use-research-context

use-research-context — router

This is a dispatcher, not an implementation skill. Each atom category has its own sub-skill with a tight tool whitelist and a context: fork policy island. Pick the right sub-skill and hand off.

Procedure

1. Load the structured paper context via the MCP tool — do not Read research/research.md directly:

   mcp__paper-compiler__get_paper_context()
   

   Returns paper title + id, atom counts by category, top-priority atoms, top communities, open missing_details count, recent session count, recent decisions count. This is the only one-call summary you should need before dispatching.

2. Check for resumable work. mcp__paper-compiler__list_sessions(limit=3). If a recent session exists and the user's prompt hints at continuation ("where were we", "continue", "next step"), invoke /paper-compiler:use-research-context continue (sub-skill) before doing anything else.

3. Dispatch by atom category. Either the user named a category (loss / encoder / dataset / optimizer / evaluation / baseline), or you can infer one from their prompt. Run the dispatcher:

   ${CLAUDE_PLUGIN_ROOT}/scripts/select-playbook.sh "<user phrase>"
   

   Output format: <category> <sub-skill> <slash-command-path>.

   Then invoke the named sub-skill. Each sub-skill picks up with its own forked context.

4. Sub-skills available:

   • implement-method — algorithmic/structural unit (ML architecture, physics scheme, chem route, bio protocol)
   • implement-objective — loss/Hamiltonian/yield/fitness function
   • implement-data — dataset/measurements + preprocessing
   • implement-procedure — optimizer/integrator/protocol + parameters
   • implement-evaluation — metrics/diagnostics/statistical tests
   • implement-baseline — published comparison method
   • debug-divergence — when implementation disagrees with paper
   • continue — resume mid-implementation (Phase D)
   • port — port to a different repo (Phase D)

Hard rules (enforced by .claude/settings.json deny patterns)

• Do not Read / Glob / Grep research/wiki/atoms/, research/wiki/papers/, research/wiki/communities/, research/evidence/, research/graph.json, research/research.db. Use the MCP tools — they return structured snippets and are autoApproved.
• research/research.md is allowed but unnecessary — get_paper_context() gives you the same structured information in less context.
• Cite atom_uid (the v1.0 stable id) in code comments, not the sequential atom_id (resh­uffles on rebuild).
• Significant choices and gotchas → record_decision(...) (requires user approval; lands in research/decisions.md and is visible to future sessions).