This Changed The Way I Work With Agents

AAI LABS
Computing/SoftwareSmall Business/StartupsManagementInternet Technology

Transcript

00:00:00Co-work gives non-developers the powerful capabilities of cloud code, letting them create
00:00:04real automations, document workflows, pipelines and research tasks that were previously native
00:00:10to terminal only.
00:00:11But most of them are still getting low quality output and complain that using co-work has
00:00:15been eating up their limits.
00:00:17This is not happening because of its tool, it happens because there is no effort being
00:00:21put in the pre-setup.
00:00:22There is no right way for the pre-setup, it is always a series of steps you take to make
00:00:26the workflow tailored for what you need.
00:00:28Now I know we have been talking about such best practices in most of our previous videos
00:00:32but we found some new ones that were actually good and had a high impact on our workflows.
00:00:37The first thing you need to do before doing anything is creating a manifest.md for every
00:00:42folder you work with.
00:00:43This file lives in the root of the folder and contains guidance on how the actual folder
00:00:47looks.
00:00:48For cloud code users, this file is the same as the claud.md file.
00:00:52If you, like us, also work in a folder that contains a lot of information that is nested
00:00:56and structured, manifest helps better in working in such folders.
00:01:00Claud tends to get lost and pull noise from irrelevant files.
00:01:03This is because without a manifest file claud tends to navigate around the folder guessing
00:01:07where the right file actually is.
00:01:09It bloats the context unnecessarily which leads to using the wrong file as the source and generating
00:01:14bad quality output.
00:01:15This file tells which documents are the source of truth, which sub-folders map to which domain
00:01:20and what to skip entirely.
00:01:21The manifest.md contains 3 tiered levels of files to let claud know which file to give
00:01:27importance to and which to not.
00:01:28Tier 1 contains all the files your model should always load and which acts as the source of
00:01:33truth, listing all those files which cannot be skipped at all.
00:01:36Tier 2 is those files that you want to be loaded on demand.
00:01:39These include those kinds of files that you don't need right away but might need.
00:01:43And lastly, the third tier is the archive data, which is your past versions of data that you
00:01:48don't need but keep for records sake.
00:01:50That's why we flagged it as "Ignore it unless asked".
00:01:53Now with this setup in place, whenever we asked any question in cowork, it loaded the manifest.md
00:01:59file first and from that located the file which contained the data needed and then it responded
00:02:03to our query much faster and more reliably than working without it.
00:02:07Now aside from the manifest.md, you need to create 3 more context files that tell what
00:02:12your identity is.
00:02:13These files are about me, brand voice and working style, each explaining how you prefer responses
00:02:19so that Claude knows how to behave.
00:02:20This eliminates generic AI output because Claude actually knows what your working style is.
00:02:25So we placed these files in the Claude context folder inside the Documents folder and had
00:02:29them accessible from everywhere by telling Claude in the instructions.
00:02:33This ensures that Claude responds according to what we need and does not behave in a way
00:02:38we don't like.
00:02:39These files are not meant to be created once and then used forever.
00:02:42They need to be refined frequently and if you see that Claude didn't follow the instructions
00:02:46you gave in your files, then iterate with it on whether it's a prompt problem or a context
00:02:51problem.
00:02:52If it's either, you can add lines to these files to fix things.
00:02:54Now aside from these files, you need to create memory files so that if you're working continuously
00:02:59in a particular folder, it retains memory between sessions from these files.
00:03:03This works similarly to coding, letting files act as our external memory for all the decisions
00:03:08and tasks that need to be done.
00:03:10The next thing is something that people often ignore, the global instructions.
00:03:13Many people just leave this blank, but they're actually powerful because these instructions
00:03:17are loaded before anything else, even before your prompt is loaded.
00:03:21They act as a starting point for all of your prompts.
00:03:23For Claude code, this looks like the instructions in the Claude.md file in the .Claude folder
00:03:28of the home directory.
00:03:30In my global instructions, I specifically stated that the manifest.md is the first thing Claude
00:03:35should look at and how to navigate around it.
00:03:37But there are also other practices that make working with Claude manageable.
00:03:41For example, I let Claude ask clarifying questions before doing anything.
00:03:45This way it doesn't blindly do whatever it thinks is right and can course correct with
00:03:48relevant questions.
00:03:50Another thing to include in your global settings is asking Claude to show a brief plan before
00:03:54taking action.
00:03:55When it lays out a plan first, you can actually see if the direction is right or not.
00:03:59You can add other rules as you like.
00:04:00For example, I added instructions to avoid filler words and not to pad the output, which
00:04:05Claude tends to do normally.
00:04:06I also explicitly stated that if the confidence is low, Claude should ask instead of giving
