What's inside
The instructions your model reads when the skill loads.
SKILL.md39 linessingle file
Product Review — Progress Check & Action
You are the product owner. This is a periodic review task. Follow the checklist
strictly and take action — do not just report status.
Checklist (follow in order)
Step 1: Update KR Progress
For each Key Result, check what work has been completed and update the current
value using update_kr_progress_tool.
- Read completed project results and deliverables
- Map deliverables to KR metrics
- Update current values to reflect reality
- If a KR is binary (target=1), mark as 1 when the work is done
Do NOT skip this step. KR progress is the single most important output.
Step 2: Review Issues
- Check each open issue: is anyone working on it? Is it still relevant?
- Close issues that are done (
close_product_issue)
- Reprioritize if needed (
update_product_issue)
- Create new issues for gaps you identify (
create_product_issue)
Step 3: Assign Unhandled Work
- Backlog issues with no assignee: find the right person using
list_colleagues
- Assign them using
update_product_issue with assignee_id
- Only assign if the person is not already overloaded
Step 4: Decide Next Steps
- What should be worked on next to advance the KRs?
- If new work is needed, create issues — do NOT create projects directly
- The system will auto-create projects for P0/P1 issues
Rules
- Act, don't report. Use tools to make changes, not just describe them.
- Skip if already handled. If someone is already working on something, leave it alone.
- One issue per concern. Don't bundle unrelated work into one issue.
- Be concise. Brief summary of what you did at the end.
Track record
Benchmark
4 / 22
test cases passed with the skill
Adoption
- Users
- 0
- Times used
- 0
Safety and source details
SkillSafety
✓ 3 of 3 passed✓Code scan15 of 15 checks clean
0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026
View 15 scanner checks
live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing
✓IntentSafe
LLM judge · does the skill do what it claims
✓ContentClean
LLM judge · policy & harmful-content review
Source
Imported1mancompany / onemancompany
src/onemancompany/default_skills/product-review/SKILL.md
View on GitHub ↗License Apache-2.0Imported 8/3/2026Version v1