Tournament Search

Tournament-Search

Knockout bracket search: generate diverse candidates, run head-to-head matches, synthesize a hybrid per match, advance the best of 3 (Parent A, Parent B, Hybrid) until one champion emerges. Includes a 3rd-place playoff. Faster than evo-search; quality ceiling is bounded by initial population diversity.

---

Step 0: Problem Intake & Rubric

Identify problem type, constraints, audience, and scope. If ambiguous, ask one question.

Trait Logging (off by default)

Activate when the user says things like "keep the best ideas from losers", "log traits", or "don't throw away good ideas". When active: before each elimination, extract and record the loser's single most distinctive feature into a trait log for review at the end.

Rubric Construction

Detect which mode applies:

| Mode | Trigger | Action | |------|---------|--------| | A: Auto | User gave only the problem | Generate 4–6 domain-appropriate criteria | | B: Guided | User hinted at priorities | Generate rubric, weight toward stated priorities | | C: Manual | User gave explicit criteria | Convert each into a scored rubric entry with anchors |

In all modes: augment vague criteria into scorable definitions, and always add: > Overall Fitness (30%): "Would a knowledgeable expert prefer this over a competent > but unremarkable response?" Scored holistically. Prevents narrow-criteria gaming.

Present the rubric to the user and wait for confirmation before proceeding.

Rubric format:

| Criterion      | Description              | Weight | Max |
|----------------|--------------------------|--------|-----|
| [Name]         | [Definition + anchors]   | X%     | 10  |
| Overall Fitness| Expert holistic score    | 30%    | 10  |
Weighted Total = Σ(score × weight)  [max = 10.00]

---

Step 1: Initial Population & Bracket Draw

Default population: 8 candidates (3 rounds + playoff). Fast mode: 4 candidates (2 rounds + playoff). Population must be a power of 2; round up if the user requests an odd number and generate an extra candidate.

Generate all candidates using these diversity frames before scoring any:

| # | Frame | |---|-------| | 1 | Conventional / mainstream | | 2 | Contrarian / challenges assumptions | | 3 | First-principles / bottom-up | | 4 | Analogy-led / draws from another domain | | 5 | Risk-focused / emphasizes what could go wrong | | 6 | Synthesis / combines multiple angles | | 7 | Minimalist / most parsimonious | | 8 | Maximalist / most comprehensive |

Score each candidate and display:

INITIAL POPULATION
| # | Frame        | Fitness | Key Strength  |
|---|--------------|---------|---------------|
| 1 | Conventional | X.X     | [one phrase]  |

Bracket draw is random. Shuffle candidate numbers and pair sequentially. Show the bracket to the user before starting matches.

BRACKET DRAW
Match 1: [A] vs [B]
Match 2: [C] vs [D]
Match 3: [E] vs [F]
Match 4: [G] vs [H]

---

Step 2: Run the Tournament

Every Match Follows These 3 Steps:

1. Synthesize a Hybrid Generate a third solution that inherits the key strength of each parent without their primary weaknesses. The hybrid must be a genuine conceptual synthesis, not a text blend. If trait logging is active, record each parent's most distinctive feature before generating the hybrid.

2. Score All Three Evaluate Parent A, Parent B, and the Hybrid against the rubric. Record fitness and a one-line diagnosis per competitor.

3. Advance the Winner Highest scorer advances. On a tie, prefer the Hybrid.

Match display format:

MATCH [N]: [Candidate A] vs [Candidate B]
|          | Fitness | Strength     | Weakness     |
|----------|---------|--------------|--------------|
| Parent A | X.X     | [one phrase] | [one phrase] |
| Parent B | X.X     | [one phrase] | [one phrase] |
| Hybrid   | X.X     | [one phrase] | [one phrase] |
→ ADVANCES: [winner] (X.X)
[→ TRAIT LOGGED from [loser]: [one sentence]  ← only if trait logging is active]

8-Candidate Bracket

QUARTER-FINALS    (Matches 1–4): 4 winners advance
SEMI-FINALS       (Matches 5–6): 2 winners → Final | 2 losers → 3rd-Place Playoff
3RD-PLACE PLAYOFF (Match 7)    : L1 vs L2 → 🥉
FINAL             (Match 8)    : W1 vs W2 → 🏆

4-Candidate Bracket

SEMI-FINALS       (Matches 1–2): 2 winners → Final | 2 losers → 3rd-Place Playoff
3RD-PLACE PLAYOFF (Match 3)    : L1 vs L2 → 🥉
FINAL             (Match 4)    : W1 vs W2 → 🏆

---

Step 3: Output

3.1 Podium Summary

🥇 1st — [Frame/description]   Score: X.X
🥈 2nd — [Frame/description]   Score: X.X
🥉 3rd — [Frame/description]   Score: X.X

3.2 Contrast Table (show before full solutions)

| Rank | Score | Defining Strength | Best Used When          |
|------|-------|-------------------|-------------------------|
| 🥇   | X.X   | [one phrase]      | [context where #1 wins] |
| 🥈   | X.X   | [one phrase]      | [context where #2 wins] |
| 🥉   | X.X   | [one phrase]      | [context where #3 wins] |

3.3 Full Solutions

Output all 3 podium solutions in full, labeled with rank and score. User can override with plain language: "just show me the winner" → top_k=1.

3.4 Rubric Breakdown

Show per-criterion scores for the champion (1st place) only.

3.5 Bracket Recap

QF:    [A](X.X) vs [B](X.X) → [winner](X.X) ✓
       ...
SF:    [W1](X.X) vs [W2](X.X) → [winner](X.X) ✓  | Loser → playoff
       ...
3rd:   [L1] vs [L2] → 🥉 [winner](X.X)
Final: [W5] vs [W6] → 🏆 [champion](X.X)

3.6 Trait Log (only if trait logging was active)

TRAIT LOG: Ideas from Eliminated Candidates
- [Round 1, Candidate X]: [one-sentence description of their best feature]
- ...
These traits did not reach the final but may be worth incorporating manually.

---

Parameters

| Parameter | Default | Options | Effect | |---------------|---------|---------------------|--------------------------------------------| | population | 8 | 4 or 8 | Initial candidates (must be power of 2) | | rubricmode | auto | auto/guided/manual | How the rubric is built | | traitlogging | off | on / off | Log best traits from eliminated candidates | | topk | 3 | 1–3 | Podium solutions shown in full | | showbracket | true | true / false | Show full bracket recap | | show_trace | true | true / false | Show per-match score tables |

Users set parameters in plain language. Examples:

• "Quick tournament, 4 candidates" → population=4
• "Keep the best ideas from losers" → trait_logging=on
• "Just show me the winner" → top_k=1
• "Use these criteria: impact, clarity, feasibility" → rubric_mode=manual

---

Scoring Guidance

Score honestly. Inflated scores make match outcomes meaningless. A mediocre-but-coherent answer scores 5–6. The hybrid winning its match is the ideal outcome; if hybrids consistently lose to parents, make the synthesis more targeted. Keep candidate length appropriate to the problem. Verbosity must not inflate fitness scores.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: smkalami
• Source: smkalami/skills
• License: MIT

Install and usage instructions live in the source repository linked above.