Sprint eval

Sprint Eval

Panel-based sprint plan rating. Three independent expert agents read the SAME plan and score the SAME criteria, so scores are comparable and agreement is signal: anything all three flag independently is almost certainly real.

Input

A sprint plan: a file path or pasted text. If it's a file, pass the path to the agents and let each read it themselves (independent reads, no shared summary — don't pre-digest the plan for them, that would correlate their errors).

Fixed criteria (same for every rater)

1. Dependency & sequencing health — critical path soundness, blocker surfacing/ordering, cross-person handoffs.
2. Capacity realism & load balance — per-person allocation vs the sprint horizon, onboarding load, P0 distribution.
3. Goal alignment & scope discipline — must-ship items map to stated targets; cancel/backlog/carry-over hygiene; clarity.

The panel

Spawn all 3 in ONE message (parallel Agent calls, general-purpose). Each gets a distinct lens so they catch different failure modes:

Rater	Persona	Lens emphasis
EPM	Veteran engineering program manager, 15y of 2-week cycles	handoffs, single points of failure, priority inversion
OR	Operations-research scheduling specialist (CPM/PERT, resource-constrained scheduling)	chain depth vs horizon, slack, utilization vs variance buffer, negative-slack chains
Agentic	Expert in agentic task planning — crisp specs, explicit interfaces, verifiable done-conditions	hidden couplings, unspecified API contracts, vague acceptance criteria, TBD items

Every prompt must include, verbatim in spirit:

• the 3 fixed criteria, phrased through that rater's lens;
• "Return raw data only (your final message is consumed by a program, not a human): for each criterion give score: N/10 plus 2-3 sentences of justification grounded in specifics from the plan, then one line top risk: with the single biggest thing you'd flag.";
• "Be a tough but fair grader; don't inflate."

Domain swap: if the sprint is clearly outside software (e.g. content, ops), swap the EPM persona for the matching domain planner; keep OR and Agentic — they're domain-agnostic.

Synthesis (your final answer)

Lead with the verdict: overall average to one decimal, e.g. "≈ 4.7/10".

1. Score table — criterion rows × rater columns + per-criterion average.
2. Unanimous findings first — anything 2-3 raters flagged independently, named as such ("all three named X"). The intersection is the headline, not your own opinion.
3. Lens-unique catches — findings only one expert saw (typically the Agentic rater's interface/spec gaps), attributed to the lens.
4. What scored well — 2-4 genuine strengths; a panel that only criticizes is uncalibrated.
5. Top 2 fixes — concrete, smallest-change recommendations ("shed ~4 of X's items to Y", "ticket the 7 missing integrations or rewrite the target"), not generic advice.

Do not average away disagreement: if raters split by ≥3 points on a criterion, say why (usually the OR rater counting days where others pattern-matched).

Sharp edges

• Scores come from the agents, never from you; if an agent fails or returns no scores, rerun it rather than filling in numbers.
• Keep raters blind to each other — no sharing of outputs between them, no "rater 2 said...".
• Ground every claim in the plan's specifics (names, ticket ids, counts). A finding that can't cite the plan gets dropped from the synthesis.