All work
Global agency, Claude Cowork plugin

Multi-agent research pipeline

A Claude Cowork plugin built for a global design agency in Hong Kong. Type one command on any topic, get back a sourced, fact-checked, branded PowerPoint. Six AI agents do the heavy lifting in parallel. Hours of research, done in minutes.

What gets created Every step saves a file you can open, read, and re-run.
Brief
research_brief.json
Step 1, Research
research_raw.md
Step 2, Analysis
analysis.md balance_review.md fact_check.md
Step 3, Slides
outline content design
Step 4, Build
presentation.pptx Google Drive

Outcome

4 steps · 6 agents · 1 command
Research to branded client presentation, delivered to Drive.

How it works

Intake, 5 questions

User answers 5 questions about audience, style, and constraints. The plugin writes research_brief.json. Each agent saves a file you can read and re-run, no black box.

Researcher (Opus)

Targets 15+ Tier-1 sources: McKinsey, BCG, Gartner, WEF, gov. Outputs research_raw.md with tier-tagged citations.

Analyst + Compliance in parallel

Opus extracts 3-5 insights with a balance review (1-10 score, rewrite under 7). Sonnet runs independent fact-checking and tier compliance. Both outputs merge before slide design.

Slide Creator (Opus)

Writes the narrative. Produces slide_outline, slide_content, and slide_design JSON. Each slide is treated as a design brief, not a summary paragraph.

QA Reviewer (Sonnet)

Checklist review across compliance and structure. On NEEDS-REVISION, loops back to the Slide Creator once.

Slide Builder + Drive upload

Sonnet renders the .pptx, runs programmatic QA on every slide, uploads via gws straight to the client's Drive. Half-day of desk research replaced by one command.

Inspectable, overridable, re-runnable

Every step writes a file you can open and edit. If the analysis is off, fix analysis.md and re-run from step 3, no need to redo the research. The pipeline is a chain of artefacts, not a black box.

Why six agents

Specialisation beats a single monolithic prompt. A focused QA agent catches compliance issues before they reach the rendered deck. A slide design agent treats each slide as a design brief, not a summary paragraph. Each agent ships structured JSON or Markdown that the next agent consumes verbatim, so a failure is isolated and retriable.

Model assignment

  • Opus: Researcher, Analyst, Slide Creator. Tasks that need judgment (search strategy, strategic synthesis, narrative structure).
  • Sonnet: Compliance Reviewer, QA Reviewer, Slide Builder. Tasks that benefit from speed and checklist reliability.