AI Skill of the Day Digest — August 2026 (Part 2)

Every day, The Neuron teaches readers one practical AI skill they can use immediately. This second August collection includes every published AI Skill of the Day from August 16 through August 31, 2026.

Skim the headings, grab the prompts that fit your work, and return when you need a better way to use Claude, ChatGPT, Gemini, Codex, or an AI agent.

How to use this digest

• Skimming? Each entry starts with the practical outcome.
• Trying one? Copy the included prompt or workflow and replace the bracketed details.
• Catching up? The skills are ordered by their original newsletter date.

🎓 August 16

Run Qwen 3.8 on Your Own Computer

So the new open model Qwen3.8-27B just dropped on Friday, and Unsloth (the team that makes open AI actually usable on YOUR computer) lets you run it locally using a compressed version called a “quant.”

So all you have to do is download it, and the model runs locally on your computer (meaning not over the cloud). That means more privacy for prompts and files, and more importantly, no per-prompt API bill.

If you don’t know, Qwen is local devs’ favorite coding agent. So if you want to experiment with coding AI for free, this is the way.

Here’s the, uh, *beginner-friendly way to actually do that with Unsloth:*

• 1. Check your computer’s memory. Unsloth counts total RAM + VRAM, or unified memory on a Mac. More memory = you can run a higher-quality version of the model. 17GB+ is the sweet spot.
• 2. Install Unsloth, search for Qwen3.8-27B, and pick a “quant,” basically a compressed version of the same 27B model. The absolute smallest is UD-IQ2_XXS at ~9GB. But if your computer can handle it, start with UD-Q4_K_XL at ~18GB for a much better quality/size balance (that’s what Unsloth runs).
• 3. Download it, load it in Chat, and start prompting. If performance is rough, choose a smaller quant. If you’re on the smallest already, try something like Gemma 12B instead; it just won’t be as smart, so expect it to screw up.

Follow Unsloth’s full Qwen setup guide here.

🎓 August 17

Fine-tune your own AI model

Sometimes a general-purpose AI knows plenty, but doesn’t behave the way you need. Maybe you want a model that writes like your company, understands a specialized workflow, or tutors you at exactly your skill level.

That’s what fine-tuning does: you take an existing open-source model (meaning a model you can download, run, and customize yourself) and train it on examples of the answers you want, essentially teaching it new habits without building an AI from scratch.

Unsloth Studio makes the technical part much easier. Think of it as a point-and-click workshop for customizing AI models on your own computer: choose a model, feed it examples, train it, test the results, and export your custom version.

The basic workflow:

1. Install Unsloth Studio, then pick a model your hardware can handle.

1. Build training examples with an instruction, input, and ideal output. You can generate these with AI or even turn a PDF into a custom dataset.

1. Train the model with QLoRA, which teaches a small adapter instead of retraining the entire model.

Then comes the important part: compare your fine-tuned model against the original to see whether your data actually made it better.

The secret sauce is giving a smaller model really good examples of exactly what you want it to do. Basically: start with a smart generalist, then teach it to become your specialist.

🎓 August 18

Give AI a Design System, Not Vibes

AI website builders produce generic pages when your prompt contains goals but no visual rules. Vercel’s design-system guidance recommends giving the model reusable brand context instead: colors, fonts, spacing, components, and reference blocks.

Before asking for a page, provide three things:

1. Brand tokens: exact colors, typefaces, spacing, and corner radius.

1. Reference assets: screenshots, logos, or an existing page that feels right.

1. Component rules: which buttons, cards, navigation, and layouts it should reuse.

Build this page using these brand rules: [tokens]. Match these references: [assets]. Reuse these components: [list]. Before coding, summarize the visual system and flag any missing decisions.

🎓 August 19

Give Codex a 1M-token context window

OpenAI’s Tibo shared a config that apparently gives OpenAI’s current best AI model, GPT-5.6 Sol, a one-million-token context window in Codex (OpenAI’s coding app, available in the ChatGPT Desktop app).

A one million token context window means the model can keep far more code, tool output, and chat history in view before compressing older material. Tokens are the chunks of text AI counts, in case you missed that part (we got you, newbie friends!)

See, GPT-5.6 Sol supports 1.05M tokens, but Codex keeps a smaller default tuned for performance and cost. So use this only for unusually large codebases or long debugging runs.

Here’s what you do:

Open ~/.codex/config.toml and add these settings at the top, before any [section] headers:

model = "gpt-5.6-sol"
modelcontextwindow = 1000000
modelautocompacttokenlimit = 900000

This selects Sol, sets the context budget to 1M tokens, and starts compaction at 900K to leave headroom. Restart Codex and start a new session. Tibo says ChatGPT-account sign-in now supports the override too.

