APIs, automation, agents, MCP, evaluation and improvement loops.
Let a larger model such as Claude Sonnet 5.5 plan the job and write the final answer, and hand the many small, repetitive pieces in between to Claude Haiku 5.5, each with narrow instructions and a fixed…
4 min read
GPT-5.5 left Codex on October 14 for everyone signed in with a ChatGPT account, so any default, saved setting, custom agent, scheduled task or script that still names gpt-5.5 needs a new model, usually…
4 min read
You call GPT-6.1 Sol through OpenAI's Responses API by setting the model to gpt-6.1-sol, choosing a reasoning effort, and putting the instructions you reuse at the start of every request so they can be served…
4 min read
An agent harness is the ordinary software wrapped around a language model that turns it into an agent: it runs the loop of asking the model what to do next, carries out the tool calls the model requests,…
4 min read
Write a task brief that states the outcome, the limits and how the agent will prove it is done, then make the agent keep a progress log and save a checkpoint after each step.
4 min read
Sorting, labeling, summarizing and preparing draft replies are safe to automate when the AI has only the permissions it needs and a person reviews every draft before it is sent.
4 min read
Collect a few hundred examples of the exact input and output you want, then train a small LoRA adapter on top of a frozen open-weight model with Hugging Face's TRL and PEFT libraries.
4 min read
Give Claude Opus 5 a written plan split into small, testable pieces, have it work through one piece at a time with a commit and a progress note after each, and review at checkpoints you set in advance.
4 min read
Keep your notes as plain text files in a small set of clearly named folders, add an index file that tells the agent what is where and how to behave, and give the agent access only to that folder through a…
3 min read
Since July 9, 2026, the ChatGPT app directory has become a plugin directory, and a plugin is now a package that can hold apps, skills and app templates.
4 min read
Give the model its own database account that can only read the few tables or views it needs, and let your code, not the model, run every query, with values passed as parameters.
4 min read
Permissions are rules that decide which tools a coding agent may use freely, which need your approval and which are forbidden.
4 min read
OpenAI says Agent Builder and Evals will stop being available on its platform from November 30, 2026, so anyone with a working canvas agent should write down how it works now and rebuild it before then.
4 min read
Give both agents one set of project rules in an AGENTS.md file, which Codex reads directly and Claude Code reads through a one-line CLAUDE.md import, then give each agent its own branch in its own git…
4 min read
Design the call flow first (greeting, what the agent may look up, what it must confirm, when it hands over to a person), then connect to OpenAI's Realtime API with the model gpt-realtime-2, describe your…
4 min read
Review it like a pull request from a fast, confident newcomer: first check the scope of what changed, then whether the tests are real and still pass on your machine, then look for secrets and new…
4 min read
List the exact tasks the agent must do, then grant the narrowest permission for each one, as a scoped, short-lived token rather than your password or a full-access key.
4 min read
For most small businesses, start with RAG, because the usual need is for answers based on your own facts, and those facts change.
4 min read
Write small automated tests that describe the behavior you want, run them to confirm they fail, commit them, and only then ask the coding agent to write code until every test passes without touching the test…
3 min read
For a small collection, often not: if everything fits comfortably in the context window, pasting it all in is simpler and can work well.
4 min read
Send each request to the smallest model that can handle it: a cheap first step, often nano itself, sorts requests by difficulty, nano and mini do the routine work, and full GPT-5.4 handles only the hard cases.
4 min read
All three now let you build an AI agent that decides which steps to take, but they package it differently.
4 min read
Wrap every scheduled AI job in a small guard that writes one log line per run, refuses to start when a daily run count or monthly budget is used up, and alerts you on failures, unusual spending or missing…
4 min read
Split the job into stages that each have one clear goal, give each stage its own agent with narrow instructions and a fixed output format, and let your code pass the result of one agent to the next.
4 min read
Agent memory is the information an AI agent keeps beyond the current conversation, such as your preferences, ongoing projects and past decisions, stored outside the model and brought back when relevant.
