uranium328/claude_project_template

This repository is a template package of governance files meant to let a lower tier AI model, something like a Sonnet or Haiku level model rather than a top tier one, reliably take over and run a software project on its own. It is built for Claude Code or any similar setup that uses a CLAUDE.md file along with a subagent mechanism. The package includes a README that can be deleted once it is deployed, a template CLAUDE.md file kept deliberately short, and a folder called ops containing several documents. These cover a first time bootstrap process that inventories the environment and checks for three common failure points, rules for how work gets assigned to different models and when to escalate or step down in model tier, a rubric for judgment calls such as when to escalate, when a task counts as done, when to ask a human, and where the quality floor sits, five templates for common types of work like searching, implementing, refactoring, researching, and reviewing, a maintenance protocol covering who is allowed to change what and how lessons learned get folded back in, a letter meant for whoever picks this project up next, and an empty log file for recording pitfalls as they come up. The core design idea is that the main CLAUDE.md file should act as a short router pointing to detail elsewhere, not a full reference on its own, staying under 60 lines. Every rule is meant to be concrete enough for a less capable model to follow exactly, since anything written abstractly is treated as if it were not written at all. The whole system is designed to run using models around the Sonnet tier, with the harder, more subjective calls, like resolving an unclear requirement or making a taste based judgment, handled by an explicit escalation path rather than left to guesswork. The author is upfront that this template cannot solve ambiguous problem definitions or subjective taste calls on its own, and that it was written outside the target project, which is exactly why the bootstrap step exists: to check real environment facts on site rather than assume anything from the template.