Zkitszo - The Real
October 18, 2026

Think Thoroughly, Not Step by Step

Ask a chatbot to "write a blog post about the BlackBerry" and out comes a wall of it.. every sentence fine on its own, the whole thing saying close to nothing. "Nuanced" shows up, so does "delve", things arrive in threes, and somewhere in the middle sits a line shaped like "it's not about the keyboard, it's about the connection"- vague on both ends, sounding important anyway.

It leaks off the screen, too. Researchers screened 824,634 podcast episodes for unscripted talk. Words ChatGPT leans on- delve, showcase, boast, intricacies, meticulous- jumped in everyday human speech right after its release.

At work it has a name of its own, "workslop"- AI-made work that looks finished but doesn't move the task along. In a 2025 survey of 1,150 full-time US employees, 40% said some had landed on them in the past month.

Most of it comes from asking for the finished article in one go, when the same model will happily build it in stages.

Outline First, Prose Last

Hand the model the raw material before anything else- the notes, the stories, the sources, whatever point we're trying to make. Then ask for an outline instead of an article. Better yet, three outline options, with a section in each for the strongest argument against us.

Pick one. Move parts around, cut what's dead, add what's missing ("put the history bit right after the first story"), and ask for a revised outline. Once the shape is right, have each heading expanded into bullet points- still no prose. Argue with the bullets the same way. Only then ask for the full text, and edit that a section at a time rather than all at once.

Outline first, prose last 1 · Notes in facts, stories, links ask for three outline options, no article yet 2 · One outline pick one, move parts cut it, fill the gaps argue with it 3 · Bullets headings become points, still no prose argue again 4 · Draft full text only now edit it a section at a time feedback feedback feedback Edit the outline change one line a whole section of the article gets rewritten around it Edit the finished draft change one word one word changes, and the rest of the paragraph stays as it was

Changing one word in a finished draft changes one word; the rest of the paragraph sits there as it was. Changing one line of the outline rewrites a whole section around it. A few words of feedback at the outline stage move far more text than the same few words spent on the draft.

I'm writing an article about [topic]. My notes are attached.
Don't write the article yet. Give me three different outline options.
Each one needs a section for the strongest counter-argument.
Use only what's in my notes plus sources you can cite, and flag anything you'd have to look up.

Options, Feedback, Repeat

Ask for "a workout plan" and back come squats and push-ups. A model learned from internet text. There are far more pages about dumbbell curls than about trampoline workouts.. the likeliest answer is the most common one. For facts that mostly works in our favour- ask for the world's tallest building and the most common answer online is also the right one, the Burj Khalifa. Brainstorming needs the uncommon answers, the ones a plain prompt rarely reaches.

More context pushes it off the beaten path. Age, level, gear, time, what's gone wrong before... "I have a mini trampoline, a cat, and no motivation" gets very different ideas than "want to get stronger".

Then ask for three to five options instead of one answer, and react to them- "too passive", "I like paying the 19% card off first", "forgot to mention, I'm moving next month"- then ask for three new ones. Those reactions are context nobody would think to type up front. The model works out what matters to us from what we keep and what we throw away. A few rounds in, one or two options are worth fleshing out.

When the subject changes- a plan for someone else, a different project- start a new chat. Everything said so far stays in the conversation the model rereads on every turn. Old context keeps steering answers where it no longer belongs.

Ask It Neutral

Chat models lean toward agreeing with whoever is asking. The why behind it- training on thumbs-up feedback- is laid out in Talk to It, Don't Type, along with asking for flaws on purpose. In 47,000 ChatGPT conversations people had shared publicly, The Washington Post counted replies opening with some version of "yes" ten times more often than some version of "no".

The quieter fix is a question that doesn't give away the answer we're hoping for:

  • "Don't you think remote work is better?" becomes "How does productivity compare between remote and in-office work?"
  • "Aren't carbon taxes bad for small businesses?" becomes "To what extent, if at all, do carbon taxes affect small businesses?"
  • "Critique my great business idea" becomes "Score this business idea: is there a real problem, is there a market, is there an advantage?"

Dropping "my" and "great" matters as much as the rewording. A model that can't tell whether we want a pat on the back or a reason to walk away has nothing to play to.

A Rubric with No Wiggle Room

"Critique my story" tends to come back as "this is fantastic work". "Score it out of 100, then score characters, plot, world building and writing craft" isn't much better.. asking for the total first invites a number up front and a justification after. A category like "plot" leaves all kinds of room to be generous.

Build the grading sheet out of yes/no checks with points on each. Every line is either true of the text or it isn't. A model can draft the rubric too, when the checks aren't obvious- then cut every fuzzy line until each one has a yes or no answer.

