nitro-research

A question needs an answer that is backed by sources, not by recall. Produce a Markdown file that leads with the answer, cites every claim, and names what stayed unknown. Do it without burning the expensive model's context on raw reading: the lead session frames and judges, cheap subagents read. The steps live in nitro agent tasks so a dead session can be resumed, and the answer lands in nitro agent memory so nobody researches it twice. Command mechanics for both live in the nitro-task and nitro-memory skills; this skill covers how research composes them. If nitro itself is not found, the CLI is not installed — stop and tell the user to install it: https://chillicream.com/docs/nitro/cli/installation. Do not attempt to install it yourself.

Core principles

• The lead never reads sources. You (the main session) write the brief, dispatch, judge the synthesized result, and own every tracker write. Raw pages, files, and search results stay inside subagent contexts. Every raw token that reaches the lead is money spent twice.
• Files are the channel, not chat. Subagents write their output to the research folder and return a summary of at most ten lines. The next stage reads the files, not the previous agent's reply. The lead pastes that summary into the step's task comment, which is how the tracker becomes the log.
• Cheapest model that can do the role. Locating and extracting are cheap; judging and verifying are not. Give each role the lowest tier that does it reliably and escalate a single role only when its output fails the check below, never the whole pipeline.
• Memory first, sources second. A question the workspace already answered costs one search. Check before dispatching anything; save the answer when done.
• Primary sources only. Official docs, source code, specs, first-party APIs, RFCs. A blog post or a forum answer is a lead to a primary source, never a citation. Every claim in the findings names the source that owns it.
• Facts, not decisions. Research reports what is true and what is unknown. If the result forces a choice, the choice belongs to the caller (a grilling ticket, the user), not to this skill.

The roles