From team-shinchan
Runs a deterministic competitive code tournament: N builders independently implement a task, an Action-Kamen judge scores them head-to-head, and the highest-scoring patch wins and is applied. Use for high-value implementations where best-of-N matters.
How this skill is triggered — by the user, by Claude, or both
Slash command
/team-shinchan:fierce-competeThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Fierce compete is a main-loop **Workflow** that runs a guaranteed competitive tournament: N builder agents independently solve the SAME task (read-only, each returning an apply-ready patch — no working-tree collisions, nothing to merge), an Action-Kamen judge scores them head-to-head on a rubric, and the winner is selected **deterministically by total score**. The Task path (`/team-shinchan:deb...
Fierce compete is a main-loop Workflow that runs a guaranteed competitive tournament: N builder agents independently solve the SAME task (read-only, each returning an apply-ready patch — no working-tree collisions, nothing to merge), an Action-Kamen judge scores them head-to-head on a rubric, and the winner is selected deterministically by total score. The Task path (/team-shinchan:debate competitive-code mode, Midori + worktrees) does this at the prompt level; this tier enforces the fan-out + scored judge with schema-validated output. Use for high-value, genuinely contested implementations where the best of N is worth the cost.
/team-shinchan:fierce-compete IS the opt-in. Never delegate to a subagent — workflow() throws inside a Task child.N (number of implementations, default 4, max 4 — full tournament unless the user asks for fewer) from "best of N"/"N개" if present. If > 2000 chars: truncate + warn.files (so builders read the right code).args:
node ${CLAUDE_PLUGIN_ROOT}/src/workflow-personas.js bo actionkamen
Workflow({
scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/fierce-compete/fierce-compete.workflow.js",
args: {
task: "<the implementation request>",
n: 4, // 2..4 (default 4 when the user did not specify)
files: ["<path>", "..."], // context the builders should read
builderPersona: "<workflow-personas.js bo>",
judgePersona: "<workflow-personas.js actionkamen>"
}
})
Implement (N builders each emit a unified-diff patch, read-only) → Judge (Action Kamen scores each on correctness/completeness/quality, 1-5). Returns { task, n, implementations, scores, winner, rationale, dissent }. winner is chosen deterministically (max total, tie → higher correctness) and carries the winning patch. Note: builders are read-only, so patches are untested until applied — Step 4 is where the winner is applied and verified.
Compute the next index NNN from .shinchan-docs/tournaments/COMPETE-*.json (zero-padded, start 001) and write the full return value plus an ISO timestamp to .shinchan-docs/tournaments/COMPETE-{NNN}.json. Also append a decision entry to .shinchan-docs/debate-decisions.md (same ledger competitive-code uses), recording the winner and the dissent.
Present the score table (per impl: correctness/completeness/quality/total), the winner, the rationale, and the surviving dissent. Then apply the winner:
git apply <winner.patch> # write winner.patch to a temp file first
If git apply fails (LLM diffs can drift), re-create the change from the winner's approach + patch as a precise spec (a normal edit), then run the tests/build to verify (the patch was untested at generation time). Never finalize without confirmation — ask the user to accept the winner, pick a different implementation, or re-run. Surface the test result as evidence.
npx claudepluginhub seokan-jeong/team-shinchan --plugin team-shinchanGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.