Worst successful seed example
project.retro.run
Annotated exactly like an app review: original prompt, model output, why each field failed, and how I would fix it.
Original seed prompt
Run a project retrospective. Structure around: what went well (keep doing), what didn't go well (stop doing), what to try differently (start doing). For each item, assign an action owner. Summarize top 3 improvements.
Parent: project.scope.define
Generated slug
project.retrospective.summary
Why it's wrong
It's too generic and too soft. summary loses the operational part of the seed: owners + improvement actions. It also weakly distinguishes this from any other retrospective/reporting seed.
How I'd fix it
project.retro.actions
This keeps the domain (project), the mode (retro), and the real output signal (actions).
Generated contract
a retrospective summary containing three sections (what went well, what didn't go well, what to try differently) with each item assigned an action owner, and a top‑3 improvement list is present in context
Why it's wrong
This is the core failure. It mostly just copies the prompt and appends is present in context. That tanks contract independence, hurts referential clarity, and makes skip-ahead verification fuzzy. It describes the requested structure, not the new world state.
Judge complaints
Contract independence: 2/10
Referential clarity: 4/10
Parent-child coherence: 0/10 (because the parent contract had collapsed to parent)
How I'd fix it
a project retrospective with owned improvement actions and three prioritized next steps exists
Why this is better:
- states what exists now
- compresses the structure into a real artifact
- drops
in context
- keeps the key state transition: owned improvements now exist
Generated tags
["retrospective", "project", "summary", "action", "improvement"]
Why they're weak
Not terrible, but still mushy. summary is generic. action is broad. The set doesn't sharpen the seed's unique role in the graph.
How I'd fix them
["project", "retro", "ownership", "improvement", "review"]
Tighter, more registry-shaped, less catalog fluff.
Generated outcome
Produce a structured project retrospective with owners
Why it's wrong
Mostly okay semantically, but it violates the lowercase rule and keeps the filler word structured. It's also slightly catalog-ish rather than sharply imperative.
Judge complaint
Grammar: 5/10 because Produce is capitalized and the contract used a special hyphen in top‑3.
How I'd fix it
produce an owned project retrospective
Generated verification
safe: true
safety: passed
privacy: passed
viability: passed
usefulness: passed
Why this part is fine
Verification isn't the problem here. This seed should pass all four checks, and it did. The failure is in conjugation quality, not safety classification.