competitive-dominator

name	competitive-dominator
description	Dominates competitive coding challenges, hackathons, Kaggle, data science competitions, AI/ML contests, CTFs, buildathons, and any timed or scored coding event. Use when user says "competition", "challenge", "hackathon", "Kaggle", "submission", "leaderboard", "scoring", "evaluation criteria", "help me win", "solve this problem", "beat the score", "improve my score", "my score is", "top score is", or "prepare presentation for competition". Also use when user uploads competition datasets, sample submissions, or evaluation scripts. Do NOT use for general coding tutorials, creative writing, factual questions, or tasks unrelated to competitive/scored events. Do NOT use for routine software development without a scoring rubric or deadline.
license	MIT
compatibility	Works on Claude.ai, Claude Code, and API. Zero external dependencies for scripts. Enhanced when used with anti-slop-design skill for UI/presentation work and document creation skills (docx, pptx, xlsx, pdf) for competition deliverables.
metadata	{"author":"ankitjha67","version":"5.2.0","category":"competitive-coding","tags":["competition","hackathon","kaggle","data-science","algorithmic","ctf","ml","coding-challenge"],"repository":"https://github.com/ankitjha67/competitive-dominator"}

Beast Mode: Competitive Dominator

22 agents. 7 frameworks. 7 scripts. 11 references. 4 asset templates. Zero-hallucination war room for winning competitions.

Instructions

New to competitions? Start with assets/quickstart-cheatsheet.md for exactly what to say.

Step 1: Challenge Intelligence (Phase 0)

Before writing code, become the expert on this challenge.

1. Initialize state tracking:

python scripts/challenge_state.py init "[challenge name]" "[platform]" "[deadline]"
python scripts/memory_engine.py init "[challenge name]" "[platform]" "[deadline]"
python scripts/experiment_tracker.py init "[challenge name]"

2. Create the Competition PRD by reading agents/competition-prd.md — this living document becomes the single source of truth all agents reference.

3. Activate the Dataset Analyst (agents/dataset-analyst.md) to deeply understand every column, distribution, target variable, and expected output format. For data competitions, run automated EDA first: python scripts/auto_eda.py train.csv --test test.csv --target [target_col] --output eda_report.md

4. Read references/challenge-taxonomy.md to classify the challenge type. Consult references/platform-guide.md for language selection and compute limits. Review references/common-pitfalls.md to preempt the bugs that waste the most time.

5. Run the War Room Commander multi-lens scan (agents/war-room-commander.md): For EACH of the 22 agents, score relevance 0-10 and decide:

   • ACTIVATE NOW (7+): Deploy immediately with assigned task
   • DEFER (4-6): Document trigger condition for later activation
   • PASS (0-3): Document why — may be called later if conditions change

6. Research (15 min max): Domain insights, past winning solutions, external data sources.

Step 2: Architecture (Phase 1)

The Solution Architect (agents/architect.md) designs three versions:

• V1: Correct brute force that passes all samples (submit THIS as safety net)
• V2: Optimized for time/memory limits
• V3: Winning solution with advanced techniques

For ML competitions, grab a ready baseline from references/quick-baselines.md. For CP problems, grab templates from references/algorithm-templates.md. Follow frameworks/time-management.md to allocate time by competition type. For team events, follow frameworks/team-protocol.md for role assignment, git workflow, and standup cadence.

The CPM (agents/cpm.md) aligns every feature to scoring criteria with point-per-effort math.

Step 3: Implementation (Phase 2 — TDD)

1. Test Engineer (agents/tester.md) writes tests BEFORE code: Sample I/O, edge cases, stress tests (10,000+ random), format validation. For ML: set up CV pipeline using references/cv-strategies.md — wrong CV = shakeup disaster.

2. Lead Developer (agents/developer.md) implements module by module in the best language for this problem. Any language — Python, C++, Java, JS, Rust, Go.

3. Code Auditor (agents/code-auditor.md) runs after EVERY code change:

   • Correctness: off-by-one, overflow, edge cases → auto-fixes bugs
   • Complexity: verifies O(?) fits within time limit
   • Runtime: measures actual wall time with 50% safety margin
   • Quality: no debug prints, no magic numbers, clean code

4. Track experiments: python scripts/experiment_tracker.py log --name "desc" --cv [score]

5. Submit V1 immediately as safety net. Submission Strategist (agents/submission-strategist.md) manages daily limits — never waste a submission on untested code.

Step 4: Optimization (Phase 3)

1. Gap analysis: python scripts/challenge_state.py gap-analysis

2. If user shares a score, activate Leaderboard Hacker (agents/leaderboard-hacker.md): Decompose gap → generate hypotheses → execute by priority → measure → iterate. Use references/feature-engineering.md for the #1 score improvement technique.

3. Apply advanced techniques from frameworks/advanced-techniques.md: SHAP analysis, adversarial validation, Bayesian encoding, null importance, feature interactions, pseudo-labeling, stacking, test-time augmentation. Use references/hyperparameter-guide.md for Optuna configs by model type.