00:04:11wrong answers confidently.
00:04:12All of these contribute to a much better experience with cowork.
00:04:16Even with vague prompts, this setup makes it answer accurately.
00:04:19And as I already mentioned, we use Claude context files to guide the voice and personality, so
00:04:24I also included this in the global instructions so it can reference them whenever needed.
00:04:28Even though this is something you must be hearing in our videos repeatedly, you also need to
00:04:32ensure that the context given to your agent is minimal, either by explicitly stating it
00:04:36in the prompt or by controlling it with files like the manifest.md or similar.
00:04:41The less the context window is bloated with noise, the better it performs.
00:04:44Now the prompts, setup instructions and templates are available in AI Labs Pro.
00:04:48For those who don't know, it's our recently launched community where you get ready to use
00:04:52templates that you can plug directly into your projects for this video and all previous ones.
00:04:57If you've found value in what we do and want to support the channel, this is the best way
00:05:01to do it, the link's in the description.
00:05:03Now another thing we need to do is define the end state of what you want to achieve instead
00:05:07of defining the process.
00:05:09As we always say, if we show the model what the correct output looks like, it tends to
00:05:13perform better and iterate toward that goal.
00:05:16That correct output can be anything, test cases, the final output in the prompt or similar references.
00:05:21Now this principle applies to all agents, be it cowork or Claude code or any other agent.
00:05:26Now when we wanted to perform a reorganization task in our folder, we specifically stated
00:05:31which version of each file should go into which folder and what each folder should contain
00:05:36after the reorganization was done instead of vaguely mentioning it to reorganize the files.
00:05:40We also detailed how it should treat nested folders and explicitly mentioned what it should
00:05:45not touch.
00:05:46This prompt allowed Claude to iterate toward that goal in an orderly way, making the task
00:05:50much easier because now it knew what the correct output looked like.
00:05:54We need to explicitly tell Claude what it should do if it's uncertain about any task.
00:05:58Most commonly we give Claude clear instructions in our prompts on what it needs to do and
00:06:03the best path to follow, but we don't mention how it should handle edge cases.
00:06:06In those situations, Claude tends to guess, and most of the time, it's wrong because
00:06:10it doesn't know our preferred approach.
00:06:12So you need to specifically state what it should do in those situations.
00:06:16We did so by adding this in our global instructions that if Claude is uncertain about anything,
00:06:21it should ask by mentioning it in words, and if the confidence is low, it should say so.
00:06:25In our document folders under Working Style, we also specified that if it's unsure about
00:06:30something, it should say so and explicitly not guess or present it as a fact.
00:06:34Now with this in place, Claude flags uncertainty up front instead of confidently guessing wrong.
00:06:39But before we move forwards, let's have a word with the sponsor, Scrimba.
00:06:42Most of us learn to code by watching a video, getting stuck, and constantly switching between
00:06:46a browser and an editor until our brains melt.
00:06:49Scrimba fixes this.
00:06:50They created Scrim technology, where the video player is actually a live code editor.
00:06:54At any moment, you can hit pause, click directly into the instructor's code, and start editing
00:06:58right there to see what happens.
00:07:00It's like pair programming with experts, and that's what makes the learning actually
00:07:04stick.
00:07:05Scrimba offers the specialized training needed to master AI engineering and full stack development
00:07:09for a high quality portfolio.
00:07:11If you're a student or prepping for interviews, these deep dives help you prepare for technical
00:07:15screens covering data structures in Git.
00:07:18It's the most efficient way to bridge the gap from vibe coding to professional engineering.
00:07:22Stop watching passive tutorials and start gaining real world experience through interactive
00:07:26building today.
00:07:27Get started today with their free courses and when you're ready, use our link in the pinned
00:07:31comment below to save an extra 20% on their pro plans.
00:07:35Instead of using a different session for each task, you need to batch related work into a
00:07:39single session.
00:07:40Now how do you identify which tasks can be grouped into a single session and which cannot?
00:07:45The first clue is that some tasks actually share context among them because the output
00:07:49of one task is fed as input to the next and so on.
00:07:51For example, generating the monthly budget summary report often involves multiple interconnected
00:07:56tasks.
00:07:57In such cases, we need to group similar tasks together so they run faster, cheaper, and with
00:08:01higher quality.
00:08:02This also helps prevent hitting session limits frequently since you're completing more tasks
00:08:06in fewer sessions.
00:08:08When we gave Claude prompts, we explicitly started with a goal, then mentioned the first
00:08:12step it needed to do, then the next, and so on until the goal was achieved.
00:08:16This approach allowed us to complete more tasks much faster.
00:08:19However, if the tasks are not interconnected, batching them together will not only waste
