AI systems / orchestration
Agent mode is the controlled workflow layer where AI decomposes tasks, uses tools, retrieves context, validates output, and keeps humans in review loops.
What this system does
Breaks complex product/data tasks into steps with inputs, tools, intermediate state, and outputs.
Uses MCP-style tool access, plugins, skills, automations, and prompt/instruction files as reusable operating components.
Uses retrieval, memory/context, background execution, and status updates to support long-running work.
Keeps human review and validation visible before decisions or generated artifacts are trusted.
How I use it
Ghost-AI uses Trigger.dev background tasks for AI design generation and spec generation.
Ghost-AI design agent updates a collaborative Liveblocks/React Flow canvas and broadcasts status/presence events.
Chorus is an AI-agent learning product concept focused on platforms, reusable skills, lessons, and saved agent workflows.
Markdown-based notes and instruction patterns are treated as workflow documentation, not hidden magic: define the task, context, review criteria, and expected handoff.
Related projects
Evidence or examples
Ghost-AI feature specs describe design-agent API wiring, TaskRun tracking, public tokens, canvas updates, AI presence, and status events.
Ghost-AI project overview defines AI generation from prompt, collaborative refinement, and Markdown spec output.
Chorus schema includes platforms, skills, lessons, and content sources for an agent-learning product.
The portfolio frames MCPs, plugins, skills, automations, reusable agent workflows, and AI operating procedures as concepts and workflow patterns, not unsupported expert claims.
Plan -> retrieve context -> call tools -> update workspace -> broadcast status -> validate output -> human review.MCP/tool access + plugin/skill instructions + markdown procedure docs + automation trigger + memory/context + approval checkpoint.Prompt -> Trigger.dev task -> OpenRouter model -> structured nodes/edges -> Liveblocks room patch -> AI_STATUS event.task.md -> objective, context, allowed tools, output format, QA checklist, human review owner, handoff notes.Generated output is treated as draft architecture or draft analysis until checked against source data and product intent.