New AI tools and features arrive every week, and working out which ones deserve your attention has become a job of its own. Advising animal protection organizations on their AI skills, we keep coming back to the same nine features. Learn these and you will have the large majority of the value available to an individual advocate today.
| Step | What it is | Why it comes here |
|---|---|---|
| 0. The right app | An agentic coworker app rather than a plain chatbot | Everything below assumes an assistant that can reach your files and act |
| 1. Memory | Standing context about you and your organization | Improves every conversation from now on, whatever you are doing |
| 2. Connectors | Access to your calendar, inbox, files, and tasks | Brings your real work into AI, which is what builds the habit |
| 3. Browser control | AI that can see and act in your browser | Reaches everything the connectors cannot |
| 4. Projects | A space per strand of work, with its own context | Stops you re-explaining the same campaign every time |
| 5. Skills | Saved, repeatable instructions for a task | Grows naturally out of projects and hard-won conversations |
| 6. Automations | A skill that runs on a schedule | Only trustworthy once the skill itself is |
| 7. Meetings | Your assistant connected to a meeting recorder | The one connector big enough to deserve its own step |
| 8. Interactive pages | Small interactive pages as a work output | Most useful once the assistant already has your context |
One practical prerequisite: this path assumes a paid seat on one of the major assistants, Claude, ChatGPT, or Gemini (we recommend Claude or ChatGPT over Gemini). Free tiers are substantially weaker. See what AI actually costs a small organization for the numbers, including nonprofit pricing.
Step 0: Work in an agentic app on your desktop
Each of the major providers now has a work-oriented version that goes beyond answering questions: Claude Cowork, ChatGPT Work, and Gemini Spark. The details differ, but the shared idea is that you give the assistant a place to work, either a folder of your files or access to your Workspace. It reads what it needs, produces real documents, and carries a multi-step task through instead of handing you text to copy somewhere else.
One availability note for this movement: Gemini Spark now runs in most countries on Google AI Pro and Ultra, but it excludes the UK, the European Economic Area, Switzerland, and Nigeria. If your team sits in any of those, Spark is not currently an option.
We recommend these apps over the simpler chat apps (regular Claude, ChatGPT, or Gemini), which have fewer features. There are also code apps: Claude Code, OpenAI Codex, and Google Antigravity. These are the most capable of the three tiers, though they were built for programmers and can be intimidating if you are not one. If you already work happily in a code app, stay there. For everyone else, the work apps are a good middle ground. Our FAQ on Claude, Claude Code, and Cowork compares the tiers in more detail.
Two habits that make all nine better
Talk to AI instead of typing. Many people get better results by speaking to the models rather than typing at them. The reason is that you naturally give more information when you talk. Typing makes most of us terse, and the assistant can only work with what you gave it.
Do not worry about sounding fluent. Modern speech-to-text tools clean up the “ums” and “ers”, add punctuation, and notice when you are speaking a list. You can circle back, contradict yourself, or dictate your third point first and then say “actually, go back to the beginning”. The models are good enough to sort it out, so you can dump a full stream of consciousness and let it organize.
Every major AI app has a microphone button, and every phone and laptop has system dictation, so this costs nothing to try. Some people prefer a dedicated app such as Wispr Flow, Monologue, or SuperWhisper.
Match the effort to the task. Most AI apps let you choose which model you work with, from fast and less capable to slow and more capable. Claude offers Haiku, Sonnet, Opus, and Fable in roughly that order. ChatGPT has Luna, Terra, and Sol. Gemini has Fast, Thinking, and Pro.
Some apps also let you set how much effort the model spends on a task. Putting a ChatGPT model on “high” effort makes it think for longer before answering; “max” can keep it thinking for ten minutes or more. When a task needs the model to try several approaches or work through a long chain of reasoning, a higher effort setting produces a better answer. For everyday drafting and summarizing, it mostly adds delay.
Picking the model and effort level to match the task takes practice, and it is one of the habits that separates confident users from frustrated ones. Model names and tiers change often, so check your tool’s current documentation. Our guide to picking the right AI model and effort level covers when the extra effort is worth it.
Step 1: Memory
Memory improves everything else. Once your AI holds an accurate picture of you and your work, every future conversation flows more smoothly, and your assistant should never need you to introduce yourself twice.
Filling out that memory is what lets short, vague prompts work. “Message Steven to say I’ll be late for the meeting” only lands if your AI already knows who Steven is, which organization he belongs to, and which app your team talks in.
Two ways to fill it, and they complement each other. Ask the assistant to research you and your organization on the web, then save what it learns, which covers the public facts. Then ask it to interview you one question at a time about your role, your goals, and your constraints, which covers everything no website mentions. Read both summaries before you save them, and correct anything wrong. One caution: memory that has gone stale is worse than memory that is thin, because a colleague listed in a role they left last year will send your assistant confidently in the wrong direction.
Step 2: Connectors
Connectors let your assistant read from, and sometimes write to, the systems where your work lives: email, calendar, task managers like Asana or Todoist, your document store, your meeting recordings. In most tools you will find them under a settings page called Connectors, where you pick a service and log in.
This is the step where AI stops being a separate app you visit and starts sitting inside your work, which is what builds the habit. It also pairs with memory. Connectors without memory give you a very fast intern who has to ask you everything; memory without connectors gives you a well-briefed assistant with nothing to look at. With both, it stops asking and starts doing. Our how-to on turning meeting notes into action items is a worked example.
