Repository documentation

User-directed work is durable project authority

A conversation can explain intent, but it cannot be the only place a project remembers what remains to be done. Before substantive implementation, Literate AI agents use the provider-neutral skills/agent/record-user-directed-work/SKILL.md skill to turn direction into a reasoned, evidence-gated item in docs/roadmap/active-work.md. Every litai init project receives the same skill, queue, and changelog structure.

flowchart LR
    U[User direction] --> I[Inspect current authority]
    I --> C[Reasoned conclusion and owner]
    C --> T[Unchecked durable work item]
    T --> X[Implementation from dependency order]
    X --> E[Build, test, run, or review evidence]
    E -->|fails or teaches| T
    E -->|passes| L[Update narrow spec, Flavor, or skill lesson]
    L --> H[Changelog and checked item]
    H --> G[(Git history)]

The queue is deliberately compact. It records desired outcome, conclusion, ownership, dependencies, implementation subtasks, and completion evidence. Detailed designs live in focused roadmap documents and are linked rather than copied. Git preserves prior states, so the project does not maintain a second timestamped task ledger.

This loop does not convert every suggestion directly into source. Diagnostic inspection may establish what the suggestion means; the recorded conclusion then selects the narrowest owning Component, Flavor, skill, workflow, routing policy, documentation area, or framework rule. Reusable lessons change that authority and are rebuilt. One-off model mistakes remain compact run evidence rather than becoming permanent instructions.

Secrets, worker hostnames, credentials, and raw private prompt journals never enter the queue. Test workers are supplied through ignored per-user configuration, while the work item records only the portable evidence requirement and durable report identity.