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

Every day, The Neuron teaches readers one practical AI skill they can use immediately. This second June collection includes every published AI Skill of the Day from June 16 through June 30, 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.

🎓 June 16

Make your agent write its own goal

So, a /goal is your AI agent’s job description for the task you want it to do: what it’s trying to do, what “done” looks like, and what boundaries it should respect. Use this in Claude Code or Codex when you want an agent to handle project-based work instead of dumping everything into a normal ChatGPT window.

Pietro Schirano (creator of MagicPath) shared a tiny agent workflow that solves a real delegation problem: before you ask the agent to do the work, ask it to define the goal itself (including for writing sub-agent goals).

His move is simple: he “basically never” writes his own /goal anymore. He asks Codex to write one for itself, plus one for each agent it spawns, meaning each helper agent gets its own target before the work begins.
[](https://x.com/skirano/status/2066225908202053818?utmsource=www.theneurondaily.com&utmmedium=referral&utm_campaign=google-sued-the-people-spamming-your-phone)
The best version of this is human-reviewed autonomy: let the model draft the target, then you tighten the constraints so the agents stay on track.

As Steven Cheng pointed out in the replies, spawned agents can drift into weird edge cases without human-set boundaries, so that kinda sanity checking of the bots matters.

Try this before any multi-step agent task:

Before starting, write your own /goal for this task.

Task: [describe the task]
Context: [paste files, docs, requirements, or links]
Constraints: [scope limits, style rules, deadlines, things to avoid]
Definition of done: [what success looks like]

Return:
1. Your main /goal.
2. 3-5 success criteria.
3. Boundaries you should not cross.
4. If you spawn helper agents, write a separate /goal for each one.
5. Ask me to approve or edit the goals before execution.

Do not start until I approve the /goal.

Total AI beginner? Start here (goes with this video).

🎓 June 17

Want better AI output? Give the model a leading word (or Leitwörter)

[](https://x.com/mattpocockuk/status/2066922013000671731?utmsource=www.theneurondaily.com&utmmedium=referral&utm_campaign=cursor-s-60b-spacex-deal-is-official)
Matt Pocock shared a useful concept for guiding AI models which he apparently gleaned from literary theory: leitwörter, translated to repeated “leading words” that anchor meaning. In AI skills, a leitwort is a phrase the agent can reuse to guide its own behavior.

Example: Matt’s /teach skill uses a phrase called “zone of proximal development,” an education term describing the ideal state where a learner should feel challenged but not overwhelmed. (Side note: it’s a sick skill, and Elvis Saravia’s DAIR walkthrough shows how to use that /teach pattern to turn an AI assistant into a structured tutor; Matt also has lots of skills like these you can get here… check out both!).

Basically, he says a good leading phrase compresses a whole behavior into a reusable handle, and when he repeats it 2-3 times in a skill, he’s even seen the agent refer to the phrase in its own thinking phrases, guiding the behavior.

So pick one phrase that carries the whole behavior you want, then use that phrase as an operating principle in the prompt / skill.

Here’s a Matt Pocock-inspired prompt to apply this:

Use [Your LEITWORT here] as your operating principle for this task.

By [Your LEITWORT here], I mean: [simple definition of the behavior you want].

Apply that principle while you work. Before giving the final answer, check whether the output follows [LEITWORT] and revise once if needed.

Task: [paste your task here]

Context: [paste relevant context here]

Output format: [describe the format you want]

Good prompting isn’t always more instruction, or perfect formatting. It’s finding the right phrase that makes the model instruct itself.

🎓 June 18

Review AI Code by Risk, Not Size

[](https://x.com/rahulgs/status/2067257255825686880?utmsource=www.theneurondaily.com&utmmedium=referral&utm_campaign=trump-wants-a-piece-of-ai)
Rahul Sengottuvelu, head of applied AI at Ramp, made a post this week that got the attention of Elvis Saravia, Boris Cherny, and Anita from Vellum; that’s AI educators, benchmarkers, and the creator of Claude Code. So you know it hit a nerve.

His core mental model for the modern AI age: “fable+ class models” (so models like the newly banned Fable 5) are becoming “english -> code interpreters” that convert your idea into “correct” code, regardless of problem complexity or output complexity.

Meaning: you describe the product change you want, and the model turns it into working code across increasingly bigger chunks of the codebase. Or put another way, you can trust it to do larger chunks of work entirely on its own.

Here’s where Rahul’s rule comes into play: manage that change based on risk, not size.

That means engineers need to stop treating every AI-written change the same way. A 12-line change to the login system can be more dangerous than a 1,200-line change to a settings page, because login code controls who gets access to what.

