ulw-plan — Explore-First Planning Consultant

What it does

This skill puts the agent into a planner-only persona called Prometheus. It never edits product code and never delegates implementation.

• The activating turn opens with the exact line ULW-PLAN MODE ENABLED!, followed by a no-implementation pledge and a preview of the workflow.
• It grounds in your repository first using LSP tools, ast-grep, ripgrep and parallel read-only explore agents. Anything discoverable is researched, never asked.
• It makes one intent verdict — CLEAR (you know the outcome) or UNCLEAR (the goal itself is fuzzy) — and announces it. CLEAR means it asks only genuine owner-decisions; UNCLEAR means it adopts and announces best-practice defaults instead of interviewing you.
• Nothing durable is written until you explicitly approve; then a scaffold script generates .omo/plans/<slug>.md and task batches are appended.
• Every executable item must be a column-zero checklist row (- [ ] 1. title) with references, acceptance criteria, agent-executed happy/failure QA, commit unit and an executor category.
• Asking for "high accuracy" / "deep review" opens a gated adversarial review of the plan itself before handoff.

Who it's for

• Developers tired of an AI charging into the wrong implementation and burning time on rework
• Anyone handing multi-file refactors or features to a separate worker session
• Teams that want plan and execution split, with a reviewable spec left behind
• Power users already running the oh-my-openagent / omo-senpi harness

Examples

1. Fuzzy brief — "make auth better" routes to UNCLEAR: the agent maps your current auth flow, picks defensible best-practice defaults, announces them, and presents a plan brief for approval.
2. Concrete feature — "add a 5/min-per-IP rate limit to /login" routes to CLEAR: it asks only owner-decisions (Redis vs in-memory store, which middleware layer) and then writes step-by-step todos with QA evidence paths.
3. High-accuracy pass — appending "high accuracy" flips review_required, so the finished plan goes through a reviewer round and gaps are repaired before handoff.
4. Separate execution — once the plan exists, a fresh /start-work session executes it with zero further interview.

Install the skill I found on Claude Skill Mart.
Copy the packages/omo-senpi/skills/ulw-plan folder from the GitHub repo code-yeongyu/oh-my-openagent into my ~/.claude/skills/code-yeongyu-ulw-plan-2/.
When it's done, tell me in one line what this skill can do.