Add ollama-multi-model-writer skill for local GPU paper writing

[adityak74]

Summary

Introduces a new ollama-multi-model-writer skill that enables research paper writing using locally-running Ollama models on GPU, with no cloud API dependencies. The skill intelligently routes paper sections to specialized models optimized for different writing tasks.

Key Changes

• New skill: ollama-multi-model-writer (skills/ollama-multi-model-writer/SKILL.md)

  • Comprehensive documentation for multi-model paper writing workflow
  • Three-phase pipeline: Reasoning (DeepSeek-R1) → Structure (Phi-4-reasoning) → Synthesis (GLM-4.7-Flash)
  • Section routing table mapping paper sections to optimal models
  • Prerequisites verification and error handling for missing Ollama/models
  • No-fabrication policy with [PLACEHOLDER] and [RESULT NEEDED] tags for missing data

• Prompt templates for consistent model invocation

  • templates/abstract-template.md — 5-part abstract structure for Phi-4-reasoning
  • templates/methods-template.md — Detailed methods section guidance for DeepSeek-R1

• Bash runner script (scripts/ollama-runner.sh)

  • Validates Ollama installation and model availability
  • Streams model output with clear error messages
  • Prevents execution with missing models

• Integration with lead-researcher pipeline

  • New Stage 7c option for local GPU workflows
  • Added to teammate roster in config.json with proper reporting chain
  • Task completion gate in TaskCompleted.sh ensures draft manuscript is produced
  • Updated lead-researcher SKILL.md with new stage documentation and shortcut routing

Notable Implementation Details

• Model specialization: DeepSeek-R1 handles technical/scientific sections (Methods, Results), Phi-4-reasoning handles narrative sections (Abstract, Introduction, Conclusion), GLM-4.7-Flash performs cross-section synthesis and consistency checking
• Graceful degradation: Each section has a backup model if the primary is unavailable
• Data integrity: Strict policy against fabricating results, citations, or benchmarks — all quantitative claims must come from user's actual experiments
• Output structure: Produces paper-{topic}-draft.md with model attribution table and data gaps section for transparency

https://claude.ai/code/session_01DYJVNVqTVq7dDzTjzEZHrr