| name | review-agentic-capabilities |
| description | Review each conversational agentic capability from a discovery catalog and write per-capability findings in Markdown. |
Review Agentic Capabilities
Review every conversational capability in the discovery catalog from a product and user-experience perspective.
Conversational means the user talks to the agent directly — chat panel, copilot, or assistant UI. Skip background/headless capabilities (queue, status, results only).
Read the code to understand behavior, but write findings in plain language — no file paths or implementation detail in the output.
Requires a discovery catalog. Each review file references its catalog entry.
When To Use
Use this skill when the user asks to:
- Review discovered agentic capabilities
- Audit each agent flow in the repo
- Evaluate agents after a discovery pass
- Produce per-capability findings from the capabilities catalog
If no catalog exists, ask the user to run discover-agentic-capabilities first or provide a catalog path.
Input
Default input: agentic-product-review/agentic-capabilities.md
Read each catalog entry by capability slug (e.g. agent-chat) and the implementation pointers listed for it.
Rubric
Apply references/review-rubric-conversational.md for every capability in scope. Set Assessment using references/assessment-criteria.md.
Review Process
- Load
agentic-product-review/memory/project-context.md if it exists.
- Load the capabilities catalog.
- Review each conversational capability one by one — skip headless/background entries; do not skip in-scope entries unless the user passes a capability ID.
- Read enough code to understand what people actually experience; apply the rubric.
- Write one file per capability at
agentic-product-review/reviews/<capability-slug>.md (e.g. agent-chat.md) using references/output-format.md. Create the folder if needed. Per file: capability name as title, intro sentence, then What works, Findings, Open questions, Notes — each a top-level ## section. End every file with Next step, horizontal rule, and footer.
- Tell the user how many review files were written.
If the catalog is empty or stale, say so and suggest re-running discovery.
Recording open question answers
When the user answers an open question from a review — in conversation, without naming a workflow — update memory inline. Do not describe this as a separate skill or command.
Required inputs:
- The user's answer (in the message or a follow-up)
- Which open question — capability slug and question slug (e.g.
agent-chat / empty-state-intentional)
Process:
- Load
agentic-product-review/memory/project-context.md. Create agentic-product-review/memory/ and the starter file from references/project-context-format.md if missing.
- Load the matching review:
agentic-product-review/reviews/<capability-slug>.md. Locate the open question by slug under Open questions.
- Update
agentic-product-review/memory/project-context.md — add or extend the capability section with a short distilled bullet from the answer.
- Remove the answered entry from Open questions in the review file. Re-read the review; if the answer confirms a real UX problem, add a Findings entry instead. If it confirms intentional design, leave it out of findings.
- Tell the user which files were updated (paths only — no workflow jargon).
Example user message:
Recorded answer for agent-chat/empty-state-intentional.
Updated:
- agentic-product-review/memory/project-context.md
- agentic-product-review/reviews/agent-chat.md
Focused Re-Review
Pass a capability ID as a positional arg:
/review-agentic-capabilities agent-chat
Update only agentic-product-review/reviews/<capability-slug>.md. Note what improved, what remains, and any new risks.
Do not implement code unless the user explicitly asks.
What This Skill Covers
- Whether each capability works well for its users
- What works — what people can rely on today
- Findings — confirmed UX problems with stable slugs (e.g.
no-first-run-guidance), what was observed, and impact (no recommendations)
- Open questions — behaviors that may be intentional, each with a slug ID, What we observed, and Question
- Assessment — strong, acceptable, weak, or incomplete (on the intro line)
- Review-specific caveats in Notes
What This Skill Does Not Cover
- Cataloging what exists — that is
discover-agentic-capabilities
- Repeating Capability, Experience, Surface, or User persona — the intro sentence and catalog cover identity
- Structural facts already in the catalog (HITL, tools, scope)
- Recording open question answers — see Recording open question answers above
Next Step
Tell the user:
Reviewed N capabilities. Output written to agentic-product-review/reviews/.
Next step: run `agentic-product-review` for the full workflow, or `recommend-agentic-product-improvements` to turn findings into codebase-grounded recommendations.
