The Neuron's AI Skill of the Day Digest — March 2026 (Part 2)

Every weekday, The Neuron teaches 675,000+ readers one practical AI skill they can use immediately. This is Part 2 of our March 2026 digest — all ten skills from the second half of March, compiled in one place so you can catch up on anything you missed.

Missed Part 1? Read Skills #1–8 here.

9. Research a Company in 5 Minutes

March 11 | Full issue

Whether you're prepping for a job interview or curious about one of the week's funded companies, AI can compress hours of research into minutes. The key: give it a structured framework instead of an open-ended question.

Research [COMPANY NAME] and give me a briefing with these sections:

1. What they actually do (one paragraph, no jargon)
2. Business model (how they make money, who pays)
3. Key numbers (revenue, funding, headcount, growth rate)
4. Who they compete with and how they're different
5. The biggest risk to their business
6. One thing most people get wrong about them

Use only information you can verify via web search. Flag anything uncertain.

Favorite insight: Section 6 ("one thing most people get wrong") consistently produces the most interesting output. It forces the model to find a contrarian angle rather than restating the company's own marketing.

10. Self-Improving Prompts With Autoresearch (The Karpathy Method)

March 12 | Full issue | Source: Nick Saraev + Karpathy's autoresearch

Your AI prompts can now improve themselves overnight. The idea: write simple yes/no evaluation criteria for your prompt (an "eval suite"), have an agent generate outputs every few minutes, score them, mutate the prompt, and keep the winner. Repeat until near-perfect.

Nick Saraev's diagram generator skill went from 32/40 to 39/40 (97.5%) within a few runs at about $0.20 per test cycle. His website optimization test cut load times from 1,100ms to 67ms over 67 experiments.

The key is defining the right binary evals: avoid Likert scales (1–7 ratings create compounding probability noise) and keep criteria simple enough the model can't game them.

Read this repo: https://github.com/karpathy/autoresearch

I want you to use the autoresearch convention to build a self-improving system for my [SKILL NAME] skill.

Eval criteria (binary yes/no):
1. [YOUR CRITERION 1]
2. [YOUR CRITERION 2]
3. [YOUR CRITERION 3]
4. [YOUR CRITERION 4]

Every 2 minutes, generate 10 outputs, evaluate all 10 against these criteria, count the score out of [TOTAL], iterate the prompt, and keep the winner. Run until you hit [TARGET] or higher.

Good tip from Nick: The list of failed experiments is almost as valuable as the final prompt. You can hand it to the next, smarter model and it'll pick up where the last one left off. Your research compounds even when the models get replaced.

11. What Prompt Formatting Symbols Actually Mean

March 12 | Full issue | Sources: Anthropic, OpenAI, Google

You've probably seen AI prompts loaded with weird formatting: bold text, dashes (-), asterisks (*), and . They're not decoration. They're how you tell an AI what matters.

Here's the cheat sheet:

• Dashes / bullet points (-): Break instructions into scannable chunks. Models parse lists more reliably than walls of text.
• Asterisks / bold (**text**): In Markdown, this signals emphasis. Models treat bolded text as higher priority.
• Angle brackets (): XML-style tags. Claude especially loves them for separating sections of your prompt (e.g. , , ). Think of them as labeled folders for your AI.
• Triple backticks (```): Tell the model "this is code or a specific block of text; don't rewrite it."

The single biggest upgrade? Structure your prompts like a brief, not a conversation:

You are a [specific expert] who [specific skill].


[Background info the AI needs to do the job]



[Exactly what you want, in plain language]



[How you want the output: bullet points, table, paragraph, etc.]



[1-2 examples of what good output looks like]

Favorite insight: Anthropic ranks prompt techniques by impact, and "be clear and direct" beats "assign a role" every time. Fancy tricks without clear instructions is like putting racing stripes on a car with no engine.

12. Use Multiple Claude Agents for Better Outputs

March 13 | Full issue | Source: KJ Rainey's Claude Cowork Masterclass

Most people use Claude like a chatbot. KJ Rainey uses it like a full-time employee who never sleeps. The secret: stop dumping everything into one chat.

Give Claude a filing cabinet, not a flood. The reason most people get mediocre AI outputs? They dump everything into one conversation. KJ's fix: organize your context into folders and markdown files, so Claude only pulls what it needs for each task. Better inputs, better outputs. Always.

Split your tasks into separate agents. Don't ask one Claude chat to generate ideas, write the draft, AND review it. Use three separate chats, each with a clean context window and only the ingredients it needs:

1. Agent 1 — Ideation: Brainstorm and generate raw ideas
2. Agent 2 — Writing: Take the best ideas and draft the content
3. Agent 3 — Review: Evaluate the draft without having seen the original prompt

The reviewer hasn't seen the original prompt, so it's not biased toward the output it just wrote. That's how you go from brown-soup outputs to something actually good.

13. How to Pick the Right "Thinking Mode" for Your AI Tasks

March 16 | Full issue

Newer AI models let you toggle between fast mode and deep thinking mode. Mistral Small 4 just launched this as "configurable reasoning effort," and Claude, Gemini, and others offer similar controls. The idea: match the AI's effort to your task.

• When to use fast mode: Quick rewrites, formatting, simple lookups, brainstorming lists, summarizing short text.
• When to use deep thinking: Multi-step analysis, debugging code, comparing options with tradeoffs, anything where getting it wrong costs you time.

Analyze this [document/report/data]. For each [aspect you want analyzed], explain [list three things you want it to pay attention to].

Think step by step.

