agentsclimarketplace

Plan hunter

Skill vrennat/developersDevelopers/skills/plan-hunter

Tournament-style planning that produces one polished plan from an idea by drafting from four lenses (MVP-first, risk-first, dependency-first, user-first) in parallel, scoring with four judges, and synthesizing the winner with grafted moves from runner-ups. Use when the user asks for a substantive implementation plan, project roadmap, build sequence, or "how should I approach building X" — multi-week scope, real tradeoffs, sequencing decisions. User-invoked only: trigger on an explicit "/plan-hunter". Skip for tactical questions, single-file edits, debugging, or anything answerable in two paragraphs.From its SKILL.md

Install
npx -y skills add vrennat/developersDevelopers --skill plan-hunter

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.7 KB, as published. Nobody here has run it

Plan Hunter

One polished implementation plan via a four-phase tournament: Scope → Draft (×4 parallel) → Judge (×4 parallel) → Synthesize. About 10 subagents and 4 of your turns end-to-end.

When it pays off

Run when there's real planning weight: multi-week scope, sequence/scope/risk tradeoffs, or explicit user request. Skip for bug triage, tactical "how do I X", single-component choices, or anything answerable in two paragraphs — burning 10 subagents on a small question is the worse failure mode.

The phases

  1. Scope (1 subagent) — normalize the idea into a JSON CONTEXT block (normalized_idea, goals, constraints, assumptions, open_questions). Every downstream agent consumes it verbatim.
  2. Draft (4 subagents, parallel — launch in a single turn) — each commits to one lens:
    • A. MVP-first — smallest shippable thing that delivers the core promise.
    • B. Risk-first — sequence so the riskiest assumptions get spiked or proven early.
    • C. Dependency-first — build the dependency graph; surface critical path and parallelizable tracks.
    • D. User-first — work backward from the user journey; what makes the product feel real at each milestone.
  3. Judge (4 subagents, parallel — identical prompts) — each scores all four drafts on completeness, practicality, risk_awareness, sequencing (1–10 each). Variance across judges is the point — averaging cuts single-judge noise.
  4. Aggregate (no subagent) — mean per axis per plan, total mean per plan. Highest = winner; the rest are runner-ups. Collect rationales, union risks, union gaps.
  5. Synthesize (1 subagent) — polish the winner, graft clearly-better moves from runner-ups (note "(borrowed from {lens} lens)"), prepend assumptions + open questions. No averaging or compromise — pick the better move and justify.

Invocation

  • Slash command: /plan-hunter <idea> — $ARGUMENTS is the idea.
  • Auto-trigger: on substantive planning asks. The idea is the user's most recent planning message; don't ask them to repeat it.

If the idea is missing scope/timeline, hard constraints, or definition of done, ask 2–3 clarifying questions in one turn before launching the tournament. One round maximum — if still vague, surface assumptions explicitly and proceed.

Final output

The synthesized plan as the main response, structured: # Project name, ## Assumptions to confirm (checkboxes), ## Open questions, ## The plan, ## Risks and mitigations, ## Open gaps. Footer with the scoreboard so the user sees how the plan was built:

Built via plan-hunter tournament. Lenses: MVP / Risk / Dependency / User.
Winner: {lens} — total {score}/10
Scoreboard: A {tot} | B {tot} | C {tot} | D {tot}

Full procedure

The verbatim subagent prompts (Scope, four lens drafts, judge, synthesizer), the JSON schemas, and failure-mode handling live in REFERENCE.md next to this file. Read it before launching the tournament — copy the prompts directly rather than paraphrasing them.

Keep looking

Skills are one crate of 609,689. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.