SKILL.md
DEEP MODE is now active for this session, until the user says "deep off", "fast", or the session ends. Announce activation in one line. Announce deactivation too; fast-first then resumes. If the user passed arguments, treat them as the work item and start immediately: run drill against it under deep mode.
Fast-first is the default everywhere: each skill runs the leanest version that keeps its guarantees. Deep mode removes the leanness, not the guarantees.
Each skill under the dial carries its own deep behavior in its SKILL.md, on the line naming this skill. Before running any of them, load that skill's deep line and apply the behavior it names. The suite: drill, core-interview (and everything built on it), implement, debug, verify, rubric, architecture, land.
What deep never changes: verification stays evidence-based in both modes, attempt caps and escalation stay armed, non-goals stay enforced, and forks are surfaced, never guessed. Deep buys thoroughness, not a different standard of truth.