Every non-trivial task follows the same lifecycle (defined in ~/.openclaw/platform/WORKFLOW.md):
request (Telegram) → Jarvis routes → lead writes Task Card →
specialists execute (worktree / draft) → deterministic gates →
human approval (HITL) → execute / deliver → log outcome →
distill lesson to MEMORY.md → lesson repeats >2x → becomes a skill
The Task Card
Mandatory before work starts:
| Field | Question it answers |
|---|---|
| Spec | What exactly should happen? |
| Verification | How do we check it happened? (exact commands) |
| Evaluation | How do we measure it was good after it ran? |
Self-improvement loop
Specification → Verification → Evaluation runs continuously; eval regression
gates every prompt/model upgrade; Fury reviews the whole system weekly.