DeepResearch skill

URL: https://mkdocs.justinsforge.com/memory/general/reference_deepresearch_skill/

/DeepResearch <question> [--depth light|standard|deep] [--model sonnet|opus|haiku] [--fg]

Forge-native deep-research orchestrator. Built 2026-05-02 to fill the "agentic multi-step web research" gap (Google's Deep Research equivalent). Renamed from the lowercase research skill on 2026-05-03; the plain-English name was getting auto-triggered on natural-language research requests. Capitalized name + explicit warning in skill description forces invocation by name only.

How it works

1. Skill at forge/.claude/skills/DeepResearch/SKILL.md parses the question + flags.
2. Slugifies and stages output path: forge/memory/research/<slug>-YYYY-MM-DD.md.
3. Builds an XML-bounded worker prompt (system_rules / reference_data / active_payload per Section 11 of FORGE-DOCTRINE.md).
4. Spawns claude -p background worker with --allowed-tools WebSearch,WebFetch,Read,Write,Bash.
5. Worker loops decompose → search → fetch → re-search until coverage is solid or depth budget hit.
6. Worker writes the full report and exits. Wrapper fires forge_notify.sh info on completion (or warning if no file produced).

Depth budget

Depth	Iterations	Sources	Words	Wall time
light	1-2	3-5	~300	2-4 min
standard	3-5	6-12	~800	5-10 min
deep	6-10	12-25	~1500	10-20 min

Output contract

Every report has: header (Date/Depth/Model), TL;DR, Findings with inline [N] citations, Disagreements section, Sources with full URLs, Search trail. Indexed by /recall since memory/ is in scope. Renders at https://mkdocs.justinsforge.com/memory/research/<slug>-<date>/.

Cost model

Subscription quota only. No external API. Per cost-aware automation defaults.

Related

• /strategy <brand> for brand-specific deep thinking (different surface, internal context heavy)
• /watchyoutube to ingest a video as input before researching
• /followup to schedule a /DeepResearch for later
• /recall for searching internal forge knowledge (no web)