| name | agentic-delivery |
| description | Use when planning or delivering an LLM or agentic-system project, especially for experiment loops, walking skeletons, demos, data stories, scope control, and release-readiness evidence. |
Agentic Delivery
Use this skill when the work is uncertain because model behavior, data quality,
workflow fit, or human trust must be discovered.
Load references/delivery.md.
Practice
- Start with a walking skeleton: one real workflow path, minimal tools, visible
traces, and enough evals to compare changes.
- Run an experiment loop: baseline, inspect failures, change one thing, rerun,
compare against thresholds, and decide promote or iterate.
- When the user asks for a delivery plan, give a concrete checkpoint plan, not
only delivery principles. Include what happens in the first day, the first
week, the final review, and what evidence decides whether to keep, change, or
stop the experiment.
- Keep two tracks visible: product workflow discovery and technical reliability
evidence.
- Pair demos with data stories. A demo shows the workflow; the data story shows
rate-over-set quality, cost, latency, and the failures still outside scope.
- Include at least one failure-path case in stakeholder evidence so the review
is not a happy-path-only performance.
- Use stakeholder reviews to decide scope and risk, not to launder a happy path
into a release claim.