4. Activate Ensemble Strategist + Grandmaster (agents/ensemble-strategist.md, agents/ensemble-grandmaster.md) for diversity-maximized multi-model ensembles.

5. Activate Post-Processor (agents/post-processor.md) for threshold optimization, calibration, rank normalization — the last 0.1-1%.

6. Activate External Data Scout (agents/external-data-scout.md) if competition allows external data. 40+ free API/dataset sources cataloged.

7. MANDATORY checkpoint: python scripts/experiment_tracker.py checkpoint

Step 5: Submission (Phase 4)

Read references/submission-checklist.md for platform-specific checks. Validate format before submitting: python scripts/submission_validator.py check submission.csv sample_submission.csv

Before EVERY submission, triple gate:

• Code Auditor: All tests pass, complexity verified, no bugs
• Tester: Format validated character-by-character
• Submission Strategist: Worth spending a submission on this?

Submit 30+ minutes before deadline. Never submit untested code.

Step 6: Presentation & Documentation (Phase 5 — if scored)

Activate Presentation Engine (agents/presentation-engine.md):

• Slide deck (10-12 slides for hackathon pitch)
• Technical report (.pdf or .docx via document creation skills)
• Dashboard (React/HTML with real data and charts)
• One-page executive summary

The Competition PRD can generate BRD/FRD/SRS on demand.

After the competition ends, fill out assets/post-mortem-template.md to compound learnings. Review top winning solutions and extract techniques for next time.

Agent Directory

Always-On (9 agents — every challenge)

Agent	File	Role
War Room Commander	agents/war-room-commander.md	Orchestrates all agents. Multi-lens scan.
Chief Product Manager	agents/cpm.md	Scoring dictator. Time budget.
Solution Architect	agents/architect.md	Algorithm design. V1→V2→V3.
Lead Developer	agents/developer.md	Implementation in any language.
Test Engineer	agents/tester.md	TDD. Stress tests. Format validation.
Code Auditor	agents/code-auditor.md	Complexity. Runtime. Auto-fix bugs.
Competition PRD	agents/competition-prd.md	Living requirements doc.
Domain Scientist	agents/domain-scientist.md	Cross-domain expertise.
Dataset Analyst	agents/dataset-analyst.md	Deep data understanding.

Conditionally-Activated (13 agents)

Agent	File	Trigger
Data Scientist	agents/data-scientist.md	ML/data challenges
ML Engineer	agents/ml-engineer.md	Model training, LLM
Advanced Analytics	agents/advanced-analytics.md	SHAP, Bayesian, interactions
Ensemble Strategist	agents/ensemble-strategist.md	Multi-model ensemble
Ensemble Grandmaster	agents/ensemble-grandmaster.md	Stacking, blending
Post-Processor	agents/post-processor.md	Thresholds, calibration
Leaderboard Hacker	agents/leaderboard-hacker.md	Reverse-engineer scores
External Data Scout	agents/external-data-scout.md	Free APIs/datasets
Novel Generator	agents/novel-generator.md	Research paper techniques
Submission Strategist	agents/submission-strategist.md	Daily limit management
Presentation Engine	agents/presentation-engine.md	Slides, reports, dashboards
Code Reviewer	agents/reviewer.md	Final quality gate
Perf/Security/DevOps/Writer/UX/Risk	agents/conditional-agents.md	Per-trigger

Frameworks & References

File	Purpose
frameworks/advanced-techniques.md	50+ cutting-edge methods by domain
frameworks/memory-persistence.md	Checkpoint system, experiment tracking
frameworks/tactical-playbooks.md	Debug protocol, tournament strategy
frameworks/score-gap-playbook.md	Close gaps by size and effort
frameworks/unfair-advantages.md	Disproportionate-return techniques
frameworks/time-management.md	Hour-by-hour plans for every competition type
frameworks/team-protocol.md	Role assignment, git workflow, standups, demo prep
references/challenge-taxonomy.md	6 challenge types and strategies
references/ml-playbook.md	ML competition winning framework
references/web-playbook.md	Hackathon speed architecture
references/submission-checklist.md	Platform-specific submission checks
references/algorithm-templates.md	Ready-to-paste CP patterns (DP, graphs, strings, math)
references/platform-guide.md	Language selection, compute limits, runtime budgets
references/quick-baselines.md	10-minute baselines for every competition type
references/hyperparameter-guide.md	Optuna configs for LGB, XGB, CatBoost, NN, NLP
references/cv-strategies.md	CV strategy selector — avoid shakeup disasters
references/common-pitfalls.md	16 competition-killing bugs and how to avoid them
references/feature-engineering.md	Feature cookbook: aggregation, lag, encoding, text, geo

Scripts (zero external dependencies)

Script	Usage
scripts/challenge_state.py	python scripts/challenge_state.py init "Name" "Platform" "Deadline"
scripts/stress_test.py	python scripts/stress_test.py stress --solution sol.py --brute brute.py
scripts/memory_checkpoint.py	python scripts/memory_checkpoint.py checkpoint
scripts/experiment_tracker.py	python scripts/experiment_tracker.py log --name "desc" --cv 0.85
scripts/memory_engine.py	python scripts/memory_engine.py status
scripts/submission_validator.py	python scripts/submission_validator.py check sub.csv sample.csv
scripts/auto_eda.py	python scripts/auto_eda.py train.csv --test test.csv --target col

Memory Protocol (zero information loss)

After every phase, score improvement, or 3 experiments, output a checkpoint. Read frameworks/memory-persistence.md for the full checkpoint format.