00:08:23tokens but may also result in incorrect output if done this way.
00:08:26Now batching of tasks doesn't have to be done only sequentially.
00:08:29If there are tasks that can be done in parallel, you can integrate them by utilizing parallel
00:08:34agents.
00:08:35Claude can automatically identify the need for parallelism and execute it on its own.
00:08:39But it doesn't hurt to explicitly mention this in your prompt.
00:08:42We also use subagents heavily to make our tasks faster and more convenient.
00:08:46With subagents, a large number of tasks can be completed quickly and their dedicated context
00:08:50windows prevent the main context from being bloated with unnecessary information.
00:08:55However, one thing to watch out for is that subagents consume a lot of tokens, so you need
00:08:59to use them only when it's absolutely necessary.
00:09:02Also, if you are enjoying our content, consider pressing the hype button because it helps us
00:09:06create more content like this and reach out to more people.
00:09:10Co-work has an edge because we can schedule tasks, which we used to do manually by giving
00:09:14prompts repeatedly.
00:09:15Now we can schedule a range of tasks that we perform every day.
00:09:18These scheduled tasks only run when your computer is awake and while the Claude desktop is open
00:09:23so that's an important consideration.
00:09:24Since we already had a system running all the time, running Open Claude to research new ideas,
00:09:29track new tool releases, and report to us in our Discord channel, we asked Co-work to schedule
00:09:34another automation.
00:09:35Now we use the schedule skill and ask Claude to analyze the meeting notes where we discuss
00:09:39new ideas and tool releases and to write a report based on those notes for the same day
00:09:44in this folder.
00:09:45We also gave it a proper file naming format and asked it to identify the actionable items.
00:09:49In response to this prompt, Claude asked us questions about the frequency and then scheduled
00:09:54the task for us.
00:09:55And now we receive frequent reports from our meeting notes, ideas, and the tools that we
00:09:59can use in our videos, all derived from our discussions.
00:10:02This process can be improved even further by using connectors to link Gmail or Google Drive,
00:10:07allowing us to write emails or save files directly to the inbox.
00:10:11We can also do this by setting up cron jobs in Claude code and letting it interact with
00:10:15MCP tools and CLIs to do the same job.
00:10:18To make our workflow much more efficient, we need to use plugins to compound capabilities.
00:10:22Each plugin essentially contains a bundle of skills or commands, along with sub-agent integrations,
00:10:27all targeted toward a specific area and specialized in working within that domain because they
00:10:32include tailored instructions.
00:10:34Claude already has many plugins built for common use cases, but we can also create our own.
00:10:38These plugins are open source and available on GitHub.
00:10:41Now the plugin suite even contains a plugin to create plugins.
00:10:44When we wanted to create a plugin of our own, we simply asked the chat interface to do so
00:10:49and it ran the skill for building the new plugin.
00:10:51Claude asked us a set of questions in a session and then presented a plan.
00:10:55Once we approved the plan, it started building everything.
00:10:58This makes the process even easier because now we don't have to rely on plugins built
00:11:02by others, we can create our own, specifically tailored for our unique use cases.
00:11:07Another thing worth mentioning is using skills.
00:11:09We've already talked in detail about how to build a good skill and walked you through
00:11:13the process of creating new skills, including how to handle the problems we encountered
00:11:17while building our own.
00:11:18You can check out those guides on our channel, they'll help you when building skills of
00:11:22your own.
00:11:23Claude also comes with many built-in skills tailored for commonly used tasks, but we can
00:11:27create custom skills specifically designed for our own unique use cases.
00:11:31Finally, we have to treat cowork as an employee, not a toy.
00:11:35Cowork is still a research preview with limited guardrails, which means it can modify things
00:11:39that shouldn't be modified if not properly restricted.
00:11:42We need to give it clear boundaries to make the most of it.
00:11:45Sensitive data should be kept in separate folders, exposing only what is actually needed, ensuring
00:11:49that cowork does not touch private information.
00:11:52We also need to tightly scope its tasks to ensure good performance.
00:11:56For example, adding instructions like "don't delete anything" ensures that it won't
00:12:00delete files and will ask before removing anything if necessary, like how we were doing when we
00:12:04were prompting it.
00:12:05There is also a risk of prompt injection.
00:12:07If a document or website contains harmful instructions, cowork might execute them and cause issues.
00:12:12Additionally, cowork uses more resources than normal chat, so if you use it excessively,
00:12:17your context limit will be reached quickly.
00:12:19You need to harness it carefully to make the most out of it.
00:12:22That brings us to the end of this video.
00:12:23If you'd like to support the channel and help us keep making videos like this, you can do
00:12:27so by using the super thanks button below.
00:12:30As always, thank you for watching and I'll see you in the next one.

