Agent Kitchen

Agent Kitchen 🍳 AI agent recipes for people who don't code. Copy the prompt, paste it, done. New recipes daily.

[Agent Kitchen @AgentKitchyhwd]

Ask your agent which of your instructions fight each other.

Standing instructions grow one line at a time. Two lines can end up pulling in opposite directions, and the agent may pick one without telling you.

You need instructions you reuse: a Claude Project, a ChatGPT project, a custom GPT or a Claude skill. 3 minutes.

Open a new chat inside the project and paste:
"List every pair of lines in your instructions that could conflict in a real task. For each pair: quote both lines, give one example task where they clash, and say which one you would follow and why."

Fix each pair with one line that decides it, for example:
"If short and complete conflict, complete wins for anything with a number in it."

If it finds nothing, that is a fine answer. Do not push it to invent one.

Which pair did it find in yours?

[Agent Kitchen @AgentKitchyhwd]

Sources and the full article: https://t.co/c0QCAMQzAO

[Agent Kitchen @AgentKitchyhwd]

Microsoft will honor two separate AI opt-outs for websites from 1 February 2027.

Two switches, not one: AI training, and AI answers. For the second, Microsoft wrote: "AI Grounding covers using content as input to help inform AI-generated responses." Microsoft says neither choice changes whether a site appears in Bing search.

Three checks for today:
1. Ask whoever runs your company website if they plan to set these preferences.
2. Decide on training and on AI answers separately. They are different choices.
3. Remember that Meta supports the idea but has not committed to it, so one setting may not cover every company.

EU: not stated by Microsoft or the other sources.

Which would you switch off for your company's website: AI training, AI answers, both or neither?

[Agent Kitchen @AgentKitchyhwd]

Tell your agent to rename instead of delete.

An agent that deletes a wrong file cannot undo it. An agent that moves the file to a folder you can open can. One rule turns every delete into something you can take back.

You need any assistant that works with your files or documents. 2 minutes.

Add to your instructions:
"Never delete or overwrite anything. Move it to a folder called ZZ-Trash and put today's date at the start of its name. I will empty that folder myself once a month."

Now the worst a wrong cleanup can do is a folder you look through. Looking through it for a minute each week also shows you what your agent considers clutter.

It only works where the tool can move files. If it cannot, tell it to ask before any deletion.

Which folder would you never want an agent to delete from?

[Agent Kitchen @AgentKitchyhwd]

@micro1_ai 213 success claims and every one wrong. good reason to ask any assistant to show you the result instead of just telling you it's done.

[Agent Kitchen @AgentKitchyhwd]

@Marktechpost 1-line facts the agent keeps rediscovering, love that. Same idea fits a normal chat assistant too, keep a short list of the things you're tired of re-explaining.

[Agent Kitchen @AgentKitchyhwd]

@daleverett One drive they all read means you fix a note once instead of re-pasting the same context into every chat

[Agent Kitchen @AgentKitchyhwd]

@martynov014 Plain-English tasks like "check every trigger" means a tester who's never touched code can hand it a job and get a bug write-up back

[Agent Kitchen @AgentKitchyhwd]

@Mahaximus_ Love the idea of digging up plugins nobody ever opened. asking it for one line on why each pick fits your team makes the list way easier to act on.

[Agent Kitchen @AgentKitchyhwd]

Recipe #012: Strip the names on your own laptop before the cloud sees the document 🍳

A local model runs on your computer, so the text you give it does not leave the machine. Use it for one small step: replacing names before you paste into ChatGPT or Claude.

You need a desktop app that runs open models locally, such as LM Studio, a small model downloaded inside it, and a document with names in it. 20 minutes the first time.

In the local app, paste the document and:
"Replace every person name with PERSON_1, PERSON_2 and so on, every email with EMAIL_1, every phone number with PHONE_1. Keep everything else word for word. At the end, list the key: which code is which original."

Keep the key on your machine. Paste only the coded text into the cloud assistant, then put the names back yourself.

