From tp-sadd
Competitive multi-agent generation with meta-judge evaluation and adaptive synthesis — for high-stakes tasks where quality matters more than speed
How this skill is triggered — by the user, by Claude, or both
Slash command
/tp-sadd:do-competitively Task description and optional output pathWhen to use
When user says 'do this competitively', 'best-of-N', 'generate multiple approaches and compare', 'competing solutions', 'quality over speed'. IMMEDIATELY when user wants to generate 3+ alternative solutions and select the best through structured evaluation. DO NOT use when you want implementation with verification gates — use sadd-execute instead.
Task description and optional output pathThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
IF task is trivial or low-stakes → use simpler execution (competitive overhead not justified)
IF task is trivial or low-stakes → use simpler execution (competitive overhead not justified) IF task requires novel approach discovery → competitive generation to explore solution space IF quality matters more than speed → competitive with meta-judge specification IF three or more independent approaches exist → competitive generation with all three running IF solutions exist but quality is uncertain → meta-judge evaluation then selective synthesis
Execute tasks through competitive multi-agent generation, meta-judge evaluation, and evidence-based synthesis to produce superior results by combining the best elements from parallel implementations.
This skill implements the Generate-Critique-Synthesize (GCS) pattern for high-stakes tasks where:
Skip this skill when:
Launch 4 agents in parallel: 3 generators + 1 meta-judge.
Meta-judge produces evaluation specification YAML (rubrics, scoring criteria). This runs in parallel because it only needs the task description, not generator outputs.
Generators produce independent solutions to the same problem. Each receives identical context but explores different approaches. Solutions saved with .a.md, .b.md, .c.md suffixes.
Dispatch order: Meta-judge first (needs time to collect context), then generators.
Prompt structure for generators:
Wait for all Phase 1 agents before dispatching judges.
Launch 3 judges in parallel. Each receives:
Each judge produces:
Reports saved to .specs/reports/{solution-name}-{date}.[1|2|3].md
Critical: Never tell judges the score threshold. They must evaluate without bias.
The orchestrator (not a subagent) analyzes judge outputs.
Three strategies based on results:
| Condition | Strategy | Action |
|---|---|---|
| Unanimous vote + clear winner | SELECT_AND_POLISH | Take winning solution, apply judge feedback |
| All solutions below 3.0/5.0 | REDESIGN | Return to Phase 1 with lessons learned |
| Split decision with merit | FULL_SYNTHESIS | Launch synthesizer to combine best elements |
Parse VOTE lines from judge replies (do not read report files directly).
Only runs when FULL_SYNTHESIS strategy selected.
Launch one synthesizer agent with all solutions and evaluation reports. Agent:
Create reports directory before starting:
mkdir -p .specs/reports
Report naming: .specs/reports/{solution-name}-{YYYY-MM-DD}.[1|2|3].md
Final solution at specified output path. Candidate solutions preserved with .a.md, .b.md, .c.md suffixes for reference. Judge reports in .specs/reports/.
Reply includes strategy used, files created, and synthesis decisions table.
npx claudepluginhub git-fg/taches-principled --plugin tp-saddProvides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.