Research Task Router

Read references/profile-map.md before deciding if the request is ambiguous. Prefer the bundled files above so the skill stays portable across machines and repositories.

Routing Workflow

1. Load The Base Rules

Always read references/profiles/CLAUDE.md first.

Treat it as the universal baseline for:

• think before acting
• keep solutions simple
• make surgical changes
• define verifiable goals

2. Identify The Primary Deliverable

Classify the request by what the user mainly wants at the end of the turn, not by isolated keywords.

Use these categories:

• paper: engineering master's thesis writing, chapter outline, abstract, introduction, related work, requirements analysis, system design, implementation, testing chapter, conclusion, polishing academic prose
• analysis: metric comparison, ablation interpretation, table writing, result summary, data inspection, error analysis, trend explanation, chart commentary
• coding: experiment scripts, model code, debugging, refactoring, reproduction, data processing code, plotting code, notebook logic
• benchmark: pass-the-test tasks, minimum-diff fixes, fast benchmark iteration, token-sensitive coding tasks
• agents: batch pipelines, scheduled experiments, multi-agent orchestration, structured machine-readable outputs, automation flows

If the request mixes categories, choose the category that determines the final output format.

Examples:

• If the user asks to "analyze these experiment tables and write result conclusions", choose analysis.
• If the user asks to "fix the training script that produced these bad results", choose coding.
• If the user asks to "draft the system testing chapter from these logs and tables", choose paper first, then consult analysis if numeric grounding is needed.

3. Load The Matching Task Profile

After classification, read exactly one primary task profile:

• paper -> references/profiles/CLAUDE.paper.md
• analysis -> references/profiles/CLAUDE.analysis.md
• coding -> references/profiles/CLAUDE.coding.md
• benchmark -> references/profiles/CLAUDE.benchmark.md
• agents -> references/profiles/CLAUDE.agents.md

4. Resolve Mixed Tasks With Minimal Overlap

Do not load every profile "just in case". Prefer these combinations:

• paper + analysis: thesis sections grounded in tables, metrics, logs, or experiment evidence
• coding + analysis: analysis code, plotting code, metric computation, notebook cleanup
• coding + benchmark: fast bug fixes where minimum diff and pass rate matter
• agents + analysis: automated result aggregation with structured outputs

Avoid combining more than three profile files in one turn including the base file.

5. State The Active Profile Set Briefly

Before substantial work, state the chosen routing result in one short sentence.

Use a compact pattern such as:

• Active profiles: CLAUDE.md + CLAUDE.coding.md
• Active profiles: CLAUDE.md + CLAUDE.paper.md because the goal is engineering-thesis drafting
• Active profiles: CLAUDE.md + CLAUDE.paper.md + CLAUDE.analysis.md because the chapter depends on test evidence

Do not turn the routing explanation into a long preamble.

Decision Rules

• Prefer the user's explicit instruction over automatic classification.
• Prefer the final deliverable over surface keywords.
• If the user explicitly names a profile, obey that unless it clearly conflicts with the requested output.
• If confidence in classification is below moderate, say so briefly and choose the safest narrow profile.
• For paper-writing tasks, never invent citations, numerical claims, deployment facts, or experiment conclusions.
• For coding tasks, never switch to analysis-only mode just because metrics or logs appear in the prompt.

Output Contract

After routing, continue with the actual task. Do not stop at classification unless the user asked only for routing.

When relevant, include:

• active profile set
• any important assumptions caused by ambiguity
• the final task output requested by the user