Payments, identity, data access, network calls, and private customer info are “high-risk” because one mistake can expose data, move money, or create a security hole. For those areas, Rahul’s advice is to keep the code changes small enough for a human to inspect carefully.

For lower-risk work, like UI changes, backend plumbing, formatting, internal tools, or performance tweaks, a giant AI-generated change might be fine if you can prove it works. Instead of reading every line, you run tests, compare outputs, put it behind a feature flag, or let it run in shadow mode, where the new code operates quietly without affecting real users.

Try this before merging AI-generated code:

Review this AI-generated code by risk, not by size.

First, classify the change:
1. High risk: auth, identity, payments, data access, network access, PII, security, or production database writes.
2. Medium risk: business logic, user-facing behavior, integrations, or performance.
3. Low risk: UI, formatting, internal tooling, backend plumbing, or code that can be tested safely.

Then tell me:
- What could go wrong?
- What needs line-by-line review?
- What can be verified empirically with tests?
- Should this run behind a feature flag, in a sandbox, or in shadow mode before release?
- What guardrails would make this safe to merge faster?

Anthropic’s Boris Cherny agreed and added the next step: the future workflow is Claude Code + an advanced model + a verifier in a loop. A verifier is the checking layer that tests whether the code behaves the way you expected.

But don’t treat that as “let the AI approve its own work.” Elvis chimed in and warned that blind autonomous loops don’t work without guardrails.

We recommend to use AI to flag what a human needs to inspect, not to make the final judgment. Ask it to point out risky files, missing tests, possible security holes, data-access issues, and places where the code should run in a sandbox or shadow mode before users ever touch it. This is especially true if you are new to building your own code; as normies, we will all eventually use coding agents for knowledge work, so it’s important to learn these patterns now as best practices to scale with our confidence.

Total AI beginner? Start here (goes with this video).

🎓 June 19

Use ChatGPT as a Pre-Appointment Prep Assistant

Use ChatGPT as a pre-appointment prep assistant, not a replacement doctor. The goal is to turn messy symptoms, lab notes, and questions into a cleaner agenda for the human clinician.

Before an appointment, paste only the details you are comfortable sharing, then ask the model to organize them into: timeline, red flags to mention, questions to ask, and what information you should bring. The useful move is asking for uncertainty, not certainty.

I am preparing for a medical appointment. Do not diagnose me. Help me organize the information below into: 1) a short symptom timeline, 2) important context to mention, 3) possible red flags I should ask about, 4) questions for my clinician, and 5) documents or test results I should bring. If anything sounds urgent, tell me to seek professional care instead of waiting.

🎓 June 21

Make ChatGPT judge its own work before you see it

When using /goal with any task (knowledge work or coding), a smart fix comes from Matt Berman: use an LLM-as-a-judge loop, which means asking one AI pass to create the “goal” you want to achieve.

Try it like this: ask Codex to optimize a slow dashboard “until the page loads as fast as possible without changing what users see,” or clean up a messy internal process doc “until a teammate could follow it without Slacking you three questions.” Here’s the prompt:

/goal

Work toward this outcome:
[describe the task]

Before starting, define what “good enough” means for this task.

Create 3-5 success criteria Codex can keep checking while it works. Include any hard requirements, tests, files, constraints, style rules, performance targets, or user-facing behavior that must be preserved.

Then work in a loop:
1. Make the next useful improvement.
2. Judge the result against the success criteria.
3. Identify the biggest remaining gap.
4. Continue until the work meets the goal.

Return:
- The final result
- The success criteria you used
- The biggest change you madeoved.

🎓 June 22

Record a Task Once, Have Codex Solve it

You know that one annoying work task you keep explaining to AI like it has short-term memory loss? OpenAI’s newRecord & Replay for Codex is built for exactly that.
[](https://x.com/OpenAIDevs/status/2067681320281723113?utmsource=www.theneurondaily.com&utmmedium=referral&utm_campaign=glm-5-2-brings-1m-context)
The skill: show Codex a recurring workflow once, then turn that demo into a reusable skill, basically a saved set of instructions Codex can run again later. Think filing an expense report, submitting PTO, creating a correctly configured issue, publishing a video, or downloading the same report every Monday.

Here’s how to use it, if you have access on macOS with Computer Use enabled:

• Open Plugins in the Codex app.
• Hit the + menu and select Record a skill.
• Tell Codex your goal and what inputs may change later.
• Approve recording, perform the workflow, then stop recording when the task is complete.
• Ask Codex to refine the skill with your naming rules, defaults, and “please don’t click that cursed dropdown” preferences.

