mishafyi How to Win
3 rounds. 3 agents per round. Same angles every round. Each round goes deeper.
Research skill that figures out what winning means for a specific situation. Deploys 3 research agents (Codebase Archaeologist, Domain Scout, Adversary) across 3 rounds. Each round reads previous findings and asks: "what did you miss?"
Model requirement: Use model: "opus" when dispatching researcher agents.
The Loop
ROUND 1: Deploy 3 researchers (Archaeologist, Scout, Adversary)
→ Combine findings
ROUND 2: Deploy same 3 researchers, they read Round 1 findings
→ "Here's what was found. What was missed? Go deeper."
ROUND 3: Deploy same 3 researchers, they read Rounds 1+2 findings
→ "Final pass. What's still missing?"
→ Synthesize all 9 reports into WINNING BRIEF
Fixed at 3 rounds. Same 3 angles every round. Each round builds on the last.
The Three Angles
| Agent | Angle | Focus | Key Question |
|---|---|---|---|
| A | Codebase Archaeologist | Code, git history, tests, dependencies, patterns, tech debt | "What does the code say?" |
| B | Domain Scout | Web search, docs, best practices, benchmarks, industry standards | "What does the world say?" |
| C | Adversary | Failure modes, risks, hidden constraints, breaking changes | "What will go wrong?" |
History File
All findings are persisted in .claude/how-to-win-history.local.md. This file:
- Is created at Round 1 (empty)
- Is appended after each round with all 3 agents' findings
- Is read by every agent at the start of their work
- Persists across sessions
Format:
---
question: [the research question]
rounds_completed: [N]
---
## Round 1
- Archaeologist: [key findings summary]
- Scout: [key findings summary]
- Adversary: [key findings summary]
## Round 2
- Archaeologist: [new findings, what was missed]
- Scout: [new findings, what was missed]
- Adversary: [new findings, what was missed]
## Round 3
...
Round 1 — First Research Pass
- Create
.claude/how-to-win-history.local.mdwith question in frontmatter, empty body - Dispatch 3 researcher agents with
run_in_background: true,model: "opus":
Agent A brief:
RESEARCH BRIEF
==============
QUESTION: [user's question]
YOUR ANGLE: Codebase Archaeologist
ROUND: 1
Investigate from the CODEBASE angle. Dig into the actual code, git history,
tests, dependencies, and patterns. Answer: "What does the code say about this?"
HISTORY FILE: .claude/how-to-win-history.local.md
Read this file FIRST. It will be empty for Round 1.
Provide evidence for every finding (file paths, git output, code snippets).
Agent B brief:
RESEARCH BRIEF
==============
QUESTION: [user's question]
YOUR ANGLE: Domain Scout
ROUND: 1
Investigate from the DOMAIN angle. Search the web, read docs, find best
practices, benchmarks, and industry standards. Answer: "What does the world say?"
HISTORY FILE: .claude/how-to-win-history.local.md
Read this file FIRST. It will be empty for Round 1.
Provide evidence for every finding (URLs, quotes, benchmark numbers).
Agent C brief:
RESEARCH BRIEF
==============
QUESTION: [user's question]
YOUR ANGLE: Adversary
ROUND: 1
Investigate from the ADVERSARY angle. Find reasons things will fail, edge cases,
hidden constraints, dependency risks, breaking changes. Answer: "What will go wrong?"
HISTORY FILE: .claude/how-to-win-history.local.md
Read this file FIRST. It will be empty for Round 1.
Provide evidence for every finding (file paths, dependency versions, known issues).
- Wait for all 3 to complete
- Append findings to
.claude/how-to-win-history.local.md
Round 2 — What Did You Miss?
- Append Round 1 findings to history file
- Dispatch same 3 agents, same angles, but add to each brief:
ROUND: 2
PREVIOUS FINDINGS are in .claude/how-to-win-history.local.md.
Read them carefully. Your job this round:
- What was MISSED in Round 1?
- What was WRONG in Round 1?
- Go DEEPER on your angle — don't repeat what's already known.
- Wait for all 3 to complete
- Append findings to history file
Round 3 — Final Pass
- Append Round 2 findings to history file
- Dispatch same 3 agents, same angles, but add to each brief:
ROUND: 3 (FINAL)
ALL PREVIOUS FINDINGS are in .claude/how-to-win-history.local.md.
Read them carefully. This is the final pass. Your job:
- What is STILL missing after 2 rounds?
- Are there any CONTRADICTIONS between findings?
- What's the single most important thing from your angle?
- Wait for all 3 to complete
- Append findings to history file
Synthesis — The Winning Brief
After Round 3, synthesize ALL 9 reports into one brief:
WINNING BRIEF
=============
QUESTION: [original question]
ROUNDS: 3
AGENTS_DEPLOYED: 9
SITUATION:
[evidence-backed current state, from Archaeologist]
REAL GOAL:
[refined and measurable, informed by all 3 angles]
SUCCESS METRICS:
- [specific, measurable, with industry context from Scout]
VERIFICATION_COMMAND:
[concrete command to prove success, if applicable]
RECOMMENDED STRATEGY PATTERN:
[from references/strategy-patterns.md, informed by research]
WHAT WILL WORK:
- [approach — evidence from Scout + Archaeologist]
- [approach — evidence]
WHAT WON'T WORK:
- [approach — evidence from Adversary + Archaeologist]
- [approach — evidence]
RISK MAP:
- [risk]: [evidence] → [mitigation]
- [risk]: [evidence] → [mitigation]
CONSTRAINTS:
- [hard limits all strategies must respect]
CONFIDENCE: [High / Medium / Low]
REMAINING UNCERTAINTY:
- [anything still unclear after 3 rounds]
Output
Present the Winning Brief to the user. They can:
- Use it to inform a
/winning:launchwith a refined goal - Use it as standalone research
- Ask follow-up questions