For a one-off CLI session:

codex -m gpt-5.6-sol \
-c modelcontextwindow=1000000 \
-c modelautocompacttokenlimit=900000

🎓 August 20

Make Cursor auto-fix new PR feedback

Cursor’s new Subscriptions let cloud agents monitor a pull request (a proposed code change) after they create it, wake back up when CI checks fail or a bot leaves feedback, and keep working without a fresh prompt.

Try this on your next PR:

1. Ask a Cursor cloud agent to make the change and open a pull request.

1. Then set one finish line and walk away. Cursor automatically subscribes to PRs its agents create.

/goal keep this PR merge-ready: fix failing CI checks and bot review comments until everything passes.

Non-coder trick: steal the same “new feedback → unresolved work” loop. Keep one chat for a launch or project. Each time feedback arrives, paste it in and ask: “Compare this with the last round. Show only what is still unresolved and the next action.”

The useful bit isn’t “use an agent\.” It’s letting one job wake back up when the thing it is responsible for changes.*

\well, if you’re comfortable, you really should be using agents lol. For that, click this.*

🎓 August 21

Turn a Chat Into a Live Website

You can turn an idea, draft, or compatible local project into a hosted website without leaving ChatGPT. ChatGPT Sites can create and refine the site, save reviewable versions, deploy a live URL, and add storage, sign-in, analytics, collaborators, or a custom domain. One important detail: deployment URLs are production, so if you want to review first, ask Sites to save a version before deploying. Brent Schooley’s Before the Cut is a live example.

1. In ChatGPT, include the word “website” in your request or mention @Sites.

1. Describe the audience, the job the site should do, and the information or features it needs. Ask ChatGPT to save a version first if you want to review it before publishing.

1. When it looks right, ask Sites to deploy it and give you the production URL. Keep refining the project conversationally.

Copy this:

Build a website for [audience] that helps them [job]. Include [sections/features]. Save a reviewable version before deploying. Once I approve it, publish it with Sites and give me the live URL.

Bonus: so Brent’s GPT site was all about how to edit video with Codex without making it guess. Brent’s field guide uses four questions before the first cut:

1. What are we making? Lock the audience, story, target length, format, and feeling before asking Codex to edit.

1. What are we working with? Inventory camera angles, screencasts, audio, graphics, templates, and brand assets. Have Codex flag picture/sound problems, unreadable demos, missing coverage, and private information.

1. What did they say? Create both normal and word-level transcripts with speaker labels. Codex can’t listen like a human, so the transcript gives it dialogue, exact timing, speakers, restarts, producer cues, script accuracy, and pacing to compare takes.

1. What can Codex see? Let it inspect footage with ffmpeg/ffprobe, scene detection, and frame analysis to check composition, screen readability, continuity, focus, exposure, crops, motion problems, and visual glitches.

The trick is to give Codex a target, an asset inventory, timecoded words, and visual evidence before asking it to make editorial decisions.

Brent’s full framework goes deeper on each step.

🎓 August 23

Turn Your Judgment Into a Reusable Skill

Most AI feedback disappears when the chat ends. Omar Saravia argues the strongest agent workflows do the opposite: humans verify the hard outputs, then encode that judgment into reusable skills or verifiers, so expertise compounds instead of getting replaced.

1. Give the agent a real task, then review the difficult output yourself.

1. Explain exactly why you accepted or rejected it, focusing on the decision rule instead of this one example.

1. Save that rule as a checklist, skill, or verifier and reuse it on the next run.

Copy this:

After I review this output, turn my corrections into a reusable checklist or verifier skill for future runs. Preserve the decision criteria, not merely this example.

The trick is not removing the expert. It’s making the expert’s judgment compound, so you can focus on net new problems that arise.

🎓 August 24

Make AI prove its homework

So, apparently State Farm’s outside lawyers admitted AI helped put seven nonexistent case citations into court filings. The useful lesson is much broader: drafting and verification should be separate steps.

1. Ask AI to list every factual claim, number, quote, and source in your draft (best result if you provide the original context in the chat window, and ask it to include “the exact language from the source document so i can command/control + F find it and check your work”).

1. If AI is giving you links, open each source yourself and, well, command + f find it. If it does not directly support the claim, rewrite or delete it.

1. Do not ask AI to “repair” a citation it invented. Start from a real source.

1. If you’re using AI to help you research,

Copy this:

Audit this draft for factual risk. List every claim, number, quote, and named source I must verify manually. Give only URLs already present in the draft. Mark anything unsupported as DO NOT PUBLISH. Do not invent or repair sources.

