AgentWitch Docs
Browse docs

Workflows and prompts

A workflow is a skill with several steps. The app calls it a multi-step skill. You can install ready-made ones from AgentWitch, build your own, and answer at the points where a run pauses for you. The prompt check is a separate tool that makes one prompt shorter and better. This page covers all of these.

#Install a ready-made skill

  1. Open Skills and go to the From AgentWitch tab. The page is titled "Ready-made skills".
  2. Search, or filter by Type and Publisher. "Official" is made and checked by AgentWitch. "Teammate" comes from people in your projects. Teammate skills show under "From teammates".
  3. Open a listing. "How to use" and "Before you start" tell you what it needs.
  4. Select Install. Choose the project and the computer that should run it ("Which computer should run this?").
  5. Select Start a task when the install finishes. The task window opens with the skill chosen. Fill in its form and select Start.

#Checkpoints in ready-made skills

An official multi-step skill runs as a fixed list of agent steps and human steps. AgentWitch sends one agent step at a time to your computer. At a human step the run stops and a window opens.

  • The window shows how far along you are, "What to do", and "From your assistant" with the earlier output. Type your answer under "Your reply" and select Continue.
  • Not now closes the window and keeps your place. A banner says "{skill} is waiting for your reply." with a Continue button.
  • Skip this step appears only when the step is optional.

If an agent step fails, a window says "Something went wrong". It shows "What the assistant printed" and "Technical details". Select Try again to run the same step again. Earlier answers stay. Not now snoozes it. The banner then says "{skill} hit a snag on step {step}. Your progress is saved." and offers Try again.

#Build your own multi-step skill

Open a project, then Skills. Owners select New › Multi-step skill. Members select New skill, which opens the same builder.

  1. Enter a Skill name and an optional description. The description shows in Skills only.
  2. Under "What should the assistant ask?", add questions. Each becomes a form field when someone runs the skill. Pick an Input type: Text, Paragraph, Number, Phone, Email, Link, Date, Yes / no, Choice list, or File (PDF or image). Tick Required to run where needed.
  3. Under "How this skill runs", add a Human step or a Specialist. A specialist is extra instructions copied to the computer. Steps run in order. Conditional branches are not available yet.
  4. Optionally add "Standing instructions", such as tone or format, and declared outputs. Declared outputs are not checked yet. Runs still return free text.
  5. Under "Try before you publish", fill in sample answers and select Try on computer. You do not need to publish first. Show prompt preview displays what the assistant will receive.
  6. Select Publish skill.

In a skill you build, a human step tells the assistant to stop and ask you. You answer in the task window ("Assistant needs your input"). The pause-and-resume checkpoint window above is only for the ready-made skills from AgentWitch.

Everyone in the project sees published skills. Drafts stay private to their author.

#Run a skill on a schedule

A saved skill can run by itself. Open Tasks › Schedule, select New repeating task and choose the skill in the Workflow field. See Tasks and planning for schedules and webhooks.

#Prompt check

The prompt check scores a prompt with judges and revises it until it passes (the default pass score is 70) or your limits are reached. Open a project, then Health › Prompt check. The work runs on your computer in the project's folder. Prompts and outputs stay there.

  1. Select New run.
  2. Pick a Folder. Optionally pick a "Skill to improve".
  3. Set a Goal, or use a quick fill such as "Save tokens", "Shorter prompt" or "Add guardrails".
  4. Paste the Prompt.
  5. Choose a writer for each role: Judge, Improver and Runner. Open "Instructions, pass scores and cost limits" to change limits or to stop early when scores stop improving.
  6. Select Run.

The wizard has four steps: Generalize, Evaluate, Separate and Optimize modules. At each step you select Continue. While it works you see "Tokens so far", which is an estimate. Stop ends the run. If it hits your ceiling it says "Budget exceeded. The run stopped at your ceiling."

When it finishes, "Prompt that passed" appears. Select Copy this prompt, or Download report (.md). Past runs are listed on the same tab, and you can delete them.

#Not available yet

  • Conditional branches in a multi-step skill.
  • Checked outputs for a multi-step skill.
  • A Prompt Lab that compares workflow runs side by side. It is planned, but there is no screen for it.

#Query aliases

  • multi-step skill workflow in AgentWitch
  • human checkpoint continue skip not now
  • workflow step failed try again
  • prompt optimizer prompt check
  • install ready-made skill from AgentWitch