Research Project

Run the complete PRD creation pipeline as a single invocation. Chains three skills in sequence — diverge (multi-perspective research), converge (structured debate), premortem (failure analysis) — passing file outputs between them. Produces a risk-annotated PRD ready for /prd-build or /scaffold.

Arguments

$ARGUMENTS — format: [N] "research question or topic" where N is optional agent count (default: 3)

Parse Arguments

Extract:

• agent_count: the optional leading integer (default 3, min 2, max 6)
• topic: the research question, feature description, or problem statement

If the topic is missing or unclear, ask the user to clarify before proceeding.

Phase 1: Diverge

Invoke the /diverge skill with the provided arguments. This spawns N independent research agents — each with a different lens (technical feasibility, user experience, risk/failure modes, prior art, first principles) — to explore the topic from uncorrelated perspectives.

Wait for the PRD file to be created (format: prd_{slugified_topic}.md).

The diverge output is a synthesis of all agent findings plus a draft PRD. Do NOT proceed until the file exists.

Phase 2: Converge

Invoke the /converge skill, passing the PRD file path from Phase 1.

This runs a structured debate: agents advocate competing positions from the diverge output, challenge each other's assumptions, propose compromises, and reach consensus. The result is a refined PRD with tensions resolved and trade-offs made explicit.

Wait for the convergence report and updated PRD.

Phase 3: Premortem

Invoke the /premortem skill, passing the refined PRD from Phase 2.

This spawns N agents who each write a narrative from the future where the project has failed — each for a different root cause (technical, operational, security, scope, organizational). Synthesizes into a risk registry with severity ratings and mitigations, then annotates the PRD with the top risks.

Wait for the risk-annotated PRD.

Phase 4: Present Results

Present the final risk-annotated PRD path and a brief summary:

• Key decisions from convergence (what was debated, what was chosen, and why)
• Top 3 risks from premortem (with severity and recommended mitigations)
• Recommended next step (usually /prd-build for parallel implementation, or /scaffold for build-order planning)

Rules

• Do NOT ask questions between phases — flow continuously
• Do NOT skip phases — all three must run
• If a phase fails, report the error and stop (do not proceed with incomplete input)
• Each phase produces files — pass file paths between phases, not inline content
• The three phases MUST run in sequence (each depends on the previous output)

Pipeline Position

• Before: /brainstorm for ideation when the problem space is vague, or start here when you have a clear question
• After: /prd-build to decompose the PRD into parallel work units and implement, or /scaffold for build-order planning, or /diverge-prototype to spike implementations before committing

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sjarmak
• Source: sjarmak/agent-workflows
• License: MIT

Install and usage instructions live in the source repository linked above.