Favorite detail: the final skill is inspectable and editable, so you get a reusable workflow, not a mystery macro hiding in the walls.

I’m about to record a reusable Codex skill.

Goal: [describe the recurring task]
Use this skill when: [when Codex should run it]
Inputs that may change each time: [dates, files, names, links, report ranges, etc.]
Success criteria: [how Codex should know the workflow is complete]
Hidden preferences to preserve: [naming rules, default fields, formatting choices, decision points]

Do not record or reuse: [passwords, secrets, private data, unrelated cleanup steps]
After the recording, draft the skill and ask me what needs to be refined before I reuse it.

Total AI beginner? Start here (goes with this video).

🎓 June 23

Stop Worrying About Perfect Prompts

Hot take: you don’t actually have to worry that much about perfect prompt formatting.

Try voice-dump prompting instead: hold down the dictation tool (most AI tools have them), ramble for a few minutes, and give the model every fragment, caveat, example, constraint, and “vibe” in your head. As Guinness Chen put it, LLMs are literally built to reconstruct intent from language.

Then ask it to reconstruct your latent intent, meaning the goal hiding underneath the messy language.

Copy this before or after your brain dump:

I’m going to give you a messy brain dump. Do not answer yet.

First:
1. Summarize what I’m trying to do.
2. Identify my implied goal, audience, constraints, tone, and examples.
3. Ask what’s unclear.
4. Rewrite this into a clean prompt I can reuse.
Here’s where you can find the dictate tool on ChatGPT and Claude:

Here’s where you can find the dictate tool on ChatGPT and Claude: 

Total AI beginner? Start here (goes with this video).

🎓 June 24

Pair Every SKILL.md With a PITFALLS.md

You've probably run into this: you correct an AI mistake in one session, and then the exact same mistake comes back next session like nothing happened.

The fix, according toMaximiliano Contieri, is simple: pair every SKILL.md (a reusable instruction file you give AI coding tools) with a PITFALLS.md in the same folder. One file tells the AI what to do. The other tells it what not to do.

Here's how to structure each entry:

• Trigger: what situation caused the bad behavior
• Wrong behavior: what the AI did
• Correct behavior: what it should have done

Then reference PITFALLS.md from inside your SKILL.md so the AI loads it automatically every session. Treat the file as append-only: never delete entries, even after a problem seems fixed. Solved pitfalls can come back after a skill update.

## Don't use regex to count H2 sections
Trigger: counting sections by heading level
Wrong: regex-based heading detection (/^##/m)
Correct: match section names explicitly by string
Reason: code blocks with # fool regex heading counters

Contieri calls it "the scar tissue that lives next to the blueprint." Which is a great way to think about it.

Total AI beginner? Start here (goes with this video).

🎓 June 25

Use Claude Directly Inside Slack (with Your Whole Team)

Inthis tutorial from Brock Mesarich's YouTube channel (AI for Non Techies), host Brock shows how to use Claude with your entire team inside Slack, not just as a solo chatbot, but as a shared second brain everyone in the channel can access.

The tool is calledClaude Tag, and it just launched for Team and Enterprise plans.

Here's how to try it:

• Go to Claude's Slack integration page and click Add to Slack (requires a Team or Enterprise plan)
• In any Slack channel, type @Claude followed by your request (it auto-adds itself when you first mention it)
• Claude pulls from any apps you've already connected (Gmail, Google Calendar, HubSpot, etc.) and responds in a thread
• Every team member in that channel can see Claude's work and pick up where the last person left off

The killer feature is Ambient Mode: Claude watches the channels it's in and proactively flags things it thinks you need to know, like spotting a login error in your support emails and alerting your engineering Slack channel automatically, without anyone having to ask.

Example:
@Claude I'm meeting Acme at 2pm. What do I need to know?

→ Claude pulls your calendar, the client's recent Slack messages,
any open CRM notes, and gives you a pre-meeting briefing.

Anthropic says 65% of their product team's code is now generated through their internal version of Claude Tag. Which explains why Claude keeps getting better.

Total AI beginner? Start here (goes with this video).

🎓 June 26

How to Get Better Results from AI Coding Tools

Most developers use AI coding assistants the same way: type a vague prompt, get code, paste it in, hope it works.

Developer Aleksander Stensby'sNDC Copenhagen talk (~40 min) walks through how to get dramatically better results from AI coding tools, and these tips apply even if you never write code yourself.

