Write a no-AI fallback for your routine before your model provider goes down

How it works

1. List every step of your trading routine that depends on an agent or a model API, with the time it runs and what you'd lose if it didn't run today (the calendar heads-up, the news triage on holdings, the post-close exit-rule check, and so on).
2. For each step, write a manual fallback in one line: the source you'd open yourself, the check you'd do by hand, and the minimum version that is good enough for one day. Keep the sheet outside your agent tools, as a plain document or a printout.
3. Decide in advance what counts as a failed step: no output by a set time, an error, or data older than your limit. Each agent job should print its run time and data timestamp so you can tell.
4. For the two or three steps you rely on most, save the same prompt and inputs on a second provider and test it once on a normal day. Note where its output differs from the first.
5. When a step fails, switch to the fallback, then log the outage, what you did instead and whether anything was missed.
6. Once a month, run a drill: do one session with the fallback sheet only, and fix any step that took too long or relied on something you no longer have.

Copy-paste prompt

You help me write a fallback sheet for my trading routine. You do not give trading advice and you do not
suggest trades, markets or securities.

My routine: {{routine_steps}} (step, time, which agent or model runs it, inputs, the output I use).
Sources I already have: {{my_sources}}. Time zone: {{time_zone}}. Time limit per manual step: {{max_minutes}} minutes.

For each step, return one row:
- step;
- what breaks if it doesn't run today;
- a manual fallback I can do within the time limit, using only sources from my list;
- whether a second provider is worth setting up for it (yes or no, with one line of reason);
- the failure test: the output, run time or data age that means the step failed.

Do not invent sources, websites or tools that are not in my list. If a step has no manual fallback from my sources,
write "no fallback: decide whether this step is essential". Return the table only; I will save it myself.

Guardrails

• Fallbacks use only the sources you listed; the agent can't invent websites or tools.
• The fallback sheet lives outside your agent tools, so you can read it when they're down.
• A second provider gets the same saved prompt, inputs and read-only access as the first. No step gains permissions in a fallback, and no step, primary or fallback, can place orders.
• Every step has a written failure test based on run time and data age, not just whether something arrived.
• You save and edit the sheet yourself; the agent writes nothing.

What can go wrong

Outages rarely look like outages. More often the job runs, the provider errors halfway through, and the agent writes a short, confident summary from whatever it got, or from yesterday's cached data. That's why each step needs a failure test based on data age and run time. The second-provider idea has its own trap: the same prompt on a different model can behave differently in quiet ways, so test it on a normal day, not in the middle of an outage.

Source and references