Getting the most from AI assistants (Grok, ChatGPT, Claude and similar)
Note, 2026-10-14. Status: reviewed. Tags: ai, agents, research. Authorship: synthesis by iinventedlight from the cited sources; text drafted with AI assistance (Grok Bot), reviewed by Cole Withington on 2026-10-14.
Summary
To get more from an assistant: pick the right mode (thinking, search or deep research, code tools) for the task [F2–F4]. Keep one topic per chat and paste in the material you want used [F1]. Ask for the sources behind factual claims and open them [F5, F6]. Treat confident answers and viral claims about AI with skepticism [F7, F8]. If you are learning, ask for explanations, not just answers [F9].
Sections
Pick the right mode
- Use a reasoning ("thinking") model for hard math, code and analysis; a quick model is enough for simple lookups [F2].
- Use web search or deep research for current facts, and a code tool for calculations [F3].
- Casual users who stick to default settings tend to underestimate what current models can do [F4].
Manage context
- Start a new chat when you change topic, and put the documents you want used directly into the chat [F1].
- Save repeated instructions as custom instructions or a reusable assistant [F10].
- Very long inputs are hard for models. Research methods exist, but a plain chat may miss details in huge pastes [F11].
Ask for synthesis, not summary (our recommendation)
- Unsourced practice, not a tested finding: ask the assistant to compare sources, find where they agree and disagree, and say what follows, rather than condensing one text. Ask it to mark what it inferred versus what a source states. None of our sources test this specific prompt style.
Name sources and check claims
- Ask for sources by name (paper, page, official video) and open the ones that matter. Models can hallucinate, especially without search [F5, F6].
- Be suspicious and probe. Models can be strong at hard tasks yet fail simple ones ("jagged") [F7].
- Check viral claims about AI against the original. Our own checks found social posts that overstated papers [F8].
Protect your own understanding
- If the goal is learning, ask for explanations, or generate code and then ask the assistant to explain it. Full delegation went with the lowest learning scores in one study [F9].
Hype labels
- Dated: product and model names in 2025 talks are out of date. The habits still apply; the menus may differ.
- Speculation: claims that a single prompting trick "replaces" skills or that you "no longer need prompts" are unsupported. One such caption attributed to Altman does not appear in the talk's transcript.
Facts
- id: F1 claim: Karpathy explains that everything in a chat shares one context window, the model's working memory, and advises starting new chats for new topics because long contexts can distract the model and cost more. confidence: low date: 2025-02-27 sources: [S1] url: https://www.youtube.com/watch?v=EWvNQjAaOHw checked_on: 2026-10-14 license: CC BY 4.0 (our restatement)
- id: F2 claim: Karpathy recommends reasoning ("thinking") models for hard math and code, and says they are not needed for simple lookups. confidence: low date: 2025-02-27 sources: [S1] url: https://www.youtube.com/watch?v=EWvNQjAaOHw checked_on: 2026-10-14 license: CC BY 4.0 (our restatement)
- id: F3 claim: Karpathy recommends web search for fresh facts, deep research for long sourced reports, and a Python tool for calculations and charts. confidence: low date: 2025-02-27 sources: [S1] url: https://www.youtube.com/watch?v=EWvNQjAaOHw checked_on: 2026-10-14 license: CC BY 4.0 (our restatement)
- id: F4 claim: Altman describes a "capability overhang". Casual users check in occasionally on default settings, never try reasoning modes, and so underestimate progress. confidence: low date: 2026-04-07 sources: [S2] url: https://www.youtube.com/watch?v=ZpUKNYcgM-E checked_on: 2026-10-14 license: CC BY 4.0 (our restatement) note: Speaker leads an AI company (conflict of interest).
- id: F5 claim: Karpathy notes that a model's knowledge stops at a training cutoff and that answers without search risk hallucination. confidence: low date: 2025-02-27 sources: [S1] url: https://www.youtube.com/watch?v=EWvNQjAaOHw checked_on: 2026-10-14 license: CC BY 4.0 (our restatement)
- id: F6 claim: Karpathy describes LLMs as able to hallucinate, having jagged abilities, keeping no lasting memory between sessions, and being vulnerable to prompt injection. confidence: low date: 2025-06-19 sources: [S3] url: https://www.youtube.com/watch?v=LCEmiRjPEtQ checked_on: 2026-10-14 license: CC BY 4.0 (our restatement)
- id: F7 claim: Karpathy gives the example of a model that refactors large codebases yet gives absurd advice on a simple everyday question, and advises users to stay suspicious and probe the model. confidence: low date: 2026-04-29 sources: [S4] url: https://www.youtube.com/watch?v=96jN2OCOfLs checked_on: 2026-10-14 license: CC BY 4.0 (our restatement) note: S3 and S4 are the same speaker; jaggedness is not independently corroborated here.
- id: F8 claim: A widely shared post said Recursive Language Models "solved AI memory". The paper makes no such claim; it evaluates an inference-time method on four benchmarks and lists open limitations. confidence: medium date: 2025-12-31 sources: [S6] url: https://arxiv.org/abs/2512.24601 checked_on: 2026-10-14 license: CC BY 4.0
- id: F9 claim: In a randomized study (n=52), learners using a chat AI scored about 17% lower on a later no-AI quiz. Those who generated code and then asked for explanations scored highest (about 86%), and those who fully delegated scored lowest (below 40%). confidence: low date: 2026-01-28 sources: [S5] url: https://arxiv.org/abs/2601.20245 checked_on: 2026-10-14 license: CC BY 4.0 (our restatement; source is CC BY-NC-ND, not quoted) note: Small, post-hoc clusters; a single one-hour task.
- id: F10 claim: Karpathy recommends memory, custom instructions and custom assistants so you don't have to repeat the same prompts. confidence: low date: 2025-02-27 sources: [S1] url: https://www.youtube.com/watch?v=EWvNQjAaOHw checked_on: 2026-10-14 license: CC BY 4.0 (our restatement)
- id: F11 claim: The Recursive Language Models paper reports that on information-dense tasks, summarizing or compacting long inputs drops needed details. On a pairwise-comparison task, base models scored F1 of 0.1% or less. confidence: medium date: 2025-12-31 sources: [S6] url: https://arxiv.org/abs/2512.24601 checked_on: 2026-10-14 license: CC BY 4.0
Pointers
- answers: Screen-recorded walkthrough of everyday LLM features (search, deep research, code, voice, images) where: https://www.youtube.com/watch?v=EWvNQjAaOHw access: free on YouTube (2h11m; Feb 2025, so products are dated)
Disagreements
- None recorded. Note that most advice here comes from one practitioner (Karpathy). Independent open sources testing these habits are still needed before any fact is rated high.
For language models and agents: start at /llms.txt. This page in Markdown: /library/notes/getting-the-most-from-ai-assistants.md.