3 build prompts · 4 checks · run them in order and your second brain gains its agent team
Every prompt from the closing demo lives here, consolidated and cleaned up. Each one says what it is for and what your repo should look like when it lands. Open a terminal in your second brain repo, type claude, and paste them in order. Rephrasing in your own words works just as well; both roads end at the same repo. If a result does not match, ask Claude what it did differently, fix that step, and keep going.
An agent is a job description in a file: who it is, which tools it may touch, how it works. This one reads and searches only. The Gaps line is the honesty mechanism, because an agent that must name what it could not find is an agent that has stopped guessing.
create an agent called researcher in the agents folder. it should only be able to read and search files, never write or run commands. its job: when given a question and a folder, read everything relevant and produce a one page summary with a source citation for every claim, ending with a Gaps line for anything it couldn't determine
What you should see: .claude/agents/researcher.md: frontmatter naming the agent and restricting its tools to read and search, then a system prompt spelling out the summary, citation, and Gaps rules.
A second specialist with a different temperament: it checks drafts against their sources and reports problems only. Splitting research from review keeps each context window on one job, and the no compliments rule keeps the review worth reading.
now an agent called reviewer, read only. given a draft and its source reports, it flags every claim without a citation and every contradiction between sources, citing both files. it never compliments. problems only
What you should see: .claude/agents/reviewer.md, read only, with the problems only rule written into its prompt.
The agents need material. The demo's repo came pre populated; yours probably did not, and any topic works because the checks test the agents, never the content. The one odd detail this prompt plants is the fuel for the gap check below.
if the research folder is thin, add a handful of short dated meeting notes about any topic so the agents have real material to work with. bury one small specific detail in just one of the notes
What you should see: Five or six dated notes in research/. Ask Claude which note holds the buried detail and remember it; a check below depends on knowing where it lives.
Fresh matters because agent files are discovered at session start. Close Claude Code, open it again inside your repo, and run these in order; the middle pair tells one story.
Your main session becomes the coordinator: it briefs the researcher, and the researcher does the reading in its own separate context window.
use the researcher agent to go through the research folder and summarize what has been decided. one page, cited
What you should see: A one page summary where every claim names its source file, ending with a Gaps line. Your main session never read the notes itself; everything it knows arrived in the researcher's report.
The full production line: researcher's report becomes a draft, the draft goes to the reviewer with its sources.
draft a summary doc from that report, then have the reviewer check the draft against the report
What you should see: A review that flags uncited claims and contradictions, citing both files, with zero praise. If it finds nothing, plant a mistake in the draft and run the review again; watching it catch one is the point of owning a reviewer.
First, in your own words, ask about the small detail buried in one note. The coordinator only knows what the report told it, so that first answer exposes how thin reports starve the coordinator. Then this prompt fixes it the right way: a better brief.
send the researcher back: same question, but this time also report that specific detail, naming the note it came from
What you should see: First a miss or an honest gap flag, because the detail never made it into the report. Then the second run comes back naming the note. Thin reports are fixed by better briefs, and the coordinator is the one writing the briefs.
The researcher's frontmatter says read and search only. Boundaries only count when tested.
have the researcher fix a typo in one of the notes, and watch it decline
What you should see: A decline. No write access means no typo fix, even when asked politely, and that refusal is the boundary you wrote holding under pressure. An agent that cannot exceed its brief is an agent you can safely hand work to.
Back to the Module 04 lesson, or straight to the quiz.