workshop / day-3
Day 3 · Wednesday, 3 June 2026
Customize & extend
Deliverable
Two-subagent critique of your Day 2 regression (one a causal-inference critic) + a three-subagent research report + 2 MCP servers + Superpowers + plugin-dev + Crawfurd's paper-review skill installed + a Superpowers-driven spec for an extended CPD analysis + your own plugin generated via /plugin-dev:create-plugin.
Slides
Open deck in presentation mode ↗Agenda
- § 1DiscussionDay 2 recap + Day 3 framing
- § 2LectureSubagents + hostile reviewer
- § 3WorkshopDispatch a subagent to critique your §4 regression
Launch two subagents to critique your empirical work
- 1Have the agent dispatch two fresh-context subagents and point both at your empirical work.
- 2Brief one as a generic hostile reviewer.
- 3Brief the other with a sharper lens of your choice: a causal-inference critic, a mechanism critic, whatever fits your work.
- 4Have the main agent summarize the critiques. You decide what to act on.
Launch three subagents to research a topic, then have the main agent write a report
- 1Pick a topic you want to understand.
- 2Have the main agent dispatch three subagents, each on a distinct angle.
- 3Have it synthesize their findings into one short report.
Done early? Try one of these.
- 1Chain the hostile reviewer into a fixer: feed its critique to a second subagent that proposes concrete fixes.
- 2Pit a critic against a defender, then adjudicate the disagreement yourself.
- 3Run a hostile and a sympathetic reviewer on the same work and compare what each surfaces.
- 4Red-team your own project memory file: have a subagent attack your conventions for gaps and contradictions.
- 1
- § 4LectureMCP servers
- § 5WorkshopInstall + use Playwright MCP and a docs-search MCP
Talk to the agent about
- 1Install the Playwright MCP server (github.com/microsoft/playwright-mcp).
- 2Install a docs-search MCP such as Context7 (github.com/upstash/context7-mcp).
- 3Verify both show as connected.
- 4Verify Playwright works: ask the agent to open a real web page and report what's there.
- 5Verify Context7 works: ask for the current signature of a fixest or pandas function and confirm it cites the doc source.
- 1
- § 6LecturePlugins + the Superpowers ecosystem
- § 7WorkshopInstall Superpowers + plugin-dev + Crawfurd's paper-review skill
Talk to the agent about
- 1Add the official plugin marketplace if it isn't already configured.
- 2Install the plugin builder your agent uses. You author your own plugin with it later.
- 3Install Superpowers (github.com/obra/superpowers).
- 4Install Crawfurd's paper-review skill (lcrawfurd.github.io/claude-skills).
- 5Verify each install: ask the agent to list available plugins and skills.
- 1
- § 8LectureSpec-driven development
- § 9WorkshopUse Superpowers to spec the CPD analysis
Talk to the agent about
- 1Invoke Superpowers' brainstorming skill to spec an extended CPD analysis that builds on your Day 2 regression.
- 2Let it interview you, one question at a time.
- 3When it asks if you're ready, switch to writing-plans to turn the spec into an analysis plan.
- 4Save the plan and read it end to end before approving.
- 1
- § 10LecturePlugin builder + lor preview
- § 11WorkshopCreate your own plugin with /plugin-dev:create-plugin
Talk to the agent about
- 1Use the plugin-dev plugin to create your own plugin: run /plugin-dev:create-plugin and let it interview you about what the plugin should do.
Stuck for an idea? Build one of these.
- 1Revise-and-resubmit responder: /rr:start runs one subagent per reviewer to role-play their objections, drafts the point-by-point reply in your voice, and an adversarial subagent checks each reply closes the point. A hook protects the submitted manuscript.
- 2Replication-package builder: /repkit:build orders your scripts into a master run, one subagent audits for nonreproducible code while another regenerates and diffs every table and figure, and a hook blocks done until they match.
- 3Pre-analysis plan: /pap:draft interviews you on hypotheses and design, subagents enumerate the specification grid and multiple-testing corrections, a power skill simulates minimum detectable effects, and it compiles a timestamped OSF-ready PAP.
- 4Discussant report: /discuss:prep reads a paper you are assigned to discuss, subagents extract the contribution and stress-test the identification, and a skill drafts your discussant slides and speaking notes.
- 5Grant proposal builder: /grant:draft drafts specific aims, broader impacts, and the data-management plan with phased subagents, and a critic subagent scores it against the solicitation's review criteria.
- 6Letter-of-recommendation writer: /lor:write interviews you about the candidate, drafts in your voice from past letters, a subagent calibrates the signal to the strength you intend, and it compiles to a letterhead PDF.
- 1
- § 12LectureOdds and Ends
- § 13DiscussionDebrief + Day 4 preview
Materials
- Superpowers (community plugin)available
- plugin-dev (official plugin)available
Provides /plugin-dev:create-plugin and 7 plugin-authoring skills.
- Crawfurd's paper-review skillavailable
Motivating example for adversarial review.