research

name	research
description	Use when exploring unfamiliar codebases, investigating bugs, checking external documentation, or learning unfamiliar technology before acting

Research

Investigate at the shallowest depth that answers safely, then stop: Observe → Explore → Verify.

1. Observe — read the cheapest orienting source first (CLAUDE.md, README, the one named file). Name what you don't yet know.
2. Explore — answer the open questions. Do it directly only when it's shallow; otherwise delegate (below).
3. Verify — close the loop once: run it, trace the call, or cross-check a second source before you act on the finding.

Delegate Wide Research to a Subagent

Orient first (the one cheap read), then decide — orientation often shrinks a "wide" question to a one-file answer, so don't delegate before you've looked.

Once oriented, if what's still open would read several files or sources — code, web, or both — delegate to Agent: subagent_type="me:researcher". Its context absorbs the dead ends; you get only the summary, keeping the main context clean.

Do it yourself only when delegation overhead isn't worth it:

What's left open	Do
One named file / known symbol	Direct Read / grep / LSP — you can cite file:line in a step or two
Behavior claim, runnable locally	Reproduce with a command or test yourself
Wide/unknown code, web docs, or a mix	Delegate to me:researcher

Dispatch independent research in one message — several me:researcher calls (or a direct read alongside one) run concurrently. Only go sequential when a call needs the previous result. Match the agent count to the question's real branches — one subagent per genuinely independent sub-question, not per file. Each spins up a fresh context, so over-splitting burns tokens for no extra signal. Give each subagent one objective and the exact fields you want back.

Evidence Standards

• Match evidence depth to risk; quick lookups need one cited source.
• Bug causes, recommendations, comparisons, and external facts want 2-3 independent signals — static config and runtime artifact count as two; two reads of the same file do not.
• Behavior claims need observed output, not code reading alone.
• Test the premise, not just the question — the report ("X is broken") may itself be false; check the actual state.
• Negative evidence is evidence: say where you searched and what was absent.

Stop When

The next search won't change your decision. You have a citable answer at the depth the risk demands. Beyond that, more reading is confirmation, not signal.

Red Flags — STOP

• 3-source research for a file-location question.
• Reading file after file in the main context when me:researcher should absorb that work.
• Dispatching subagents one per message when several are independent.
• Fanning out more subagents than the question has independent branches — context spin-up is the cost, not the search.
• Claiming behavior from code when a runnable check exists.
• Reading code for a live URL/UI bug before inspecting the running page (me:browse).
• Trusting outdated or unsourced web content.