Deep Research

Orchestrate parallel research sub-agents, then write one coherent cited report. Research is parallel; writing is single-point, never let multiple agents write report sections.

Step 0, Always first

1. Run date +%Y-%m-%d via Bash. Never assume the current year from training data.
2. Triage:
   • Answerable with 1-2 searches? → STOP, just search directly. Do not use this skill.
   • Enumeration task (N items × M fields, e.g. "compare 20 frameworks")? → still this skill, but use table-oriented decomposition (one sub-agent per item batch).
   • Open-ended investigation? → continue below.
3. Pick depth (default standard; user can override with words like "quick"/"exhaustive"):

Mode	Sub-agents (round 1)	Max follow-up rounds	Sources target
quick	2-3	0	8+
standard	3-5	1	15+
deep	5-8	2	25+

These are hard budgets. Reflection (Phase 4) can spend them but never exceed them.

Workspace

All state lives on disk at ./research/<slug>/, never only in context (survives compaction):

research/<slug>/
├── brief.md         # research brief, the single contract for all phases
├── findings/        # F1.md, F2.md ... one per sub-agent, structured evidence
└── REPORT.md        # final deliverable

On resume: re-read brief.md + list findings/, skip completed angles, continue.

Phase 1, Scope

Ask at most one round of clarifying questions (AskUserQuestion), only if genuinely ambiguous: audience, time frame, region, decision at stake. If the user said "just run it" or intent is clear, skip asking and write assumptions into the brief instead.

Then write brief.md: refined question, scope boundaries (in/out), assumptions, depth mode, today's date. This brief, not the raw conversation, is what every later phase measures against.

Phase 2, Plan