Zkitszo - The Real
October 19, 2026

Sloppy Jalopy- When a Trusted Prompt Goes Sideways

Further to what had been talked about in Think Thoroughly, Not Step by Step- I mean, I could talk about this all day. Doom scrolling? When I am on a roll, I feel like I've gotten lost in my creativity- my brain can't come up with ideas fast enough- even finding myself frustrated- I want to create something epic- this is the opportunity- then, like a brick to the face (pretty harsh eh? I know.. because it is... harsh.).

What I Pick Is What I Tip

Everything will be going so smooth and I am checking off every little app idea, every simple game- sharing with friends and the like- they love it.

Due to my accident, food prep is difficult.. it's more a cop out- expensive.. we all need an excuse to be lazy.. no no.. truly, it's difficult for me to do certain tasks- plus dangerous. I order DoorDash- I tip. Most of my girlfriends worked in the restaurant industry... I understand that everything is expensive and ok ok, a good tip brings that lil extra to the day- when the Dasher delivers according to directions, I don't mind tipping, tipping generously- sometimes even tearing up pieces of paper, writing down $ amounts ranging 10%- 20% with a few outliers, a couple low ball tips and some high rollers... I pull the piece of paper outta a hat which I had shaken up in front of the Dasher, explaining the simple game- what I pick, is what I tip.

Yes, I've had a couple which were like $50.. $75.. maybeee $100... The Dasher is pretty taken aback- I am true to my word.... I've had repeat Dashers show up some time later- a different delivery- and they are glowing- can't wait to express their gratitude for that last tip... my issues with poor Dasher deliveries have gone down. It's not every time/ all the time. But, it is some days....

So, I made an app- a simple web app- no logins or accounts, just a fun game to play with the Dasher, though one needs to be loyal and pay out the value picked. A good portion of the time it's like that..

DoorDash says Dashers get 100% of the customer tips it receives. Cash at the door is fair game too.

Next prompt... maybe it's after that usage limit time- a prompt regularly used.. results predictable and desired- you start to review the result and shake your head- disbelief- What? Had? Happened? That is impossible.. everything is a mess... filled with AI sloppy jalopy- "Fable?! Are you drunk??"

Maybe, having gotten comfortable with a regular, familiar prompt, you forgot that a certain part was key and trimmed the fat yourself. Or that usage reset, which already felt like an eternity- your AI agent seems to run some fancy calculation determining what section of the prompt gets tossed.. even if it was important to you.

Goal, Input, Output

An app that small can come out of a single prompt. Three building blocks carry most of it- the goal (what the thing is for), the input (what the person using it types or taps) and the output (what it shows back). The ideas that come out working on the first go share a shape- one well-defined task, nothing to upload, no outside data to fetch.. open it, use it for a minute, close it. A Pomodoro timer, a bill splitter, a flash-card quiz. A hat full of tips fits right in with that crowd. A multiplayer game played over the internet sits a whole level harder.

Shaped like the tip hat, the whole prompt fits in a few lines:

Build a single-page web app. No logins, no accounts, no server- everything runs in the browser.

Goal: a tip game to play at the door with a delivery driver.

Input: the order total; a low and a high tip percentage (default 10% and 20%); how many slips go in the hat; how many low-ball and how many high-roller slips to add.

Output: a hat that shakes when tapped, then draws one slip at random and shows the tip as a percentage and as a dollar amount, big enough to read at arm's length.

Built as a Claude artifact, sharing it gets complicated- a new artifact shares by link only on a paid plan. Opening one takes an account too- "Everyone needs a Claude account".. a no-login app with a login in front of it. The same code saved as a plain .html file opens in any browser.

Brief It Like a New Hire

Everything the model reads before it replies is its context- the vendor's system prompt, descriptions of its tools, our files, the whole chat so far. A one-line prompt leaves our part of it nearly empty, so the gaps get filled with the most common answer going.

What the model reads, every single turn CONTEXT WINDOW System prompt the vendor's: date, model name, house rules Tool descriptions web search, code, files... how to use each Saved instructions project instructions, CLAUDE.md Files we hand it docs, screenshots, spreadsheets Chat history every message so far, its replies too This prompt the newest message Model Reply added to the history LONG SESSION, WINDOW NEARLY FULL Chat history hours of messages, every detail compacted Summary early details can drop out

Picture handing these exact words to a smart new hire who knows nothing about us- would they do a good job? If not, add what that person would have to ask for. Typing is where a lot of those details fall off- one more reason to talk to it instead of typing.

Give it the reason along with the rule. Anthropic's prompting guide shows a bare "NEVER use ellipses" doing worse than the version with the why- the reply gets read aloud by a text-to-speech engine that won't know how to pronounce them. With the reason, the model can work out the related rules nobody wrote down.

A workout plan for mom, asked for in the same chat that built ours, drags our schedule and preferences along with it. An unrelated question gets a fresh chat.

"Fable?! Are You Drunk??"

Anthropic's docs describe nothing like a calculation that picks part of a prompt to throw out. A handful of documented things end up in the same place, though.

Long sessions get squeezed

When Claude Code's context window fills up, it compacts on its own- older tool output gets cleared, then the conversation gets summarized. The docs say it straight:

Claude compacts automatically, but instructions from early in the conversation can get lost.

