What it does
THROUGHLINE is a spec-driven development framework for AI coding agents. It is a folder of markdown prompts that you copy into a project, with git as the memory, so there is nothing to install: no runtime, CLI or dependencies. It works with Claude Code, Codex and Cursor, and ships in English and Korean. The problem it targets is an agent forgetting earlier decisions, so a later session overwrites something agreed in an earlier one. Four prompts (kickoff, adopt, develop and audit) generate and maintain files such as an architecture contract, a plan, a progress log, a history and notes, so the next session can pick up where the last one stopped. Personas review each spec before code is written, and an audit prompt checks for drift between documents and code. A separate edition covers teams where several developers and agents build the same codebase together. The repository includes two small benchmark pilots that the author describes as showing direction rather than magnitude. In one, the framework's record kept an early decision intact across seven sessions. The README also lists limits: it cannot force an agent to follow the rules, and the documentation upkeep may not pay off for small one-off tasks.
Capabilities
- Markdown and git only, with nothing to install
- Prompts for new projects, existing codebases, ongoing development and audits
- Progress and history files that carry decisions between sessions
- Persona review of each spec before code
- Audit prompt to catch drift between documents and code
- Solo and team editions, in English and Korean
When to use it
- Your AI-assisted project runs over many sessions and decisions keep getting lost
- You use Claude Code, Codex or Cursor and want a shared written contract for them
- You want to add a spec and memory structure to an existing codebase