4 min read
Sign in with your ChatGPT account, add a project folder, and start one thread per task; each agent can work in its own copy of the code, and you review its changes as a diff before keeping them.
4 min read
A self-improving loop lets a model propose a better version of an agent's prompt, scores each version on a fixed set of test cases, and keeps the best one.
3 min read
Write it as a short operating manual in labeled sections: the goal and its limits, the context the agent needs, when to use each tool, the rules it must never break, when to stop, when to hand the task to a…
4 min read
A connector is an authorized link between an AI assistant and another service, such as Google Drive or Google Calendar, so the assistant can search your files or schedule and cite what it found.
4 min read
Collect about 30 real tasks you have already done, write down for each one the input and what a good answer must contain, and choose how each will be graded, by exact match, by a short rubric, or by a model…
4 min read
AI agents are most reliable on short, well-defined tasks with clear feedback, such as running code and checking whether tests pass, or gathering information from a few sources into a fixed format.
4 min read
Make a folder with a SKILL.md file whose front matter gives the skill a name and a description of when to use it, then write the steps below it and add any scripts or templates the steps refer to.
3 min read
Put the agent's script in a GitHub repository and add a GitHub Actions workflow with a schedule trigger, so GitHub's servers run it at a set time each day without your computer being on.
3 min read
You build an MCP server that offers one or more tools, register a small HTML component for ChatGPT to display, and link each tool to that component.
3 min read
In Agent Builder you draw your agent as a workflow on a visual canvas: a start point, guardrail checks, one or more agent steps with their tools, and an end, then you test it with preview runs and evals…
4 min read
Turn on Codex code review for a GitHub repository, then either let it review each pull request automatically when it is marked ready, or ask for a review with a comment such as "@codex review for security…
4 min read
A sub-agent is a helper agent that a main agent hands one focused job to; it works in its own fresh context with its own instructions and tools, then returns a short result.
4 min read
A good spec tells the agent what to build, what it must not do, and how everyone will know it is finished, ideally as tests it can run.
3 min read
Build a short chain of n8n nodes: a Schedule Trigger set to once a week, a step that fetches the week's data, an AI step that turns it into a summary, an approval email to you, and a final send to your…
4 min read
Install it with npm (it needs Node.js 20 or newer), type gemini in a project folder, sign in with a personal Google account, and describe what you want in plain language.
4 min read
Add a step that pauses the workflow after the AI has done its part and before anything leaves your hands, sends the draft to a person, and continues only if that person approves.
4 min read
Write a set of realistic tasks, each with a check on the final state the agent should leave behind (the outcome) and, where it matters, a check on the tools it called along the way (the trajectory).
4 min read
Write a short plain-text Markdown file at the root of your repository that tells a coding agent what the project is, which commands build and test it, and which rules to follow; name it AGENTS.md for OpenAI's…
4 min read
Use OpenAI's Codex when you want to hand off several well-defined tasks to cloud sandboxes and review the results later, and Anthropic's Claude Code when you want an agent working beside you in your own…
4 min read
A coding agent is an AI system that doesn't just suggest code but carries out a programming task itself: it reads your project, edits files, runs commands and tests, and repeats until the job is done or it…
4 min read
Treat it like any other change to software you depend on: confirm the change with a fixed set of test prompts, adjust your instructions or switch versions if needed, and report what you saw.
3 min read
Install it with npm, set your OpenAI API key, open a terminal in your project folder and type codex followed by what you want done, such as "fix the failing test." It reads your files, proposes changes and…
4 min read
Write the tests first, have the model write code, run the tests automatically, and send any failures back to the model with a request to fix them, repeating until the tests pass or a round limit is reached.
4 min read
No MCP server is safe by default: each one is a program that runs on your computer and whose tool descriptions are read by your AI assistant.
4 min read
Install the official MCP Python SDK, write a short file that creates a FastMCP server and turns one Python function into a tool with a decorator, then open it in the MCP Inspector to check it.