A key instruction typed three hours back may live on as one line of a summary... or not at all. The project's CLAUDE.md file is the exception- Claude Code re-reads it from disk after a compaction. Compacting by hand takes focus instructions, so the summary keeps what matters:

/compact keep the slip rules and the default tip range

The middle of a long prompt gets less attention

The 2023 "Lost in the Middle" study found models do best when what they need sits at the beginning or the end of the input, and "significantly" worse when it's buried in the middle. A 2025 benchmark piled up to 500 instructions into one prompt- the best frontier models managed 68%. Most misses were omissions, instructions simply not followed, with a lean toward the early ones. A prompt that grows a line at a time over weeks pushes its oldest rules toward the middle.. from the outside, that looks a lot like a fancy calculation tossing the part that mattered.

A different model reads differently

Anthropic says its latest models are trained for "precise instruction following", and that Claude Sonnet 5 "interprets prompts literally and explicitly, particularly at lower effort levels." A prompt that leaned on the model guessing what we meant can fall apart the day the model or the effort setting changes.

The usage limit isn't one of those changes. Paid Claude plans have a session limit that resets every five hours plus a weekly one. Claude Code's model docs say rate-limit and billing errors "never trigger a switch" to a fallback model.

Sometimes it really is them

To state it plainly: We never reduce model quality due to demand, time of day, or server load.

That's Anthropic's September 2025 postmortem, which also owns up to three infrastructure bugs that intermittently degraded Claude's responses between August and early September- after users started reporting worse answers in early August. So "are you drunk??" isn't always in our heads.. though a prompt retyped from memory, or trimmed to save tokens, is the cheaper suspect to check first.

Keep the Key Part Where It Can't Fall Off

A prompt that matters lives in a file- not in our head, not in yesterday's chat. Paste it fresh. Keep dated copies when it changes, so a weird result can be checked against the last version that worked.

Rules that belong in every chat go where they get re-read. In the Claude app that's a Project- its instructions apply to every chat inside it. Free accounts get up to five. In Claude Code it's CLAUDE.md. Keep both short, for the reasons in Usage Limit Reached? Not So Fast..- every line gets read on every turn.

Long pasted material goes at the top, the question at the bottom- Anthropic's guide says "Queries at the end can improve response quality by up to 30 percent in tests." The rule that can't be dropped goes at the start or the end, never wedged in the middle of the notes. Fence the parts off with XML tags (plain labels in angle brackets) so they don't blur together:

<notes>
(everything pasted in goes here)
</notes>

<rules>
Keep the default tip range at 10% to 20%.
Every slip shows a percentage and a dollar amount.
</rules>

Using the notes, rebuild the slip screen. Follow every rule.

Say what to do, more than what not to do- Anthropic swaps "Do not use markdown in your response" for "Your response should be composed of smoothly flowing prose paragraphs." Image generators with a separate negative prompt box are the exception, built for the don't-want list.

Typos Welcome.. Mostly

Plenty of prompting advice says typos don't matter- "can you cook eggs in microwave" gets about the same answer as the tidy version, since models learned from a web full of misspellings. Good news for anybody typing on a phone in a hurry!

A 2024 study doesn't fully agree. Grammar mistakes barely moved its results, but performance was "severely influenced by typographical errors"- tested on ChatGPT-3.5 and two small open models, old by now. The scarier 20% to 33% drops quoted from PromptRobust (first published as PromptBench) come from errors engineered to break the model, not slips of a finger.

A typo in the chatty part of a prompt costs little. A typo in a name, a number, a file path or the key rule changes what's actually being asked.

Has a prompt you trust ever come back drunk? What turned out to be behind it- the prompt, the model, or a session that ran too long? Leave a comment.. would love to hear 'em.

Glossary

  • Prompt — the message we type or paste to an AI model.
  • Context — everything the model reads before it replies: the vendor's system prompt, tool descriptions, saved instructions, our files, the chat so far and the newest prompt.
  • Context window — the most context a model can hold at once, measured in tokens.
  • Token — the chunk of text a model reads and writes in; a common word is often one token, a misspelled one can split into several.
  • System prompt — instructions the vendor puts in front of every chat (the date, the model's name, house rules) before our message arrives.
  • Compaction — squeezing a long conversation into a summary so it fits back in the context window; early details can get lost.
  • CLAUDE.md — a plain text file of standing instructions that Claude Code reads at the start of a session and re-reads after a compaction.
  • Claude Code — Anthropic's coding agent that works on files and runs commands on our computer.
  • Project (Claude app) — a folder of chats that share the same instructions and uploaded files.
  • Artifact — a page, app or document Claude builds in a panel beside the chat.
  • Effort level — a setting for how much thinking the model puts in before answering; lower is faster and more literal.
  • Fallback model — a second model an app switches to when the first one is overloaded or unavailable.
  • Usage limit — the cap on how much a subscription plan can use; on paid Claude plans, a five-hour session window plus a weekly limit.
  • Frontier model — one of the newest, most capable models from the big labs.
  • Omission error — an instruction the model simply didn't follow, as opposed to one it followed badly.
  • XML tags — labels in angle brackets, like <notes> and </notes>, used to fence off parts of a prompt.
  • Negative prompt — a separate list, in some image generators, of things the picture should not contain.
  • Pomodoro timer — a work timer that splits work into timed blocks (25 minutes in the classic version) with short breaks between.
  • Web app — a program that runs in a web browser instead of being installed.
  • Dasher — DoorDash's name for its delivery drivers.

Sources

Comments

Post a Comment

← Newer