| name | recommend-agentic-product-improvements |
| description | Recommend codebase-grounded improvements for agentic capability findings. |
Recommend Agentic Product Improvements
Turn review findings into concrete improvement recommendations grounded in the repository.
Read the catalog for experience and surface, the review for findings, the code for improvements.
When To Use
Use this skill when the user asks to:
- Recommend improvements for agentic review findings
- Plan changes grounded in the codebase
- Turn
agentic-product-review/reviews/agent-chat.md into actionable engineering guidance
- Propose improvements for one capability or finding
This skill is Phase 3 of agentic-product-review. It can also be run standalone after review.
If no review exists, ask the user to run review-agentic-capabilities first or provide review files.
Inputs
Required:
agentic-product-review/agentic-capabilities.md — capability slugs, implementation pointers, audience
Review:
agentic-product-review/reviews/agent-chat.md — one file per capability from review-agentic-capabilities
Optional scope — pass a capability slug as a positional arg:
agent-chat — one capability only- A specific finding slug the user names (e.g.
coarse-multi-tool-progress) or passes with the capability
What To Ground
For each finding, read enough code to propose a realistic improvement:
- Implementation pointers, experience, and surface from the catalog entry
- Findings from the matching review file
- Routes, components, hooks, jobs, and tests related to the capability
- Similar patterns already used in the repo (loading states, empty states, approval flows, error handling)
Use capability as the stable unit. Read catalog and review for context; write forward-looking recommendations only — do not restate findings.
Improvement Process
- Load
agentic-product-review/memory/project-context.md if it exists (format: review-agentic-capabilities/references/project-context-format.md).
- Load
agentic-product-review/agentic-capabilities.md and the relevant review file(s) from agentic-product-review/reviews/.
- Select findings to address:
- All confirmed findings in Findings if the user did not specify a scope
- One capability if the user passes
agent-chat or names a capability slug
- One finding if the user passes or names a finding slug (e.g.
coarse-multi-tool-progress)
- Skip Open questions — when the user answers one, follow Recording open question answers in
review-agentic-capabilities first
- For each selected finding, read the code and find patterns to reuse.
- Write one file per capability at
agentic-product-review/recommendations/<capability-slug>.md using references/output-format.md. Create the folder if needed. Per finding: Finding ref, then Product (Recommendation only), then Tech (Implementation and files). Do not restate review or catalog copy. End every file with Next step, horizontal rule, and footer.
- Tell the user how many recommendation files were written.
Do not write code unless the user explicitly asks to implement.
Re-Run
Pass a capability slug as a positional arg:
/recommend-agentic-product-improvements agent-chat
Re-read the code, update only agentic-product-review/recommendations/<capability-slug>.md, and note which findings are addressed and what remains.
What This Skill Covers
- Forward-looking Recommendation (product) and Implementation (tech) per review finding
- Finding ref as a pointer — not a copy of the review
- File paths, patterns to follow, acceptance criteria, effort, and risk
What This Skill Does Not Cover
- Restating findings, problems, impact, or catalog experience/surface
- Implementing code unless the user explicitly asks
Next Step
Tell the user:
Wrote N recommendation files to agentic-product-review/recommendations/.
Next step: pick a recommendation to implement, or re-run review-agentic-capabilities after changes.