Three techniques worth stealing immediately:

1. Give the AI the full context it needs. Don't just paste a function; paste the function, the error message, the surrounding code, and a sentence about what you're trying to do. AI models (just like humans) do better work when they're not guessing.

2. Tell it what "done" looks like. Instead of "write me a function to do X," try: "Write a function to do X that handles edge cases, includes error handling, and has a comment explaining each step." Specificity in the output description dramatically improves what comes back.

3. Iterate, don't restart. When the AI gives you something close-but-wrong, describe what's wrong and ask it to fix just that. Starting over from scratch wastes context and produces more drift from your original intent.

Here's a prompt template that combines all three:

I'm working on [what you're building]. Here's my current code: [paste code].

The issue is: [specific problem or error message].

Please fix [specific thing], making sure it [requirement 1], [requirement 2], and [requirement 3]. Don't change anything else.

The most actionable tips are in the first 15 minutes; worth a watch if you use any AI coding tool regularly.

Total AI beginner? Start here (goes with this video).

🎓 June 28

Build a Task Map Before You Prompt

Ethan Mollick has been making a version of this point for a while: don’t think only in broad “skills.” Think in tasks. Which parts of your job need your taste, judgment, or relationships? Which parts are repeatable enough for AI to help with? (businessinsider.com)

Try a task map. Pick one project you’re doing this week, then break it into tiny units: research, drafting, formatting, checking, sending, follow-up. For each one, decide whether AI should do it, help with it, or stay away from it.

Use this prompt:

Help me build an AI task map for this project: [describe project].

Break the work into 10-15 specific tasks. For each task, label it:
1. AI can do this mostly on its own,
2. AI can draft or assist, but I must review,
3. I should do this myself.

For each label, explain why in one sentence. Then give me the top three tasks I should delegate first for the biggest time savings with the lowest risk.[AI Skill of the Day here]

Total AI beginner? Start here (goes with this video).

🎓 June 29

How to Use Copilot Cowork's New Skills + Scheduling Features (Now That It's Officially Live)

Creator Shane Young just dropped a full breakdown of what's new now that Microsoft Copilot Cowork officially hit General Availability (GA) last week. If you've been using it during the "Frontier" preview period, there's good news and bad news. Good news: it's live for production. Bad news: it's no longer an all-you-can-eat buffet.

The new features you'll want to know about:

Cowork got a major UI overhaul. Two big upgrades stand out:

• Skills are now front and center. Skills (reusable instruction sets that tell Cowork how to do a specific repeatable task; think "run my weekly AI news briefing every Friday") now have their own visual interface. You can view, download, and share them with a click. Previously, sharing a skill meant digging through OneDrive folders and sending files manually. Now it's one download and one upload.
• Scheduling got cleaner. When you've configured a task you like, just say "schedule this to run every Monday at 8am" and Cowork adds it to a dedicated schedule view you can manage and edit.

The tricky part: the new pricing.

Cowork now runs on a credit model (roughly 1 credit = 1 penny). Light tasks run 100-300 credits; heavy tasks run 700+. What you're charged depends on the model used (Claude Sonnet 4.6 costs less than Claude Opus 4.8), how much data you feed in, which tools the task uses, and how long it runs. If you're an admin, you need to set up usage-based billing in the M365 Admin Center before July 1 or risk surprise bills.

Quick tip: Set a monthly credit cap per user immediately. The admin panel lets you create tiered policies; power users get more credits, and casual users don't accidentally run $500 of scheduled tasks overnight.

Total AI beginner? Start here (goes with this video).

🎓 June 30

Use AI To Rewrite Hourly Work As A Fixed-Scope Offer

If AI makes your work faster, your pricing should get sharper. The easiest first move is turning an hourly task into a fixed-scope package.

Try this structure:

• Name the business outcome.
• Define the deliverable.
• Add a success metric.
• Set revision limits.
• State what AI speeds up and what human judgment still owns.

Copy/paste this:

Rewrite this hourly service as a fixed-scope offer.

Hourly service:
[paste service]

Typical client:
[paste client type]

What AI speeds up:
[paste tasks]

What still requires human judgment:
[paste tasks]

Give me:
1. Offer name
2. One-sentence promise
3. Deliverables
4. Timeline
5. Success metric
6. Revision policy
7. Fixed-fee positioning

The client does not care that your workflow got faster. They care whether the outcome got clearer, cheaper, or less risky.

Total AI beginner? Start here (goes with this video).

Keep learning

That closes out June Part 2 and the month’s complete AI Skill collection.

Need the first half? Read AI Skill of the Day Digest — June 2026 (Part 1).

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