4 min read
The Responses API is OpenAI's newer way to call its models: one endpoint that takes your instructions and input, can keep the conversation on OpenAI's side, and can use built-in tools such as web search.
4 min read
Install the openai-agents Python package, describe each agent with a name, instructions and the Python functions it may call as tools, and run it with Runner, which handles the loop of model calls and tool…
3 min read
Never compare the headline price alone: multiply each model's input price by the input tokens your task uses, multiply its output price by the output tokens, and add the two.
4 min read
Install it with npm, open a terminal in your project folder, type claude, and sign in once with an Anthropic Console account that has billing turned on.
3 min read
Vibe coding is building software by describing what you want to an AI coding assistant, accepting its changes without reading them, and pasting any errors back until the program works.
4 min read
Give the agent a narrow, written task with clear limits, watch it work, and take the controls yourself for anything sensitive, such as signing in, entering payment details or confirming an order.
4 min read
Test the two halves separately. First check whether retrieval finds the right passages for a fixed set of questions, then check whether the answers stick to those passages, and have a person review a sample…
4 min read
Function calling is a feature of a model's API: you describe tools in your request, the model replies with the name and arguments of the one it wants, and your own code runs it.
4 min read
You install the Claude Desktop app and Node.js, then add the official filesystem server to Claude Desktop's configuration file with the path of one folder you choose.
4 min read
The Model Context Protocol (MCP) is an open standard, released by Anthropic yesterday, that defines one common way for AI assistants to connect to outside data and tools.
4 min read
You give a second model, the judge, the question, the answer to check and a short written rubric, and ask it to reason against each criterion before giving a score in a fixed format.
4 min read
Gather the documents people really ask about, clean them into simple text, split them into short self-contained pieces, and turn each piece into an embedding so the AI can find the right passages by meaning.
5 min read
Not your everyday computer, not yet. Anthropic's computer use beta is worth exploring, but only inside a sealed-off test machine with no personal data, limited internet access and a person watching every run.
5 min read
You place an OpenAI action between a trigger and the rest of your workflow, give it a clear prompt that uses fields from the trigger, and ask for output in a fixed shape so the next steps can use it.
5 min read
Prompt caching lets an AI provider remember the opening part of a prompt you send again and again, so later requests that start the same way are billed at a lower rate for that part.
5 min read
You open a WebSocket connection to OpenAI's Realtime API, configure the session with instructions, a voice, an audio format and any tools, then stream the caller's audio in and play the model's audio replies…
4 min read
Write a small script that does one bounded job, such as summarizing the latest items from a feed in a single API call, and let cron, the standard Unix scheduler, run it at a fixed time.
3 min read
Fine-tuning GPT-4o is worth it when you have a narrow, repeated task, a few dozen to a few hundred good examples, and a careful prompt that still misses on tone, format or domain rules.
4 min read
Describe the exact shape you want as a schema, send it with strict mode turned on, and the reply will match that shape every time unless the model refuses or runs out of tokens.
4 min read
You describe the job in plain language, connect the apps the agent may use, and decide exactly which actions it can take on its own, using a no-code tool such as Zapier Central or a custom GPT in ChatGPT.
4 min read
Send most of your requests to GPT-4o mini, which costs $0.15 per million input tokens and $0.60 per million output tokens, and keep the larger GPT-4o only for the requests that truly need it.
4 min read
Build it in six decisions: a narrow goal, a short list of tools, a loop that lets the model call those tools through function calling, a memory for what it has learned, firm stop rules, and a log of every…
3 min read
An agent framework is a code library that supplies the plumbing every AI agent needs, such as the think-act-observe loop, tool definitions, memory and the hand-offs between several agents, so you write less…
4 min read
Build a GPT when the people who need your assistant already use ChatGPT and you want something working this afternoon without code.
