8000 Should the guides keep the bold AVOID marker? · Issue #156 · habit-hooks/habit-hooks · GitHub
[go: up one dir, main page]

Skip to content

Should the guides keep the bold AVOID marker? #156

Open

Description

@devill

Some generic guides end with a paragraph that starts with a bold AVOID: high-complexity, swallowed-exception and too-many-parameters. The core uncoached guide has the same thing as "AVOID AT ALL COST".

PR #150 turned all of them into ordinary sentences. For example:

  • now: "AVOID: merging conditions with and/or, or rewriting branches as ternaries, just to lower the score"
  • theirs: "Lower the decision count, not the metric: merged conditions and ternaries move the number while the decisions remain."

The trade is that the marker is a pattern an agent can match on. Dropping it makes the prohibition quieter, and we have no evidence either way about whether that matters.

Decide once and apply it to every guide, rather than letting the guides drift apart one PR at a time.

Activity

  1. added
    needs discussionOpen design questions to resolve before implementation
    on Aug 29, 2026
  2. devill commented on Aug 29, 2026

    @devill
    CollaboratorAuthor

    My personal experience is that spelling out what to avoid makes the reviewer agents more likely to spot cheating. For that reason I'm against this wording change.

  3. ghouston commented on Sep 4, 2026

    @ghouston
    Contributor

    the reasoning for dropping the "avoid" is to switch from "don't do this" style of prompts which agents tend to ignore, into more "do this" actionable prompt which the agent is more likely to perform. maybe a combination of the two prompt styles: "do this AND don't do this" and/or a checklist would work better.

  4. devill commented on Sep 5, 2026

    @devill
    CollaboratorAuthor

    I think what this depends on is what "mindset" the LLM is in. I agree that for AGENTS.md , skills and even initial prompts, it's good practice to avoid "AVOID" and "DON'T" style prompts. However when we are in review mode this flips. Review agents tend to justify things as "just an alternative solution" unless it is called out explicitly that they should not accept that.

    But I agree that a mixed variant containing both dos and don'ts might be even better.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    needs discussionOpen design questions to resolve before implementation

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      0