# Workflow

Plan → Prompt → Implement → Check → PR → Review → Fix → Validate

When the current phase is unclear, invoke `saf-route`. It recommends the next skill and points to that skill's `SKILL.md`; it does not invoke skills or change files. `saf-validate` is the terminal public verification capability.

## Where to read next

| Step | Doc |
| --- | --- |
| Full workflow with agents | [saf-skills-usage-guide.md](saf-skills-usage-guide.md) |
| Skill reference | [skills-catalog.md](skills-catalog.md) |
| How skills are invoked | [invocation-model.md](invocation-model.md) |
| Execution modes | [execution-modes.md](execution-modes.md) |
| Autonomy levels | [autonomy-levels.md](autonomy-levels.md) |
| Prompt patterns | [prompt-recipes.md](prompt-recipes.md) |

## Rules of thumb

Create specs before implementation. Resolve configuration and package identity before choosing a route. Use `saf-brainstorm` only when the problem, outcome, or consequential product direction still requires a decision; use `saf-create-spec` when the outcome is bounded enough to produce observable requirements, even when acceptance criteria or technical details remain open. Preserve an explicit package or task identity and inspect its prerequisites before treating it as ready.

Implement one task at a time unless `saf-implement-multi` has an explicit dependency graph and isolation boundary. Run `saf-check-task` on task evidence before accepting work. Run `saf-validate` after integration. The routing corpus and its structural check document expected routes; they do not prove host or model behavior.

Use the TLC baseline for planning and the [TDD baseline](tdd-baseline.md) for code tasks. Confirm required behavior and contractual seams. Work in vertical slices. Record current sensor evidence (or an explicit gap). RED is optional and diagnostic; do not fabricate it.

Use multi-task orchestration only after explicit planning identifies independent paths and dependencies. The host may provide isolation; Git worktrees remain an optional fallback. Change-review skills extend the loop; they do not replace task checking.

Use `plan`, `guided`, `apply`, `review`, and `full` to select the level of local work. These modes preserve human approval and Git/release boundaries; see [execution modes](execution-modes.md).