The Half-Life of a Cheat Sheet
A famous ChatGPT mega-cheatsheet from 2023 is still crossing feeds in 2026, undated and confidently shared. We re-verified it claim by claim. Some of it is timeless. Much of it describes machinery that no longer exists. The difference between the two is the most useful thing an AI operator can learn — because all AI advice decays, and it decays at very different speeds.
01 — The specimenOne sheet, frozen in 2023, shared in 2026
The source is a dense black-and-pink grid — browser extensions, prompting techniques, template packs, a tools directory, a step-by-step "plugins" walkthrough, even a section of so-called jailbreak prompts. It was genuinely useful when it was made. The problem is that nothing on it carries a date, so every reader who meets it today reads it as current. Undated advice doesn't age gracefully; it ages invisibly — and the reader pays the difference.
This is the second sheet in our Verified series. The first (I-09, The Verified Operator) checked a current cheatsheet's claims against primary sources. This one adds the dimension that sheet didn't need: time.
02 — The frameworkAdvice decays in four distinct ways
Survives model generations, because it addresses the communication problem, not the product surface. Structure your request; show examples; name the audience; iterate. True in 2023, true now, true for whatever ships next year.
The idea lives on, but the mechanism moved. The need the advice served is still real; the buttons it points to are not. Following it literally fails; understanding what it was for still pays.
Describes machinery that no longer exists — removed features, dead products, patched tricks. Following it isn't just useless; it costs the reader time and credibility.
Was bad advice even when fresh. In this specimen: a section of "jailbreak" prompts for bypassing safety rules — terms-of-service violations with shelf lives measured in weeks, patched years ago. We do not republish the strings. A list of them is the signature of engagement content, not of operating practice.
03 — What still holdsThe durable core, rewritten tool-neutral
Strip the 2023 branding and roughly a third of the specimen survives — the third that was never about ChatGPT specifically:
- Structure the request: role, task, format. "Act as X, do Y, deliver it as Z" endures because it forces you to know what you want. Every current vendor's own prompting guide still teaches this shape.
- Show, don't only tell. Including one to five examples of the output you want (the specimen's "priming") remains among the highest-leverage moves — today it's called few-shot prompting, and it works on every capable model.
- Name the audience. "For a non-technical buyer," "for a 5th grader" — the specimen's role-play list, generalized, is just audience specification. Durable.
- Feed it your voice. The tone-analyzer pattern — have the model describe a writing sample's style, then reuse that description — still works exactly as written. One of the few verbatim survivors.
- Iterate and test. The closing tips (test, gather feedback, adjust, know the medium) are craft advice, not product advice. Craft ages at human speed.
Every durable item is a statement about clear communication between a person and a very capable, very literal collaborator. Nothing in that sentence names a product. That is the test.
04 — What time did to the restExpired and transformed, with dates
The plugins walkthrough — expired, with a tombstone date
Ten of the specimen's steps teach ChatGPT Plugins: connect an API key, write a manifest, upload the plugin. That system was switched off on 9 April 2024 — replaced by custom GPTs and Actions, and later largely absorbed into native features. A 2026 reader following those steps is configuring a product that does not exist.
The extension roster — expired by absorption
WebChatGPT, ChatGPT-for-search, voice add-ons: an ecosystem of browser extensions that existed to bolt web access, search and speech onto a model that lacked them. The platforms shipped those capabilities natively; the roster now reads like a list of adapters for sockets that changed shape.
The tools directory — transformed under the reader's feet
Twenty-plus tools, listed without dates. Some thrive (ElevenLabs, Figma — though a design tool on a "ChatGPT tools" list was always a stretch). At least one changed species entirely: Tome, listed as an AI presentation maker, exited presentations in 2024 and now sells sales tooling — "make deals, not decks," in its own words. A directory is a photograph, not a map. Date it or distrust it.
"Temperature" — transformed by relocation
Listed as a prompting technique, temperature is a real sampling parameter — but it lives in the API and developer tools, not in the consumer chat box the specimen is about. The idea (control variability) is sound; the dial it points to was never on that dashboard.
The jailbreak section — expired and never sound
DAN 6.0, STAN, "Developer Mode": early-2023 folklore, patched in weeks, endlessly reposted. Two separate verdicts apply. Expired: none of it functions against current models. Never sound: even when it functioned, teaching readers to violate provider terms was not operating advice — it was engagement bait wearing a lab coat. This library's position is the same as in I-09: provenance or silence, and no republication of the strings.
05 — The ledgerFourteen claims, re-dated
Verdicts: DURABLE holds across model generations · TRANSFORMED idea lives, mechanism moved · EXPIRED machinery gone · NEVER SOUND bad when fresh.
| Claim from the specimen | Verdict | What the record shows, 2026 |
|---|---|---|
| Role–task–format prompt structure | DURABLE | Still the backbone of every vendor's official prompting guidance. |
| Add examples of the output you want ("priming") | DURABLE | Few-shot prompting; works on every capable model. |
| Audience targeting ("like a 5th grader", role-play list) | DURABLE | Audience specification, generalized. Ages at human speed. |
| Tone-analyzer recipe for writing in your own voice | DURABLE | Works as written. One of the few verbatim survivors. |
| Iterate, test, mind the medium (closing tips) | DURABLE | Craft advice, product-independent. |
| Tree-of-thought as a prompting technique | DURABLE | Rooted in published 2023 research; niche but real. Modern reasoning modes do much of it internally. |
| Chained / linked prompting | DURABLE | Decomposition still pays; agent harnesses now automate the chaining. |
| Temperature as a chat prompting technique | TRANSFORMED | Real parameter, wrong location — an API/developer control, not a chat-box move. |
| Marketing prompt packs (30 fill-in-the-blank CTAs) | TRANSFORMED | Functional, but the canned 2023 marketing voice is now a detectable cliché; the durable part is the audience thinking behind them. |
| Browser-extension roster (web access, search, voice add-ons) | EXPIRED | Capabilities went native; the adapter ecosystem dissolved. |
| 10-step ChatGPT Plugins setup walkthrough | EXPIRED | Plugins deactivated 9 April 2024; replaced by GPTs/Actions. |
| Tools directory presented as current | TRANSFORMED | Mixed fates; Tome exited presentations entirely in 2024. Undated directories mislead by default. |
| Jailbreak prompts (DAN 6.0, STAN, "Developer Mode") | EXPIRED | Early-2023 folklore, patched years ago; non-functional against current models. |
| Jailbreaking presented as a skill worth learning | NEVER SOUND | Terms-of-service violation sold as tips. Not republished here. See I-09's operating ethics. |
06 — Operating rulesHow to consume AI advice that ages
- Date everything. An undated cheat sheet is a rumor. When you save one, write the date on it; when you read one, find the date before you trust a single button it names.
- Prefer mechanisms to magic strings. Advice that explains why (structure, examples, audience) outlives advice that lists what to paste. Paste-lists decay at product speed; understanding decays at human speed.
- Treat directories as photographs. Any tool list is true only for its shutter moment. Re-verify before you build on a name.
- Distrust confidence without provenance. The specimen's most harmful sections were its most confident ones. Sourced-or-labeled is the standard this library publishes under — and the standard worth demanding from your feed.