Worst-seed gallery

Two more bad conjugations

Same format as the first review: original prompt, generated fields, why they failed, and the corrected versions I’d actually want.

Example 2

project.email.update

Prompt
Draft a project update email from the status report. Audience-appropriate: executives get highlights and decisions needed; team gets task-level detail. Include: progress summary, blockers, and asks.
Generated slug
email.update.project
Generated contract
a draft project update email with separate executive and team sections, each containing a progress summary, blockers, and asks, exists in the context
Generated tags
["email", "project", "update", "executive", "team"]
Generated outcome
Draft a project update email with executive and team details
Why it failed
  • Contract independence: it just re-lists the prompt requirements and adds exists in the context.
  • Distinctiveness: email.update.project is generic and close to other update/draft/email shapes.
  • Parent-child coherence: this assumes a usable status report exists, but the parent contract was too weak to guarantee that.
What I’d change

Slug
project.update.email

Contract
a project update email tailored to its audience with current progress, blockers, and explicit asks exists

Tags
["project", "update", "email", "status", "stakeholders"]

Outcome
draft a project update email
Example 3

data.report.narrative

Prompt
Write a narrative summary of the data findings suitable for a non-technical audience. Translate statistics into plain language insights. Highlight: what's notable, what's expected, and what needs attention. Under 300 words.
Generated slug
data.narrative.summary
Generated contract
a narrative summary of the data findings, written in plain language for a non‑technical audience, under 300 words, and containing sections on notable observations, expected trends, and items needing attention exists in the context
Generated tags
["data", "summary", "narrative", "plain", "nontechnical"]
Generated outcome
Generate a plain-language summary of data findings
Why it failed
  • Contract independence: again, this is the prompt copied into contract form.
  • Navigability: an agent can see the text exists, but can't cleanly verify "plain language for a non-technical audience" without subjective evaluation.
  • Parent-child coherence: it assumes findings exist, but the parent guarantee was too weak.
What I’d change

Slug
data.findings.brief

Contract
a plain-language summary of the dataset's main findings, surprises, and risks exists

Tags
["data", "findings", "summary", "plainlanguage", "insights"]

Outcome
summarize data findings in plain language
These two fail for the same core reason as project.retro.run: the model keeps turning the prompt spec into a contract instead of turning the post-state into a contract.