rlm-workflow arrives as a Skills.sh skill that translates the Recursive Language Models paper's 10M-token context technique into a kanban-style development pipeline. The MIT research showed sub-agents could shuttle information between a chat window and an external store; this implementation takes a different angle — specifications never enter the conversation. Chat becomes a command interface, not a context vessel.

The workflow moves sequentially through five locked phases: requirements, as-is analysis, to-be design, implementation summary, and manual QA. Each phase consumes the previous phase's markdown output and emits its own gated document. A user creates 00-requirements.md in an rlm/ folder, then invokes the agent. The run pauses at manual QA for human sign-off before proceeding. Phase outputs are immutable once written, preventing the drift that plagues long chat histories.

Two global artifacts anchor traceability: DECISIONS.md logs every implemented requirement with rationale and links to its phase docs, while STATE.md maintains a living architectural overview. The folder structure mirrors the workflow — rlm/00-my-feature/ holds the five phase files plus an optional addenda directory. No vector database or codebase indexing is required; the documents themselves provide progressive disclosure that steers the model.

Early testing suggests the approach cuts token spend and eliminates context rot by keeping specs out of the conversation window. The self-documenting pipeline also produces stakeholder-ready artifacts without extra effort. For teams drowning in rework from degraded chat contexts, this trades conversational flexibility for deterministic, auditable execution — a deliberate constraint that may prove more valuable than raw context length.