orchestrate

Personal Skills Marketplace

Marketplace of highly fine-tuned skills, generated by me.

Planned skills

• LATEX — highly fine-tuned LaTeX skill (beyond generic LaTeX help).
• WRITING — paper-writing skill that tracks modification history and translates it into future intent of the work (i.e. remembers where a draft is headed, not just its current state).
• MATH — proper structure for generating proofs and mathematical conceptual understanding.
• More TBD — needs further thought.

Agent-infrastructure skills

Second track (see PLAN.md "Second track"):

• orchestrate — five-tier, provider-neutral model routing for subagents, with mappings for Anthropic (Fable orchestrates; Opus deep-reasons; Sonnet implements; Haiku executes mechanical/bulk work) and OpenAI Codex (gpt-5.1-codex-max/-codex/-codex-mini with reasoning-effort settings), cross-provider routing via the codex/claude CLIs, and escalation/demotion rules.
• repo-wiki — generates a GitHub wiki from CLAUDE.md/README/docs with source-commit staleness stamps, and makes it the first-read entry point for exploring agents via pointers in both CLAUDE.md and AGENTS.md (Claude Code and Codex respectively).
• llm-harness — builds an eval harness (cases, graders, baseline accuracy) around any LLM-dependent system so prompt changes are measured, not vibes.

Status

MVP built: .claude-plugin/marketplace.json plus all three skills implemented. Distribution model: a personal Claude Code plugin marketplace, following the existing ~/.claude/plugins/marketplaces/ pattern (like caveman), not standalone skill repos.

• WRITING (flagship) — journal mechanism (.writing-journal.md), forward-looking brief loop, and helper scripts implemented and functionally tested. Not yet validated against a real personal draft (PLAN.md's success criterion) — that's the next milestone.
• LATEX — compile-log triage and bib hygiene scripts are real and tested; venue template is still a generic placeholder pending a real submission to test against (PLAN.md Phase 2).
• MATH — ships standalone: sympy_verify.py is a real, tested symbolic/numeric verification round-trip, clearing PLAN.md's ship condition rather than folding into LATEX.

See PLAN.md for the full concept: problem statement, decisions, skill specs, phased milestones, success criteria, and risks. See AGENTS.md and CLAUDE.md for how to build and test skills in this repo.