workflow
Installation
SKILL.md
Workflow — run a dynamic workflow
Orchestrate a goal end-to-end: plan → route → compose skills/tools/MCPs → gate each step → verify → fail safe. The engine for speed + quality. Pattern: references/sops/dynamic-workflows.md. Named workflows: references/sops/wf-*.md.
How to run
- Resolve the goal. Named workflow (
/workflow investor-update) → load itswf-*.mdcard. Free-form goal → first run/structureto get Context · Outcome · Goal · Success criteria, then plan the steps. - Check the autonomy gate per step. Internal/reversible/no-PHI → may run unattended. External/irreversible/regulated → stop and have the user approve (draft-never-send; never to a sensitive recipient without approval).
- Execute step by step. For each: run the action (a
/skill, atools/*.py, or an MCP call) → run its gate/verify check → branch on the result (e.g. missing data → flag + ask, don't guess). Reuse existing capabilities; don't rebuild. - Fail safe. On error: read it, retry/repair if trivial, else escalate to the user with what's blocked. Never silently drop a step.
- Verify done. Confirm the success criteria are met (evidence, not assertion) before reporting complete.
- Deliver. Internal → your sanctioned cloud / chat. External → draft to the user. Log meaningful decisions to
decisions/log.md.
Build / refine a workflow
New recurring multi-step goal that composes ≥2 capabilities and has a definable "done"? Author a references/sops/wf-<name>.md card (format in dynamic-workflows.md) and add it to the inventory. Don't workflow-ify one-offs (Simplicity). Don't overwrite a workflow card without asking.
Guardrails
All of CLAUDE.md. Especially: the autonomy gate per step, quality gate before "done", cite-don't-invent, no confidential data persisted, draft-never-send external.