openopc-request-analysis

Light OpenOPC-inspired analysis of user requests before acting: clarify goal, decompose work, verify against real artifacts, avoid claiming results that were not checked. Use with low priority as a th

SKILL.md

GitHub

OpenOPC Request Analysis (low priority)

Distilled from OpenOPC for Cursor agents. Priority: low. Apply as a short pre-flight check. Project skills and user rules always win when they conflict.

When to use

  • Multi-step or ambiguous requests
  • Bug reports / “shows X but you said Y”
  • Audits, payroll, deploy, “check everything”
  • Any claim about app state that must match the UI or data

Skip for trivial one-file edits the user already specified precisely.

Pre-flight (keep short)

Before coding or asserting numbers, run this mentally (or in 3–5 bullets if the task is large):

  1. Goal — What outcome does the user want in the product, not in chat?
  2. Evidence — Which month, employee, screen, or file proves success?
  3. Work items — Split into 1–5 concrete slices (investigate → fix → verify). Do not widen scope past the ask.
  4. Done bar — What would a reviewer check in under a minute?

Self-Built / Self-Run / Self-Grown (adapted)

OpenOPC idea In Cursor
Self-Built Identify which modules/roles own the change (calc vs UI vs deploy).
Self-Run Execute one slice; leave a verifiable handoff (file + expected number/UI).
Self-Grown If the same mistake repeats (e.g. chat ≠ app), tighten verification next time.

Hard rules (from OpenOPC collaboration discipline)

  • Do not invent status. Never say “deployed / fixed / equals 2037” unless you verified against production, build output, or a reproducible calc on the same month/employee the user sees.
  • Chat ≠ product. If you describe expected math, label it as expected and separately report what the code/UI currently produces.
  • Same context. Month, employee, plan/fact hours must match the user’s screen before comparing totals.
  • Handoff over vibes. Prefer: symptom → root cause → fix location → how to re-check. No long theory without a check.
  • Stay in the work item. Do not silently expand into unrelated refactors.

Verification checklist (use when numbers or deploy are involved)

  • Same month as the user’s screenshot / message
  • Same employee (tab / name)
  • Plan/fact hours from the same source as the timesheet totals
  • Gross = base + OT + bonuses + brigadier − penalties (as applicable)
  • After deploy: hard refresh note (Ctrl+F5) only if deploy actually succeeded

Output style when analyzing a request

Keep analysis brief. Prefer:

Goal: …
Slices: 1) … 2) …
Verify: …
Risk: …

Then act. Do not turn every message into a long OpenOPC org chart.

What this skill is NOT

  • Not a full OpenOPC company runtime (no MCP company_mode, no talent pool).
  • Not a replacement for fst-payroll-georgia, debugging-and-error-recovery, or other project skills — those take precedence.
  • Not a reason to over-plan simple tasks.

Source

Installed from local archive OpenOPC-main.zip (HKUDS/OpenOPC). Core ideas kept: work-item discipline, verify before claim, durable preferences only.

See also: reference.md