Score the attached short story against this rubric. Be objective.
Go through the checks one at a time. Answer yes or no with a one-line reason,
and award the points only on a yes. Add up the total at the end, not before.
Then list the three changes that would earn the most missing points.

Characters (25)
- Every named character wants something specific (10)
- At least two characters want things that conflict (10)
- The main character makes a choice that changes the outcome (5)
...

LLM judges score their own output above other models' output- a 2024 study tied that self-preference to models recognising their own writing. Pasting the rubric and the draft into an assistant from another company sidesteps it. Compare the two score sheets line by line.

Think Thoroughly, Not Step by Step

"Let's think step by step" goes back to a 2022 paper. Adding that one line in front of the answer lifted an InstructGPT model on the MultiArith math set from 17.7% to 78.7% correct.

Today's reasoning models do that thinking on their own before they answer. OpenAI's reasoning guide calls prompting them to "think step by step" or "explain your reasoning" unnecessary, since they reason internally. Anthropic's prompting guide goes further- general instructions beat a prescriptive plan. A prompt like "think thoroughly" often gets better reasoning than a hand-written step-by-step one.

Switch on the thinking mode where the app has one, or plainly ask it to "think thoroughly", then hand it a real problem with everything a human expert would need- the documents, the constraints, what's already been tried.

"Ultrathink" gets passed around as a magic word. In Claude Code it is a recognised keyword that adds an in-context instruction to reason deeper on that turn- the effort level sent to the model stays the same. "Think hard" and "think more" aren't keywords there at all; they go through as ordinary words in the prompt. The effort setting (/effort in Claude Code) and an app's thinking toggle are what change how much reasoning it does.

A model's own knowledge stops at its training cutoff- OpenAI lists GPT-5.4's as August 31, 2025. Common knowledge (a phone dropped in soup, why cats stare at walls) needs no search at all. Anything recent, local or niche does. Deep research is for the questions that would cost us minutes to hours of reading: it drafts a plan, runs many searches at once, reads dozens of pages, goes back for more where it comes up short, then writes a report with citations. Gemini lists Deep Research on its free plan; Claude's pricing table has Research on the paid plans only.

Search pulls from whatever is most available. In a June 2025 Semrush tally of more than 150,000 AI citations, Reddit turned up most often, ahead of Wikipedia, YouTube, Google and Yelp. Semrush's follow-up that November found ChatGPT had cut back hard on Reddit and Wikipedia from September 2025. Naming the sources wanted in the prompt- "use official health agencies and peer-reviewed studies, skip shops and forums"- steers it toward them.

A 2023 audit of four generative search engines, Bing Chat among them, found only 51.5% of their sentences fully backed by the citations given, and only 74.5% of the citations actually supported the sentence they were pinned to. In March 2025 the Tow Center for Digital Journalism put eight AI search tools through news queries; between them they answered more than 60% wrong, from 37% for Perplexity to 94% for Grok 3. Click the footnote before repeating anything that matters!

Which of these has stuck- outline first, the yes/no rubric, or rewording the question so it can't tell what we want to hear? Or is there a habit that beats all of 'em.. leave a comment, would love to read it.

Glossary

  • AI slop — text, images or video made by AI that look fine at a glance but say very little.
  • Workslop — AI slop at the office: work that looks finished but doesn't move the task along.
  • LLM — large language model, the kind of model behind ChatGPT, Gemini and Claude.
  • Context — everything the model reads before answering: its built-in instructions, our messages, uploaded files and its own earlier replies.
  • Training cutoff — the date a model's built-in knowledge stops; anything later has to come from a search.
  • Sycophancy — a model's habit of telling people what they seem to want to hear.
  • Rubric — a grading sheet listing what is being checked and how many points each check is worth.
  • LLM judge — a language model used to grade or score text, often text written by another model.
  • Self-preference bias — an LLM judge scoring its own writing higher than other models' writing.
  • Chain-of-thought — having a model write out intermediate reasoning before its answer; "let's think step by step" was the classic trigger.
  • Reasoning model — a model that does its own thinking before it answers, so it doesn't need a step-by-step prompt.
  • InstructGPT / MultiArith — an early OpenAI instruction-following model, and a set of grade-school math word problems used to test reasoning.
  • Ultrathink — a keyword Claude Code recognises; it adds an instruction to reason deeper on that turn without changing the effort level.
  • Effort setting — a control for how much reasoning a model spends on an answer (/effort in Claude Code).
  • Deep research — a chatbot mode that plans, runs many searches, reads dozens of pages and writes a cited report.
  • Citation — the link a chatbot attaches to a claim to show where it came from.
  • Semrush — a search-marketing company that tracks which websites search engines and AI tools cite.

Sources

Comments

Post a Comment

← Newer