Introducing recursive-mode

recursive-mode is an installable skill package for structured AI-assisted software development. It keeps requirements, plans, implementation, testing, review, wrap-up, and memory as a file-based workflow rather than chat, so that work context doesn't disappear even after a session ends.

The core problem is context rot. In long agent tasks, requirements, decisions, and plans remain only in the chat window and get lost when a session ends or the context window is exceeded. To solve this problem, recursive-mode treats static documents inside the repository as the single source of truth, and reduces prompts to short commands.

The key to this approach is recursion. Each development phase produces a single locked artifact, and the next phase uses the previous phase's artifact as input. Also, before a phase ends, it must satisfy exit criteria defined based on the workflow and prior documents; if it falls short, draft → audit → repair → re-audit is repeated until it's actually brought to a ready state.

The documents that get saved are organized as follows.

• .recursive/RECURSIVE.md: the baseline spec for the entire workflow
• STATE.md: the current repository state
• DECISIONS.md: a record of decisions
• run/…/: requirements, as-is analysis, to-be plan, implementation summary, test summary, manual QA, and addenda for individual tasks

Global STATE.md and DECISIONS.md are read and updated at the start and end of each run. This ensures reproducible traceability for both humans and agents at any time, instead of context scattered across chat.

The author sees this documentation as more than just record-keeping. The documents are readable by both humans and machines, and in the long run, along with code diffs, they can also be used as a dataset for fine-tuning or self-distillation. In the end, what matters is not the conversation but the evidence left in the repository.

The way it operates is also clear. Chat is used not as a place to carry on complex conversations but like a CLI for issuing commands, while detailed planning is managed through plan mode and documents. When needed, a working run folder and requirements document are created first, and the agent is instructed briefly, like "implement run 01."

This package includes the following subskills.

• recursive-mode: phase-by-phase orchestration, locked artifacts, durable memory
• recursive-worktree: separating implementation work with a dedicated git worktree
• recursive-debugging: structured debugging including Phase 1.5
• recursive-tdd: TDD that records RED/GREEN evidence
• recursive-review-bundle: packaging delegated reviews into reproducible bundles
• recursive-subagent: managing subagent handoff contracts and self-audit fallback

Installation can be done as the full package or as individual subskills. The author also describes this as an alternative to Missions, and emphasizes that it is open source and works with any IDE, CLI, agent, or model.