Skip to the content.

Getting started

Five minutes, and you need to be a Jira administrator for the first two steps.


1. Get an API key

Premortem has no AI of its own. It calls the provider you choose, with your key, and the tokens land on your bill. Pick one:

Provider Where the key comes from Sensible first model
OpenAI platform.openai.com gpt-5.5
Anthropic console.anthropic.com claude-sonnet-5
Google Gemini aistudio.google.com gemini-2.5-flash
An OpenAI-compatible endpoint yours — but see the note below whatever it serves

An OpenAI-compatible endpoint of your own works only on a domain the app has declared: Forge lets an app reach nothing it did not list in advance, so a new domain means a new app version. Ask us before you rely on one.

If you have no preference, use whichever provider you already have an account with. On OpenAI the default is gpt-5.5, the model every published measurement ran on; gpt-4.1-mini is cheaper but has not been measured. An assessment is a few thousand tokens; you are not going to be surprised by the bill on a trial. For scale: the Marketplace reviewer who tested the whole app end to end, on our key, spent $0.65 in total.

2. Put the key into Premortem

Jira → Settings (⚙) → Apps → Premortem settings.

Under AI model, choose the provider, paste the key, and leave the model field empty unless you want a specific one — empty means the default above.

Press Save, then Test connection. It calls the provider once and tells you whether the key works. Do this now rather than finding out from a failed assessment later.

The key is stored in Forge’s encrypted secret storage, never in plain settings, and it is never shown back to you — only the last four characters, so you can tell which key is loaded.

3. Run the first assessment

Open any Jira issue with a real description. In the right-hand column, expand the Premortem panel and press Assess risks.

Pick a real ticket for the first run, not a one-line one. The assessment can only reason about what the issue says: a title with no description produces risks that could apply to any project, and that is the app working correctly on thin input rather than the app being empty. An issue with a description, a parent epic and a few comments is the honest test.

It takes a few seconds. You will get risks grouped by kind, each with:

The list keeps the order the assessment wrote it in, with a count of high, medium and low at the top of each group. The ratings are there to read, not to sort by: when we measured it, sorting by them pushed the risk that actually came true further down the list, not up.

Each risk has five buttons: Accept, Mitigating, Dismiss, Ask AI, and This happened.

4. Decide what to keep

This is the part that matters. The model proposes; you decide. Nothing is written anywhere until you accept it.

If the assessment missed something, write it into the field under the list and press It missed one. It becomes a precedent for similar work too.

Then press Write accepted. Two checkboxes decide where the accepted risks go: the issue description, the Confluence register, or both.


Where to go next

Read the settings reference before turning things on. Most of the context switches are off by default because each one adds tokens to every assessment — on your key.

If the assessments feel generic, that is the expected starting state: the app knows nothing about your systems yet. Two things fix it, in this order:

  1. Point it at your architecture page — one Confluence link, one minute, and it is the single biggest quality jump available. See settings.
  2. Import your past incidents and postmortems — so risks cite what actually happened to you before, by issue key or by postmortem.