Skip to main content
Build and run

Build Automations with Your AI Assistant

The fastest, most accurate way to build an automation is to ask the AI assistant you already use. Your assistant is connected to StackJack, it can set everything a person can set in the automation…

Written By Christopher Scaminaci

Last updated 27 days ago

The fastest, most accurate way to build an automation is to ask the AI assistant you already use. Your assistant is connected to StackJack, it can set everything a person can set in the automation builder, and it does it in minutes instead of a long run of clicks. You describe the job in plain English, it works out which systems and tools are involved, writes the configuration, tests it and shows you the result. You check it in the portal before anything goes live.

This page explains why that is the recommended path, how to install the StackJack automation builder skill that makes it smooth, and gives example requests you can copy.

  • It can set everything the builder can. Every setting in the advanced builder can be read and written through your assistant: the tools, the trigger, the fixed steps, the knowledge, the limits, even editing a live automation. See Building and Maintaining Automations from Your AI Assistant for the tool-by-tool reference.
  • It is faster. One request replaces opening five sections and picking tools one at a time. Your assistant searches the whole tool catalog for the right tools, and it can look at your real data before it designs anything.
  • It is more accurate. It reads the exact shape of each tool and each setting before it writes, so it does not guess a name or a field. It tests the automation and reads the run record to see what actually happened, then fixes what is wrong.
  • You stay in charge. The things that matter are yours to decide: your assistant asks before it accepts the automation's consent, and before it takes the automation out of test mode. You can open the automation in the portal at any point and check what it can do. The portal builder is still there, and so are the guided wizard and templates, if you prefer them.

What you need

  • A connected assistant. Connect one by following Connect Your AI. Sign in with your own account when you connect it: automation changes are refused unless StackJack can tell which person is making them.
  • The automation builder tools granted to that connection. If your assistant says a builder tool is missing, open the endpoint's grant page in the portal, grant the automation tools, and refresh your assistant's tool list. A connection whose grant was a hand-picked set of tools keeps exactly that set until you grant the new ones. See Building and Maintaining Automations from Your AI Assistant.
  • The systems connected. The assistant can only use the connectors you have already set up. See How Connectors Work.
  • Credits, or your own Anthropic key. Test runs and live runs use credits, or your own key if you have added one.

Install the builder skill

The skill is a small folder of instructions that teaches your assistant how to interview you, look at your systems, build, test and improve an automation. It contains no code and no credentials.

Download stackjack-automation-builder.zip

The zip holds one folder, stackjack-automation-builder, with a SKILL.md file and a references folder of longer notes. Keep the zip as it is.

Claude on the web and the Claude desktop app

  1. Turn on code execution first. Skills need it. On Free, Pro and Max plans, open Settings > Capabilities and switch on Code execution and file creation. On Team and Enterprise plans an owner enables code execution and skills in the organization settings.
  2. Go to Customize > Skills.
  3. Press +, then + Create skill, then Upload a skill.
  4. Choose the downloaded zip. The skill appears in your list. Make sure it is switched on.

If the upload option is missing on a Team or Enterprise plan, an owner may have turned off user-created skills. Anthropic describes these steps in Use skills in Claude.

Claude Code

Unzip the download into your personal skills folder, so the file ends up at ~/.claude/skills/stackjack-automation-builder/SKILL.md:

unzip stackjack-automation-builder.zip -d ~/.claude/skills/

On Windows PowerShell, run Expand-Archive .\stackjack-automation-builder.zip -DestinationPath "$HOME\.claude\skills". Claude Code picks up changes to an existing skills folder without a restart. If ~/.claude/skills/ did not exist when your session started, run /reload-skills. Then ask for it by name or type /stackjack-automation-builder. See Extend Claude with skills.

Other assistants

Skills are an Anthropic format, and not every assistant reads them. If yours has no skills feature, open SKILL.md in a text editor and paste it into the assistant's custom instructions or project instructions, then ask it to follow the skill. We have not tested this with every assistant, and an assistant that limits how long instructions can be may not accept the whole file.

Check that it works

Start a new conversation and say: "Use the StackJack automation builder. What could I automate?" The assistant should start by checking that StackJack is connected and the builder tools are available, then ask you a question.

What the skill does

  1. Checks the setup. It confirms the StackJack connection and the builder tools. If one is missing it tells you exactly what to grant.
  2. Interviews you, one question at a time. The outcome you want, which of your connected systems to use, what starts it (on demand, a schedule or a webhook), what it may change, who is told, and how you will know it worked.
  3. Finds things to build, if you ask. It reads your connected systems, never changes them, and proposes a short ranked list of automations with the evidence behind each.
  4. Designs it. It chooses the narrowest set of tools that does the job and shows you the plan in plain words.
  5. Builds it. It creates the automation as a draft in test mode, adds fixed steps and knowledge if the job needs them, and checks it is ready.
  6. Tests it. It runs the automation, reads what happened and explains it. Writes are blocked in test mode, so nothing in your systems changes. It asks before it switches on a schedule or a webhook for a test.
  7. Goes live only on your word. It summarizes what the automation can change, points you to the portal to check, asks for your acknowledgment where one is needed, and only then takes it out of test mode.
  8. Suggests improvements after a few runs: a tighter tool set, a schedule, a quiet stop when there is nothing to do, a lower cost.

Example requests

Say these to your assistant with the skill installed. Change the names, times and systems to match yours.

1. Aging tickets with device alerts (HaloPSA and NinjaOne)

Every weekday at 8am, find HaloPSA tickets that are open and have had no update for 3 days. For each one tied to a device, check NinjaOne for active alerts on that device, and add a private note to the ticket with how long it has been idle and any open alerts. Do not change status or assignment.

