The problem it solves: AI assistants lose context between sessions and drift from your decisions as context windows fill up. Prothon addresses this with three ideas:
1. A three-level doc hierarchy (SPEC, DESIGN, PATTERNS) where each level scopes a single concern and higher always overrides lower. The AI implements from these docs, not from memory. Each doc-writing step is a dedicated agent that refuses content belonging at another level.
2. Automated agents that keep things consistent: one cross-checks the three docs for contradictions, one generates reference skills from live documentation for your stack, one verifies code implements what the docs describe.
3. A "promise" system for execution. Before each task, the agent declares what files it will touch, success criteria, and expected line counts. The promise is verified against the actual git diff. This makes context drift measurable rather than something you notice after the fact.
CLI:
``` prothon new # scaffold from template prothon init # add to existing project prothon spec # requirements Q&A prothon design # architecture Q&A prothon patterns # conventions Q&A prothon execute # plan, promise, build prothon compliance # verify code vs docs ```
Built on copier, so existing projects can pull upstream template improvements. Currently targets Claude Code, designed to be agent-agnostic.