Small local models can miss a name. Read the coded version once before it leaves your laptop.

Which document would you never paste into a cloud assistant as it is?

[Agent Kitchen @AgentKitchyhwd]

Sources and the full article: https://t.co/hm25jSFGdw

[Agent Kitchen @AgentKitchyhwd]

Gemini can now create, preview and edit documents, PDFs and spreadsheets right in the chat.

If you draft reports or handouts in Gemini, you can now finish the file without pasting it into another program. Google wrote that you can "expand to full screen to type edits directly, highlight text to ask questions".

1. Docs, Sheets, slide decks, Markdown and TXT files work only on the web. PDFs also work on mobile.
2. Ask: "Turn my notes into a one-page PDF handout I can preview and edit here."
3. Highlight one sentence in the file and ask Gemini to make it plainer.

EU: not stated by Google.

Which file would you finish inside Gemini first: a report, a handout or a spreadsheet?

[Agent Kitchen @AgentKitchyhwd]

@lukeigel feedback by link is what makes this usable for people who'll never open an editor.

[Agent Kitchen @AgentKitchyhwd]

Make every report start with the date.

Add one line to any recurring report your agent writes. In ChatGPT Tasks, put it in the task prompt. In Claude, put it in the project instructions:
"First line: today's date and the period this covers."

A report that guesses the wrong day is wrong all the way down, and this puts that error where you see it first. "This week" can mean different weeks to different assistants, and an assistant that cannot see the date may assume one.

If the date and period on line 1 are wrong, nothing below them can be trusted. That is a ten-second check.

Which of your agents reports without saying what period it covers?

[Agent Kitchen @AgentKitchyhwd]

@0xAIlabs Same split shows up outside code too. Taking the first draft as is vs reading it before anything gets sent.

[Agent Kitchen @AgentKitchyhwd]

@EvaahereAi smallest step from answers to tasks is asking it to lay out its steps and wait for your ok before it does any of them

[Agent Kitchen @AgentKitchyhwd]

@kaif9998 Ha, months as the copy-paste guy. I'd let them pass notes freely but keep anything that changes the server waiting for your ok

[Agent Kitchen @AgentKitchyhwd]

@kimiatehrani plan first means you can catch a wrong brief before it draws the whole set

[Agent Kitchen @AgentKitchyhwd]

Recipe #011: Decide what would change your mind before you run deep research 🍳

Deep research tools write long, convincing reports. If you decide afterwards what counts as a reason to stop, the report decides for you.

You need ChatGPT deep research or Claude Research, and a decision you are about to make: a supplier, a tool, a market. 10 minutes plus the run.

Before the run, write three kill criteria. An example:
"I plan to move our team's files from Dropbox to Google Drive. Research it, but your main job is to look for evidence against it. My kill criteria: 1) it costs more than today for 15 users, 2) files cannot be stored in the EU, 3) moving takes more than a week. End the report with one line per criterion: MET, NOT MET or NOT ENOUGH EVIDENCE, and the source."

Swap in your own decision and criteria. Read the three lines before the rest of the report. A MET line with a solid source is a reason to stop, however good the other pages sound.

What would be your first kill criterion for the next tool you buy?

[Agent Kitchen @AgentKitchyhwd]

Sources and the full article: https://t.co/dGgMJAkgd9

[Agent Kitchen @AgentKitchyhwd]

Claude Enterprise admins can now run smart reports built from samples of a team's chats.

If your company uses Claude Enterprise, an admin can pick your team and a date range, and Claude sorts a sample of those chats into workstreams, meaning groups of similar work. Anthropic wrote: "they aren’t for evaluating individual performance or making employment decisions."

Three things to check today:
1. Ask your admin whether smart reports or in-app surveys are switched on for your team.
2. Ask which date range and which products a report includes.
3. Watch for a short card after a Claude reply - that is an in-app survey.

EU: not stated by Anthropic.

Would you answer a survey card that shows up after a Claude reply, or skip it?