Custom workflows

Tenant-defined linear chains of prompt steps. Build at /admin/legal/custom-workflows. The same engine that runs the Playbooks — playbooks are just code-curated versions of what you can author here.

Anatomy

A custom workflow is up to 8 steps, each with:

label:          short name shown in the runner
system_prompt:  the system message for that step
user_template:  body of the user message, with placeholders
input_kind:     "text" | "vault_doc"   (step 1 only)

Placeholders

{{input}}      the input the runner provided
{{step_1}}     output of step 1
{{step_2}}     output of step 2
...

Step i can reference {{step_1}}…{{step_(i-1)}} plus {{input}}.

Example

name: "Deal-close memo generator"
description: "From a closed-deal summary, produces a partner-facing memo."
steps:
  - label: "Extract milestones"
    system_prompt: |
      You are a corporate paralegal. Extract every closing milestone
      from the input. Output JSON: [{date, milestone, status}].
    user_template: "{{input}}"
  - label: "Risk surface"
    system_prompt: |
      You are M&A counsel. Given milestones below, flag any open
      risk items requiring follow-up. Output bulleted markdown.
    user_template: "MILESTONES:\n\n{{step_1}}"
  - label: "Memo draft"
    system_prompt: |
      Compose a one-page partner-facing memo summarizing the deal
      close. Include the milestones, the risk surface, and a
      recommended next-steps section.
    user_template: |
      DEAL SUMMARY:\n{{input}}\n\nMILESTONES:\n{{step_1}}\n\nRISKS:\n{{step_2}}

Run it from the same page — paste a deal summary, get the memo.

Why trace

Every step’s LLM call is stamped on the workflow’s Why trace as a distinct llm_call event. Per-step output also persists to legal_custom_workflow_runs.steps_output_json.

Limits

8 steps per workflow (hard cap)
Input text capped at 80k chars
Names tenant-unique