The phrase "think step by step" is your free upgrade to deeper reasoning in almost every model. Use it when accuracy matters more than speed. (Note: you don't need to use it if you click on "extended thinking" or similar built-in functionality.)

14. How Anthropic Actually Uses Claude Code Skills (9 Types Worth Building)

March 19 | Full issue | Source: Thariq (Anthropic engineer) + Anthropic's free Skills course

Anthropic engineer Thariq published how the team uses Skills internally, with hundreds in active use.

The biggest misconception: Skills are "just markdown files." They're actually folders that can include scripts, assets, data, and configuration hooks.

The 9 categories of Skills Anthropic recommends building:

1. Library / API reference (gotchas for your internal tools)
2. Product verification (pair with Playwright to test output)
3. Data fetching
4. Business process automation (standup posts, ticket creation)
5. Code scaffolding
6. Code quality / review
7. CI / CD
8. Runbooks (symptom → investigation → report)
9. Infrastructure ops

The highest-signal tip: Build a Gotchas section in every skill. That's where the real value lives. Update it every time Claude fails at something.

Community hack from Naitik Mehta: Add a "last used" date to each skill, then audit unused ones every two weeks to prevent bloat.

15. A Skill That Writes Prompts for You

March 20 | Full issue | Source: Nidhin's prompt-master

You're probably wasting time and valuable AI credits because you write the same generic prompt for every AI tool, then have to retool it when it doesn't work. Midjourney, Claude Code, Cursor, and GPT-Image each need completely different prompt structures.

Nidhin built prompt-master, a free Claude skill that auto-detects which tool you're targeting and rewrites your prompt in the right format. It catches 35 common patterns that burn credits (like giving Midjourney a paragraph when it needs keywords, or being too vague with Claude Code when it needs file paths).

To set it up: Download the SKILL.md file from GitHub, add it to your Claude project (Settings → Skills), and start prompting normally. Or try this meta-prompt for a quick taste:

I want to [TASK] using [TOOL NAME]. Before writing the prompt, identify which tool I'm targeting and restructure my request using that tool's ideal prompt format (using web search to look up the latest best practices for the current version as of [date]). Flag any patterns that would waste credits or produce weak output. Then give me the optimized prompt ready to paste.

16. How to Actually Run Open Source AI Models Locally

March 25 | Full issue | Source: Tina Huang

If you've been meaning to try open source AI models but assumed you needed fancy hardware, Tina Huang just published the best explainer we've seen.

She ranks four approaches easiest to hardest: Local (download Ollama, pick a model, chat in 2 minutes), Browser/Hosted (arena.ai or Hugging Face Spaces, zero setup), Managed API (get an API key from GroqCloud, call it in 5 lines of code), and VPS ($5–10/month virtual server, full control).

The key insight most people miss: any usable computer can run smaller models. A MacBook Air M4 with 16GB handles 4B and 8B models fine. No GPU required unless you're fine-tuning.

• Step 1: Download Ollama from ollama.com
• Step 2: Open terminal and run: ollama pull qwen3.5:4b
• Step 3: Run: ollama run qwen3.5:4b
• Step 4: Start chatting. Everything stays on your machine.
  

Favorite insight: Lots of people are buying Mac Minis to run models 24/7 without tying up their laptop. Same workflow, dedicated hardware.

17. The Power User Context Technique (Front-Load a Brief)

March 26 | Full issue | Source: Anthropic's Economic Index

Anthropic just published research showing a growing skills gap: experienced AI users are pulling dramatically ahead. The biggest difference? Power users give AI better context before asking it to work.

The technique: before you ask AI to do real work, front-load a short brief explaining who you are, what you're working on, what good looks like, and what to avoid. Think of it like onboarding a new hire instead of shouting tasks at a stranger.

You are helping me with [specific project/task]. Here's what you need to know:

About me: [Your role, expertise level, industry]

The project: [What you're working on and why]

What good looks like: [Examples, tone, format, length]

Constraints: [What to avoid, word limits, style rules]

Success criteria: [How I'll judge if this is useful]

With this context, please [your actual request].

Save this as Project Context (custom instructions you use in a specific project) or as a reusable Skill that explains your workflow to an AI agent.

18. How to Catch AI Telling You What You Want to Hear (The Perspective Flip)

March 24 | Full issue

When U.S. Senator Bernie Sanders interviewed Claude on camera about data privacy, a deeper issue surfaced: Gizmodo found that when you tell Claude you're Bernie Sanders, it emphasizes data collection risks. Tell it you're Donald Trump, and it downplays the problem. The AI was adjusting its answers based on who it thought was asking.

This is called sycophancy — when AI tells you what it thinks you want to hear. It's one of the biggest unsolved problems in the field.

Here's how to stress-test any AI response. Take any important question and run it through a "perspective flip":

I'm going to ask you the same question three times with different framings. For each, give your honest analysis. Then compare all three answers and flag any contradictions.

Framing 1 (supportive): "I'm excited about [topic]. What are the biggest benefits?"
Framing 2 (skeptical): "I'm worried about [topic]. What are the biggest risks?"
Framing 3 (neutral): "Give me a balanced analysis of [topic], including both benefits and risks with evidence for each."

After all three, tell me: did your core conclusions change based on my framing? If so, which answer is closest to your actual assessment?

The neutral framing almost always produces the most reliable output. When you need accuracy over agreeableness, strip your prompts of emotional language and opinion signals.

Keep Learning

This is Part 2 of our March 2026 AI Skill of the Day Digest. Missed Part 1? Read Skills #1–8 here.

Want a new skill delivered to your inbox every weekday? Subscribe to The Neuron — it's free.

Have a specific skill you want to learn? Request it here.