2. A critical Huntress incident becomes a ticket and a page (Huntress, HaloPSA and PagerDuty)

When Huntress reports a critical incident, open a HaloPSA ticket for the right client with the summary and the recommended action, then page the on-call engineer in PagerDuty with a link to the ticket. If a ticket for that incident already exists, add a note instead.

Your assistant will ask how Huntress alerts reach StackJack: a webhook fed by your alert pipeline, or a check every 15 minutes. Paging someone changes a live system, so it asks for the destructive-action acknowledgment before this one goes live.

3. Licences you pay for against licences in use (Pax8 and Microsoft 365)

On the first of every month, compare the Microsoft 365 licences we buy through Pax8 with the licences actually assigned in each client's tenant. List clients with more than 10 percent unused, and licences held by disabled accounts. Put the report in a HaloPSA ticket for the account manager.

4. Backup failures, one ticket per client (your backup product and HaloPSA)

Every morning at 7, collect backup failures from the last 24 hours, group them by client, and open one HaloPSA ticket per client with the failing devices and errors. If that client already has one open for today, add a note instead. Skip clients with no failures.

5. Let your assistant find the work (read only)

Look at my last 30 days of HaloPSA tickets and propose the five automations that save my techs the most time. For each one give the evidence, the systems, the trigger, what it would change, the hours it would save per month, and the exact request I should make to build it. Do not change anything.

6. Triage new tickets that need a judgement (HaloPSA)

Every 15 minutes, look at the HaloPSA tickets that are still in the New status. If there are none, stop. Otherwise decide whether any of them describes an outage that affects more than one person, and if so add a private note that says so and lists similar open tickets for the same client. Skip a ticket that already has that note.

A step cannot work out "since the last run", so this one filters on something that changes once the work is done, here the New status. Pick a status, queue or assignment that a handled ticket leaves.

7. An offboarding check (Microsoft 365 and HaloPSA)

When I give you a departing employee's email address, check that their Microsoft 365 account is disabled, list the licences still assigned, and list their open HaloPSA tickets. Give me a checklist of what is left. Do not change anything.

8. A client summary before a review (HaloPSA, NinjaOne and Huntress)

When I name a client, pull their last 90 days: ticket counts by type and the slowest resolutions, devices that are offline or out of date, and any Huntress incidents. Write a one-page summary for the quarterly review.

What makes a request work well

  • Say what starts it, which systems it uses, the rule, and the result you want.
  • Say what it must not change. "Read only" and "do not change status" are clear instructions.
  • Give a real example to test with, such as a ticket number or a client name.
  • Start read only. Once a test run looks right, ask for the version that writes.
  • For work that should stop when there is nothing to do, say so: "if there are none, stop". The assistant adds a check on the step that reads, and a run that finds nothing ends as Skipped at no AI cost. See Stop when a step finds nothing.
  • For a real judgement, such as how urgent something is or which category it belongs in, the assistant can add a decision step. A decision step judges only the values it is handed, so the assistant gives it the exact facts it needs: a count, one field of one ticket, or a short list of small rows, never a whole page of ticket rows. An input that is too big, or that points at something the earlier step did not return, makes the step fail, and the run ends as Failed unless the step is set to carry on. See Decision steps.

Testing and going live

An automation your assistant creates always starts in test mode, also called dry run. In test mode the automation reads your systems for real, and every write is blocked and recorded as "would have called", so nothing changes in your tools. A test run uses credits (or your own Anthropic key), the same as any run.

  • The assistant runs it and reports back. It starts the run, waits for it to finish, reads the record of what the automation did, and explains it in plain words. If something is wrong, it fixes the cause and runs it again.
  • A schedule or a webhook does not start by itself. Testing starts a run on demand. A test needs the automation switched on, so for a scheduled or webhook automation your assistant asks first, tells you it will then run on its schedule or accept deliveries and cost credits each time, and switches it off again right after the test unless you say otherwise. If you have no preference, start with an on-demand automation and add the schedule once a test run looks right.
  • The consents are yours. Before a test can run, the automation needs its general consent. Your assistant shows you the exact sentence and what it covers, tells you the acceptance cannot be withdrawn through the assistant (archiving the automation ends it), and records nothing until you say yes. If it can use tools that delete or change things, it also needs the destructive-action acknowledgment before it goes live, asked the same way. It will not accept either one on its own. See Guardrails and Safety.
  • You verify it in the portal. Open Automations, then the automation, and check the trigger, the tools it can call, its fixed steps and its limits. The run history shows every test run. The portal is at portal.stackjack.io, or portal-eu.stackjack.io for EU accounts. See The Automation Detail Page and Run History.
  • Going live is your decision. When you tell it to, the assistant takes the automation out of test mode. For a schedule or a webhook it then asks a separate question before it switches the trigger back on, because that is what makes it run for real from then on: it tells you how often it will run or what starts it, and that each run uses credits or your own Anthropic key. If you say no, the automation stays live but switched off, and you can switch it on later from the portal or by asking your assistant. If you kept the trigger on after the test, the assistant says so when it asks to take the automation out of test mode, because every run the trigger starts from then on is real. You can switch it off again at any time the same way. Each saved change is kept as a version you can put back; see Version History and Rollback.
  • It keeps improving. After a few real runs, ask: "How is it doing, and what would you change?" The assistant looks at the recent runs, which tools were never used, what each run cost and how often it was skipped or failed, and suggests changes. It makes none without your say-so.