AI prompts vanish faster than you think—here's the timeline
The cursor blinks in the prompt box. You’re rebuilding yesterday’s GPT-4 instruction from memory because you closed the tab, cleared the cache, or just switched tools. Somewhere between your screen and the model’s server, your best work expired on a timer you never saw.
AI writing assistants cache your prompts—but the clock runs out faster than you expect
Most AI platforms store recent prompts to speed up responses, but retention windows vary wildly by tool and plan.
ChatGPT keeps conversation history until you delete it manually or hit organisation retention policies. Claude caches prompts for five minutes on the API—long enough for a follow-up, not long enough for lunch. Gemini’s context window lives for the session, then vanishes when you close the browser. If you’re using API wrappers like Zapier’s ChatGPT action or Make’s OpenAI module, the cache sits with the platform, not the model, and those logs disappear on their own schedule—often thirty days or fewer.
This matters when you’re iterating on a product description, refining an email sequence, or testing headline variants. If you don’t save the working version before the cache expires, you’re rebuilding from scratch. The friction doubles when you switch tools mid-project: a prompt optimised for GPT-4 rarely transfers cleanly to Claude or Gemini, and the cache won’t follow you. Most operators discover this the hard way—after they’ve lost the version that actually worked.
The fix is version control for prompts, not memory. Plain-text files, a shared Google Doc, or a dedicated prompt library all work. The key is timestamping iterations and tagging which model and parameters produced the result. That way, when the cache expires—or when you need to reproduce output six months later—you’re not guessing.
TACTIC
Version your prompts before you lose the ones that worked
Solo operators iterate AI prompts fast—tweaking tone, structure, and constraints until output clicks—but rarely save the working drafts. A week later, you remember the result but not the instruction that produced it, and the tool’s history has cycled out. Prompt versioning doesn’t require new software; a timestamped text file or a dedicated Notion page is enough. The trick is logging model, temperature, and max tokens alongside the prompt text, so you can reproduce results when cache and memory both fail.
WORTH READING
Token counting is inconsistent—and you’re paying for the difference
AI platforms charge by the token, but the count displayed in your dashboard rarely matches what the API logs or what your bill reflects. Some tools count whitespace and formatting; others strip it. System messages and function calls sometimes appear in usage reports, sometimes don’t. The gap becomes expensive when you’re processing thousands of requests a day. Auditing token usage means comparing three sources: the platform’s usage page, the API response headers, and your invoice line items. The discrepancies reveal where you’re being billed for overhead you didn’t expect.
FROM THE ARCHIVE
When automation steps fail, retry logic decides whether your workflow heals or quits
Most workflow platforms retry failed steps automatically, but the rules vary by tool and error type. Zapier retries HTTP 500 errors but not 400s. Make retries three times by default, with exponential backoff. Integromat (now Make) used to skip retries on certain webhook timeouts entirely. If you’re chaining AI API calls—prompt to summary to formatted output—one transient failure can cascade through the sequence. Understanding which errors trigger retries and which abort the run helps you design workflows that recover gracefully instead of failing silently at 3 a.m.
Know someone who would like this? Forward today’s email—every operator we reach is one closer to running an online business with a little less friction.