| name | spec |
| description | Entry point for paper research requests. Routes the request into one of 4 branches (search / read / analyze-compare / ideate). Triggers on requests like "find papers about X", "read this paper for me", "compare these", "any ideas from this". |
spec — paper workflow router
Process
- Classify the request. Judge the user's request into one of 4 branches:
- Search ("find papers on X", "what's recent research on X?") → call
literature_review
- Read ("read this paper", "summarize this", targets one specific paper/link) → call
go
- Analyze/Compare ("compare these two papers", "analyze this method", "which is better?") → call
check
- Ideate ("what could I build from this", "next research direction", "ideas based on this paper") → call
ship
- Ask when ambiguous. If the request spans branches (e.g. "find it and summarize it") or the branch is unclear, use AskUserQuestion or propose a sequential plan (search → read) up front.
- Delegate to the chosen branch skill. Once routed, spec does not do the work itself — it hands off to the branch skill.
- Confirm domain scope. This harness covers the full ML/AI space (LLM/NLP, RL/agents, vision/multimodal, efficiency/systems ML). If a request falls outside this, say so explicitly.
Anti-Rationalization
- "Just search everything and show it" — skipping the branch means verification skills (citation_accuracy, source_verification) never run. Always route through a branch.
- "It's ambiguous, default to search" — don't silently guess on ambiguity; ask the user.
Evidence
Routing is correct when: the chosen branch skill's output (under _papers/, _ideas/, _sessions/) matches the user's actual intent.
Red Flags
- Answering directly without invoking any of the 4 branch skills
- Silently assuming intent on an ambiguous request