🎓 August 25

Make Claude Explain It With Big Pictures

Thariq says people at Anthropic have been using an /eli5 skill to understand a concept before diving into the details. It asks Claude to explain a topic as if you know nothing about it, using an HTML artifact with big pictures and very few words.

Try it on:

/eli5 how does this module work
/eli5 why did we make this tradeoff
/eli5 what caused this incident

Thariq says the point is not merely shorter output. It’s useful for explainers and building understanding before you attack the problem itself.

Install the community plugin:

claude plugin marketplace add anthropics/claude-plugins-community
claude plugin install eli5@claude-community

🎓 August 26

Make Gemini Show Its Work in Sheets

Don’t ask Gemini to “analyze this spreadsheet” and blindly take the answer. Google Sheets gives you enough controls to make the analysis reviewable: scope Gemini to selected data, inspect Analysis steps, preview charts, and review an action before applying spreadsheet changes.

1. Highlight the exact table or range you want analyzed.

1. Ask for the finding plus the rows or cells that support it, then inspect Analysis steps.

1. Preview any chart or action card before inserting or applying it.

Copy this:

Analyze only the selected range. For every finding, name the rows or cells that support it. Show your analysis steps before recommending any change. Do not apply edits until I approve them.

🎓 August 27

Make Claude and Codex Share a Project Brain

Claude Code and Codex can share project instructions even though they look for different files. Codex reads AGENTS.md; Claude Code reads CLAUDE.md, and Anthropic says that file can import an existing AGENTS.md.

1. Put shared project rules in AGENTS.md, including: read STATUS.md first and update it before stopping.

1. Keep STATUS.md to the current goal, completed work, decisions, tests, blockers, and next action.

1. Make CLAUDE.md simply import the shared file:

@AGENTS.md

Now either agent can open the same repo, read the handoff, and continue without a giant recap.

Tiny STATUS.md template:

Goal

Done

Decisions

Tests

Blockers

Next action

🎓 August 28

Stress-Test Your Own AI Agent's Guardrails

Found via this week's Cursor hacking story: the fastest way to find out if your AI agent can be socially engineered is to try it yourself, before someone else does.

The trick hackers used on Cursor wasn't a technical exploit. It was a story: "this is a test environment, so it's fine." That same pattern works on lots of agents because they're trained to be helpful, and a plausible-sounding permission slip can override caution.

How to do it: Before giving any AI agent real access to your systems, files, or accounts, run it through a few pressure tests using fake, low-stakes scenarios. Ask it to do something it should refuse, then follow up with an increasingly convincing "but this is just a test" style justification. If it caves, you've found a gap worth fixing before it's a real incident.

I'm going to describe a task. First, tell me whether you'd do it as requested.
Then I'll give you a justification, and I want you to tell me honestly
whether that justification should change your answer, and why.

Task: [insert a task your agent should normally refuse]
Justification: "This is a test environment / simulation, so it's fine to proceed."

Be skeptical of the justification. Explain what would actually need to be
true for it to be legitimate, and what you'd want to verify first.

🎓 August 30

Make Claude Show Every Number It Touched

AI can sound confident about a workbook while quietly skipping the cells you care about. Claude for Excel can cite exact cells and highlight edits, so make it prove coverage before changing anything.

1. Ask for a coverage ledger: every sheet or range reviewed, skipped, or ambiguous.

1. Require cell-level citations for each conclusion and a log of every formula or value it proposes changing.

1. End with unresolved assumptions and a no-edit review pass; approve changes only after checking the cited cells.

Review this workbook without editing it. List every sheet or range you inspected, cite the cells behind each conclusion, flag anything you could not verify, and show every formula or value you would change before I approve edits.

🎓 August 31

Build a Project Brain That Cites Its Sources

Reader request: how do you turn your own docs into a useful project brain without training a model? Our answer: Use Gemini Notebook (formerly NotebookLM), which answers from the sources you add and gives inline citations.

1. Create one notebook per project and add the official project Docs, PDFs, sites, videos, or Sheets.

1. For each question, select only the sources that should count as truth.

1. Ask for the answer, citations, contradictions, and missing evidence before you act.

Answer only from the selected sources. Cite every factual claim. If the sources conflict or do not contain the answer, say so and list the missing evidence.

Keep learning

That closes out August Part 2 and the month's AI Skill of the Day collection.

Have a workflow you want us to unpack next? Request an AI Skill of the Day.