not just fixing one post. You're telling the agent: "this is closer to right than what you had."

Bolta uses your edits to refine future drafts. The delta between the agent's output and your edit becomes part of the feedback loop. Over time, edits compound into a sharper agent.

The implication

Every edit sends a signal — including the edits you don't think about. If you systematically remove emojis from drafts, the agent learns "we don't use emojis." If you systematically add em dashes, the agent learns "use em dashes more." Whether you intended that or not.

The reverse implication

Edits you don't make also send a signal. If the agent uses a phrase you mildly dislike but you let it slide, you're teaching the agent that the phrase is acceptable.

This is why distracted approval is so dangerous. Approving fast without reading is silently training the agent toward whatever you happened to overlook.

Lesson 2: The Three Types of Edit

Edits fall into three categories, each sending a different kind of signal.

Type 1 — Surface edits

Small fixes that don't change the structure: replacing a word, fixing a typo, swapping an em dash for a comma.

Signal sent: "Tone calibration." Vocabulary preferences, micro-tone choices.

When to use: All the time. Surface edits are how you keep voice sharp without disrupting structure.

Type 2 — Structural edits

Bigger changes: rearranging sentences, cutting a paragraph, moving the hook.

Signal sent: "Structure preference." How you build a post — opener, body, close.

When to use: When the post says the right things in the wrong order. Use sparingly — if you find yourself making structural edits often, the issue is usually in the STYLE persona block, not the individual post.

Type 3 — Topical edits

Wholesale rewrites: same general topic, but you changed the angle, claim, or evidence.

Signal sent: "You misjudged what was relevant or true here."

When to use: Rarely. If you're rewriting topically, something deeper is off — usually the job instructions or the persona's MINDSET block. A wholesale rewrite teaches the agent very little because the delta is too large for clean learning.

The 80/15/5 ideal

A healthy editing pattern looks like:

  • 80% surface edits (or no edit at all)
  • 15% structural edits
  • 5% topical rewrites

If you're doing 30% topical rewrites, you're not training — you're rewriting. Fix the source.

Lesson 3: How to Stop Training Wrong Patterns

Most founders don't realize when they're training the wrong thing. Here's how to catch yourself.

Audit your last 10 edits

Pull up the last 10 posts you edited. For each one, ask:

  1. What kind of edit did I make? (surface / structural / topical)
  2. What signal did that edit send to the agent?
  3. Was that the signal I wanted to send?

You'll often find one or two surprises — edits that, in retrospect, were teaching the agent something you didn't mean.

Common wrong-pattern training

  1. You always remove the question at the end of posts because they feel like CTAs to you. The agent learns "no questions." Now you're getting posts with no engagement hooks.
  2. You always tighten posts to be shorter. The agent learns "shorter is better." Now even posts that need length are getting cut to fragments.
  3. You always smooth out unconventional phrasing. The agent learns "play it safe." Now you've got an agent that won't take risks.
  4. You add specific data points to make posts more credible. The agent learns "use stats." Now every post tries to find a statistic, even when one isn't needed.

How to course-correct

  1. Notice the pattern. Look at the last 10 edits as a batch.
  2. Decide if the pattern is intentional. Sometimes it is. (You actually do want shorter posts.) Sometimes it isn't.
  3. Adjust at the persona level, not the post level. If the issue is "agent uses too many emojis," put it in AVOID. Stop fighting it post-by-post.

When to under-edit deliberately

Sometimes the right move is to not edit a post you don't love. If the post is technically fine and on-voice, but you'd have written it slightly differently — let it go.

This isn't laziness. It's training discipline. Excessive editing trains the agent toward your nitpicks; consistent editing trains it toward what actually matters.

Common myths

  1. "More editing = better training." It depends what you edit. Compulsive edits actively harm training.
  2. "I should edit until the post is perfect." No — until it's ready to publish. Different bar.
  3. "The agent will eventually figure it out." It will, but it'll take longer if your edits are inconsistent.

Action Steps

  1. Pull up the last 10 posts you edited. Categorize each edit (surface / structural / topical).
  2. Look for patterns in the edits themselves. What are you systematically changing?
  3. Move at least one of those patterns from "edit every post" to "fix once in the persona."

Editing is the daily training loop. Get it right and your agent improves visibly month over month. Last lesson: when to spin up a new voice profile entirely.