Product Creation Toolkit is a research-stage scaffold for a guided agent workflow toolkit: an opinionated, human-controlled method that helps a product owner move through six explicit product-creation phases with durable artifacts, evidence labels, decision records, and reopenable gates.
Status: Research and handoff only. This repository contains no working toolkit, package, CLI, MCP server, generated adapter, or supported host integration yet.
The future toolkit is intended to let an agent facilitate a defined workflow while deterministic software owns state, validation, and file operations. The accountable human remains responsible for passing, looping, stopping, or reopening each gate.
It is intended to be:
- one opinionated six-phase product-creation method;
- explicitly invoked by a person through a host-specific skill or command;
- local-first, with authoritative run artifacts stored beside each product;
- resumable from files rather than hidden chat state;
- available through a deterministic core, local store, CLI, and stdio MCP adapter;
- distributed through generated Agent Skill and command adapters that are tested per host.
It is not:
- a working implementation today;
- an autonomous product strategist or gate decision-maker;
- a general workflow engine, agent SDK, or workflow DSL;
- a universal plugin that behaves identically in every agent host;
- a web application, hosted service, remote MCP service, or cloud-sync product;
- evidence that a product idea has been validated merely because documents were generated.
The tentative architecture separates concerns:
- Methodology: the six phases, evidence labels, artifacts, gates, and human authority rules.
- Deterministic core: future state transitions, invariants, validation, migrations, provenance, and stable diagnostics.
- Local store: future path-confined, versioned, auditable run state and human-readable artifacts beside each product.
- CLI: future direct interface for people, scripts, and hosts with shell access.
- stdio MCP adapter: future typed, capability-limited access to the same core.
- Generated host adapters: future Agent Skill and explicit-command packages derived from one canonical instruction source.
These are design intentions, not implemented capabilities.
Agent Skills can carry broadly reusable instructions, MCP can expose broadly reusable operations, and a CLI can provide a stable local fallback. Host discovery paths, explicit invocation syntax, permission prompts, supported MCP primitives, cloud filesystem access, and packaging still vary.
Support must therefore be claimed for a tested host surface and version, not for every model or product using the same underlying model. A cloud agent also cannot access a laptop-local run or stdio process unless the artifacts and executable are present in its own environment.
The first implementation should prove only the Phase 1 vertical slice end to end: create and resume a run, preserve the raw prompt, record labeled claims and notes, validate the Phase 1 artifact contract, preview gate readiness, record a human decision, and reopen safely through the core, local store, CLI, stdio MCP tools, and one explicit Tier 1 host adapter.
No web UI or hosted service belongs in v1.
- Fresh-agent handoff
- Agent instructions
- New-chat starter prompt
- Six-phase product-creation framework
- Architecture and portability research
Implementation choices intentionally remain unresolved where current standards, support floors, and security policy must be reconciled first. Start with HANDOFF.md, not with package initialization.
Licensed under the MIT License.