Skip to content

Reinforce evidence-for prohibition in retro agent definition #102

Description

@fullsend-ai-retro

What happened

The retro agent definition at agents/retro.md delegates behavioral constraints to skills, stating "See the retro-analysis skill for the proposal object schema and writing guidance." The evidence-for prohibition lives only in skills/retro-analysis/SKILL.md (line 145), buried in the "Before proposing" section after several paragraphs about duplicate checking. The agent ignores this instruction frequently enough that PR #85 needed to add a post-script gate, referencing fullsend-ai/fullsend#3881.

What could go better

The agent definition is loaded first and establishes the agent's core identity and constraints. The output rules section of agents/retro.md already has a bullet list of rules ("Write ONLY the JSON file", "The JSON must be valid", etc.) — adding the evidence-for prohibition here would place it alongside other hard output constraints rather than in a procedural section of the skill.

This is a prompt engineering improvement. There is no guarantee that moving the instruction will fix the compliance issue entirely (the post-script gate should remain regardless), but making the constraint more prominent in the agent definition could reduce the frequency of violations.

Confidence: Medium — prompt placement effects are empirical. The hypothesis is reasonable (more prominent placement → higher compliance) but the effect size is uncertain. The cost of the change is very low (one additional bullet point).

Note: GitHub API was unavailable during this retro, so I could not verify whether an open issue already covers this improvement.

Proposed change

In agents/retro.md, add a bullet to the "Output rules" section:

- Do NOT file proposals that corroborate existing issues. If your analysis produces evidence supporting an open issue, mention it in the `summary` field with the issue number — do NOT create a new proposal titled "Evidence for #XXXX" or similar. Evidence proposals create duplicate noise.

This reinforces the skill-level instruction at the agent definition level, where it is more likely to be attended to during output generation.

Validation criteria

  1. The rate of evidence-for proposals reaching the post-script gate decreases over the next 10 retro runs.
  2. The change does not introduce false negatives — legitimate proposals that happen to mention evidence or existing issue numbers continue to be filed normally.

Generated by retro agent from #85

Activity

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

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions