โ† ClaudeAtlas

tandem-engineering-workflowlisted

The STANDARD engineering workflow for the tandem system. Trigger on any development, debugging, or feature work driven through tandem โ€” build, create, implement, fix, add, develop, code, debug, or when a Claude Code session's output is being reviewed, or a plan/phase is being discussed. This is the discipline the director and the manager layer follow. If anyone tries to skip steps (no spec, no plan mode, no plan review), STOP THEM. When triggered, begin your first message with ๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹ so the user knows the workflow is active.
Maxmedawar/tandem ยท โ˜… 14 ยท AI & Automation ยท score 80
Install: claude install-skill Maxmedawar/tandem
# Tandem Engineering Workflow ๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹ โ€” When you see this, the engineering workflow is active. Start EVERY response in this workflow with `๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹๐Ÿ“‹`. This is the **standard** every layer holds to. In tandem the director (chatbot), the manager, and the worker are all AI โ€” the human is only in the loop when they want to be. This workflow defines the review discipline that keeps autonomous work honest. --- ## ROLES (all AI except the human) - **Human** โ€” sets the goal, makes final calls, steps out whenever. - **Director (chatbot: Claude.ai or ChatGPT)** โ€” brainstorms with the human, writes the spec, audits final output, drives sessions over the tandem MCP. - **Manager CC session** โ€” owns the build: plans phases, reviews the worker, never rubber-stamps, escalates blockers. Follows THIS workflow. - **Worker CC session** โ€” does the hands-on work and spawns agent teams (see `tandem-agentic-engineering`). The old "human pastes between two windows" model is gone. The director drives the sessions, and the manager reports back through the browser loop (see `tandem-orchestration`). The steps below are the quality gates each AI layer enforces on the next. --- ## THE WORKFLOW ### PHASE 0: BRAINSTORM & RESEARCH 1. The human gives the goal in as much detail as they have. 2. Don't guess โ€” research: search GitHub for implementations, look for prior art, libraries, patterns; verify what's actually possible and what the best approach is. 3. Give honest takes. If something is a bad