Skip to content

Add a pattern for invisible-context defenses #198

Description

@bhekanik

The pattern

During iterative editing, a model can carry abandoned objections and alternative framings from the drafting conversation into the final prose. The reader never saw that context, so the result sounds defensive or introduces distinctions nobody was making.

Phrases often include:

  • "This isn't mainly about X"
  • "I'm not trying to do Y"
  • "You could categorize this another way, but..."

These are not always wrong. The useful test is: would a cold reader naturally arrive at the objection without seeing the drafting conversation? If yes, engage it directly. If it exists only because it came up during brainstorming, remove the defense.

Why it seems distinct

This is adjacent to excessive hedging, signposting, and diff-anchored writing, but it has a different source and test:

  • Hedging weakens a claim.
  • Signposting announces what the prose will do.
  • Diff-anchored writing narrates a previous version.
  • Invisible-context defenses answer an objection that exists outside the published text.

Example

Before:

This isn't mainly about prompt length, and I'm not arguing that documentation doesn't matter. You could categorize the problem another way, but the issue is whether the agent can use the instruction when it acts.

After:

The issue is whether the agent can use the instruction when it acts.

A false-positive guard matters here. Some objections are obvious or well established, and answering those can strengthen an argument. The pattern should target defenses that only make sense with access to discarded drafting context.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions