product-capability
Turns PRDs, roadmap items, and product discussions into implementation-ready capability plans that surface constraints, invariants, and open decisions before multi-service engineering work begins.
Baseline already passes 86% of tasks (saturated), so pass rate can't show a difference. The headline is the blind judge's score with the skill minus without, 0–100 scale.
cost of the lift+$0.0686 per task · +3,401,286 input tokens · aux $0.007 (generated media)agent cost with the skill minus without, per task, this model — excludes SkillzBench's own judging.
- Pass-rate Lift
- −43pp [−86, −14]
- Content Lift forced arm
- −3.4pts [−12.3, +6.6]
- Trigger rate
- 100%
- Trials
- 23 · 7 tasks
- Model
- GPT-6 Luna
- Measured
- 2026-10-20
Content Lift = the skill's instructions injected into the system prompt, so it measures what the content does once loaded, apart from whether the agent chooses to load it. Method →
Embed this score · live badge for your README
The badge shows the real benchmark number and updates itself when we re-bench — including the honest states (“did not trigger”, “not benched yet”). All badge options →
[](https://skillzbench.com/skills/product-capability)[](https://skillzbench.com/skills/product-capability)<a href="https://skillzbench.com/skills/product-capability"><img src="https://skillzbench.com/api/badge/product-capability.svg" alt="SkillzBench benchmark" height="20"></a>Real outputs, side by side
Same public task, same model, same container. Left: without the skill. Right: with it. Drag the handle. These are the grader's own renders that the blind judge scored; the Scrolled views are captured by a scripted scroll pass, so scroll-driven work (reveals, pins, parallax, progress) is visible mid-flight. Held-out task outputs are never shown.
public task · productivity-automation-149tsw6r
Turn the PRD at /work/inputs/prd_audit_trail.md into an implementation-ready capability plan for the Audit Trail capability, taking the consumer notes in /work/inputs/audit_consumers.md as binding.…
- Results are for the SkillzBench reference agent on GPT-6 Luna. Switch models above — a skill can help one model and not another.
- Screen tier (1 trial per arm) is directional. Verified needs a full-tier run (≥ 8 tasks, ≥ 3 trials per arm) with the CI excluding zero.
- Task coverage: 7 generated tasks for this skill's category; a skill can shine elsewhere.