4 min read
You write example conversations in OpenAI's chat format, each one a system message, a user message and the ideal assistant reply, and save them one per line in a JSONL file.
3 min read
An eval is a fixed set of test inputs, each paired with what a good answer must contain, that you run through a model to score how well a prompt or system performs.
4 min read
Ollama is a new free, open-source tool, first released as an early preview on July 8, 2023, that downloads and runs open large language models on your own computer with one short command, such as "ollama run…
4 min read
You describe a weather function to the model in JSON Schema, let gpt-3.5-turbo-0613 reply with the arguments it wants to use, run the real weather lookup in your own code, and send the result back so the…
3 min read
You split your documents into short passages, turn each passage into an embedding, and for every question retrieve the few passages closest in meaning, then paste them into the prompt and tell the model to…
4 min read
A vector database stores embeddings, the lists of numbers that represent the meaning of text or images, and quickly finds the stored items whose numbers are closest to a query.
4 min read
Prompt injection is an attack where text written by someone other than the app's builder, typed into a chat or hidden in a web page, email or document, gives a language model new instructions that override…
4 min read
An agent loop asks the model to write a thought and one action in a fixed text format, your program reads that action, runs the matching tool, and sends the result back as the next message.
4 min read
An AI agent is a program that gives a language model a goal, a set of tools such as web search or file writing, and a loop that lets it choose and carry out one step after another until it decides the goal is…
4 min read
Auto-GPT is a free, open-source Python program released on 30 March 2023 that gives GPT-4 a goal and lets it decide its own next steps, such as searching the web, reading pages, saving notes and writing…
5 min read
llama.cpp is a small open-source program, released on GitHub on March 10, 2023 by Georgi Gerganov, that runs Meta's LLaMA language models in plain C/C++ on an ordinary computer's processor.
4 min read
A system message is the first message in a ChatGPT API conversation, written by the developer rather than the user, that tells the model who it is, what it should do and how it should behave.
4 min read
Write the comment or docstring as a small specification before the code: what goes in, what comes out, the edge cases, and one or two examples.
4 min read
You show the model a handful of photos, usually three to five, of the one subject or style you want, and run a short training step that links those pictures to a new word you can then use in prompts.
4 min read
You send each piece of text to OpenAI's Moderation endpoint, a free classifier released on August 10, and it tells you whether the text is flagged for sexual, hateful, violent or self-harm content.
4 min read
In Google Sheets you can write a small custom function in Apps Script that sends a cell's text to the OpenAI API and returns GPT-3's reply into the cell.
3 min read
Zapier and Make, the new name for Integromat, both connect apps without code, but they count usage differently: Zapier charges a task for each action that runs, while Make charges an operation for every…
4 min read
An API key is a secret code that identifies you to a service such as the OpenAI API, so every request made with it is billed to your account.
4 min read
An embedding is a long list of numbers that represents the meaning of a piece of text, so that texts with similar meanings get similar numbers.
4 min read
Fine-tuning trains a private copy of a GPT-3 model on your own examples, each one a prompt paired with the completion you want, so the model learns your task and style instead of reading them from a long…
4 min read
Write your request as the opening of the file you want, in comments that name the language, the pieces and the edge cases, then let Codex continue it.
4 min read
OpenAI Codex is a descendant of GPT-3 trained on natural language and billions of lines of public code, and since August 10, 2021 an improved version is available in private beta through the OpenAI API.
4 min read
No-code automation means connecting the apps you already use so that an event in one, such as a new form entry, automatically causes actions in others, without writing a program.
4 min read
With the GPT-3 API you can build almost any tool whose input and output are text, from a style rewriter or a question-answering helper to a chatbot character or a code sketch made from a plain description.
4 min read
The OpenAI API is a new web service, announced on June 11, 2020, that lets software send text to OpenAI's GPT-3 family of models and get a continuation back.
3 min read
Sign up for the weekly newsletter and receive the free guide: what changed in AI this week, one lesson, one prompt to try. Unsubscribe in one click.