To resume: user pastes the checkpoint, Claude reads it and continues from "Next Actions." Never repeat failed experiments — always check the failed experiment registry first.

Examples

Example 1: Kaggle ML Competition

User: "I'm entering a Kaggle competition to predict house prices. Help me win."

Actions:
1. Initialize state tracking (3 scripts)
2. Dataset Analyst: Analyze train.csv, test.csv, sample_submission.csv
3. Competition PRD: Populate with problem, scoring (RMSE), constraints
4. War Room: ACTIVATE Data Scientist, ML Engineer, Advanced Analytics, Domain Scientist
   DEFER Ensemble (until baseline exists), Post-Processor (until plateau)
   PASS Security, DevOps, UX
5. Architect: V1=LightGBM with raw features → V2=Feature engineered → V3=Stacked ensemble
6. Implement V1, submit as safety net
7. Feature engineering sprint, track each experiment
8. Ensemble, post-process, submit best
9. Checkpoint progress throughout

Example 2: Codeforces Problem

User: "Solve this: Given N integers, find longest subsequence where adjacent differ by at most K."

Actions:
1. Parse: N≤10^5, K≤10^9 → need O(N log N)
2. War Room: ACTIVATE Architect, Developer, Tester, Code Auditor
   DEFER Performance (until TLE), PASS Data Scientist, ML Engineer, External Data
3. Architect: V1=O(N²) DP → V2=O(N log N) with segment tree/sorted container
4. Tester: Write tests from samples + edge cases (N=1, all equal, sorted, reverse)
5. Implement V1, verify samples pass
6. Code Auditor: Check complexity, verify N=10^5 fits in time limit
7. If TLE: activate Performance Engineer, optimize to V2
8. Submit after all tests pass

Example 3: Hackathon (24-hour)

User: "We're in a 24-hour hackathon building an AI sustainability tracker."

Actions:
1. Competition PRD: Judging criteria (Innovation 30%, Execution 30%, Impact 20%, Demo 20%)
2. War Room: ACTIVATE ALL except Security/CTF-specific agents
3. Architect: FastAPI + React + Supabase + Claude API
4. Developer: Core feature by hour 6, polish by hour 10
5. Presentation Engine: Slides + live demo + README by hour 12
6. Code Auditor: Verify deployment works, no crashes during demo
7. Submit 1 hour early with backup demo video

Troubleshooting

Skill doesn't trigger

Cause: Query doesn't match trigger phrases. Solution: Explicitly say "Use the competitive-dominator skill" or include keywords like "competition", "challenge", "hackathon", "leaderboard", "scoring".

Context getting too long

Cause: Too many agent files loaded simultaneously. Solution: The War Room Commander uses progressive disclosure — only load agent files when that agent is ACTIVATED. Defer/Pass agents to save context. Run python scripts/experiment_tracker.py checkpoint to save state, then ask Claude to summarize and continue in a fresh context.

Memory lost after compaction

Cause: Long session caused context compaction. Solution: The memory persistence framework (frameworks/memory-persistence.md) produces checkpoint blocks. Paste the last checkpoint into a new conversation and say "Continue from this checkpoint." All progress, decisions, and failed experiments are preserved.

Script errors

Cause: Scripts require Python 3.8+. Solution: All 7 scripts use stdlib only (zero external dependencies). Verify: python scripts/challenge_state.py --help should show usage. If Python isn't available, the agent instructions still work — scripts are optional accelerators, not requirements.

Code Auditor finds issues

Cause: Code has bugs, complexity issues, or style problems. Solution: Code Auditor auto-fixes issues in place and adds regression tests. If a fix is non-trivial, it flags the issue and proposes the fix for human review. Always re-run ALL tests after any auto-fix.

Submission wasted on bad code

Cause: Submitted without proper verification. Solution: Always run the triple gate before submitting:

1. Code Auditor: APPROVED status required
2. Tester: All sample + edge + stress tests pass
3. Submission Strategist: Confirms this is worth a submission slot

Performance Notes

• Take your time with each phase. Quality over speed.
• Do not skip validation steps, even under time pressure.
• Read the full competition spec before writing any code.
• The biggest competition wins come from UNDERSTANDING THE PROBLEM, not clever code.