Before you connect anything, check two things. Connecting AI to your other tools means sensitive data can be read by the model and held on the AI company’s servers, and some connectors can change or delete things rather than just read them. Plenty of organizations restrict this, so check your organization’s AI policy to see which connectors you are allowed to use. If your organization does not have one yet, our policy worksheet will get the conversation started, and data security and privacy basics covers the groundwork. For EU and UK organizations there is a further question about where that data is processed.
Step 3: Browser use
Browser control lets your assistant take over your browser, click around, and complete tasks for you. It can work on a page where you are already logged in, so it reaches anything behind a login or paywall that you can reach yourself.
For ChatGPT and Claude you install an extension for Google Chrome. For Gemini it is already built in.
Use it to summarize the page you are looking at, to find settings buried in confusing websites, or to fill out forms. It is also good at checking a source: open the page a claim cites and ask whether the claim holds up.
Browser control carries some risk. Models occasionally click the wrong button, so do not turn one loose on a page where a mistaken click would cause real damage.
Step 4: Projects
Your job probably involves several strands of work at once:
- Campaigns
- Managing employees or volunteers
- Writing research reports or funder updates
Projects, called Gems in Gemini, give each strand its own space. When your assistant enters that space it reads a custom instruction file and starts out knowing the background. In Claude and ChatGPT you can turn a folder on your computer into a project, which lets the model see everything in the folder and add new files to it.
The trigger for making one is simple. If you are pasting the same background into a fresh chat for the third time, that background belongs in a project.
Step 5: Skills
A skill is a saved set of instructions your AI can follow, closer to a recipe or a standard operating procedure than to a prompt. Depending on your tool, the same idea appears as a skill, a custom GPT, or a Gem.
You might build a brand guidelines skill that tells an assistant how to format a document to your organization’s brand standards, or one that tells it how to update your Asana board.
As you encode more of your recurring tasks this way, your assistant can take on more of the work. The best way to start is to take a conversation where you solved something the hard way, iterating and correcting until the output was right, and ask for that conversation back as a skill, including the mistakes it made along the way and how to avoid them.
Skills are also a good way to crystallize your expertise into something your colleagues can use, so they get your standard of output without booking your time.
Step 6: Automations
An automation, or scheduled task, is a prompt that runs on a timer: hourly, daily, weekly. You might have one run at 6am that reads your calendar, works out which meetings matter that day, researches the attendees, and leaves you a briefing to read beforehand.
Automations sit after skills for a reason. A scheduled task repeats whatever the underlying skill does, including its mistakes, without you watching. Run it by hand a few times first, then put the working version on a timer. It is also worth keeping a human in the loop for anything that leaves your organization: have the automation leave a draft rather than send it.
We are describing automations built inside Claude, ChatGPT, or Gemini here. More complex ones can be built in dedicated tools such as n8n or Make, which we compare in choosing between n8n, Make, Apps Script, and custom code.
Step 7: Meetings
Strictly speaking, this is another connector and belongs under step 2. It gets its own step because the payoff is bigger than any other connector you will set up. An AI meeting recorder is the only tool we consider essential alongside access to a frontier model.
You invite a recorder to your meetings. It transcribes what was said, and can then summarize what happened afterwards. Always get permission from the other attendees before you record a call.
Most of these apps produce their own AI summary of a meeting, and those are useful. What is dramatically more useful is connecting your meeting recorder to your assistant, so it can work through the entire transcript in detail.
A single Claude or ChatGPT conversation can then take on most of the follow-up a meeting generates. With the transcript in front of it, your assistant can:
- Draft emails to every attendee with the action items specific to them
- Update a task manager such as Asana, or a CRM such as HubSpot
- Update calendars
- Post an update for colleagues in Slack
All of that happens in one conversation, from one transcript.
You can also have your assistant pull several transcripts at once and look for patterns across them. When an organization transcribes most or all of its meetings, the archive becomes a rich database of organizational wisdom: what was decided, when, and why.
We have a fuller guide to choosing a recorder on the way. The short version is that every tool we have tested transcribes to roughly the same standard, so the deciding factor is whether it connects directly to your AI tool of choice. Google Meet cannot, which is why we do not recommend it.
Handing note-taking and post-meeting admin to your assistant frees advocates to be fully present in the meeting itself.
Step 8: Interactive pages
For years, most office work produced three shapes: the document, the spreadsheet, and the slide deck. Everything you made ended up as one of those. There is now a fourth: the interactive page.
A website is one. A data dashboard is one. A small web app is one. So is a piece of software you build for yourself that only runs on your computer, or a report a reader can click through.
Interactive pages used to require programming skills and plenty of time, which put them out of reach for most organizations. Now that AI is fast and capable at writing code, anyone can build one. Instead of sharing a spreadsheet with colleagues, you can have your AI build a clickable dashboard on top of it. Instead of sending a funder report as a PDF, you can send an interactive version that lets them explore the parts they care about. What used to take a developer days now takes minutes, and it brings the work to life.
Where to start
Start with memory. It is the only step that improves everything you do afterwards without you having to remember to use it, and it takes about fifteen minutes.
If you are thinking about a whole team rather than your own workflow, pair this with the six pillars of AI transformation and our guide to picking the right AI tools. Every hour these features give back is an hour that goes into the work for animals.