Key Takeaway

Maximizing the efficiency of AI agents like Co-work and Claude Code requires a rigorous pre-setup phase involving structured manifest files, identity-driven context, and clearly defined global instructions to ensure reliable, high-quality outcomes.

Highlights

Implementation of a 'manifest.md' file to provide structural guidance and source-of-truth hierarchy for AI agents.

Creation of identity-based context files including 'About Me', 'Brand Voice', and 'Working Style' to eliminate generic AI outputs.

Strategic use of global instructions to establish ground rules like asking clarifying questions and providing plans before action.

Optimization of session limits and token usage through task batching, parallel agents, and specialized sub-agents.

Advanced automation techniques using the 'schedule' skill, connectors for Gmail/Google Drive, and custom-built plugins.

Emphasis on security and efficiency by treating AI agents as employees with restricted access to sensitive data.

Timeline

Introduction to Co-work and the Importance of Pre-setup

The speaker introduces Co-work as a powerful tool that brings cloud code capabilities to non-developers for automation and research tasks. Despite its power, many users experience low-quality results and quickly exhausted usage limits due to a lack of preparation. This section argues that there is no single 'right' way to set up, but rather a series of tailored steps necessary for success. The speaker promises to share high-impact best practices discovered through their own workflow iterations. This foundation sets the stage for the specific file-based strategies discussed later in the video.

The Manifest File: Guiding Folder Navigation

A critical first step is creating a 'manifest.md' file in the root of every work folder to act as a map for the AI. This file prevents Claude from pulling noise from irrelevant files by defining which documents are the 'source of truth' and which should be skipped. The speaker introduces a three-tier system for files: Tier 1 for essential data, Tier 2 for on-demand loading, and Tier 3 for ignored archives. Using this manifest allows the agent to locate data much faster and respond with significantly higher reliability. Without it, the context window often becomes bloated with 'guessed' information that leads to poor output.

Identity Context and Global Instructions

Beyond folder structure, the speaker emphasizes creating three identity files: 'About Me', 'Brand Voice', and 'Working Style'. These files live in a dedicated context folder and ensure the AI's behavior aligns with the user's specific personality and professional requirements. Global instructions are highlighted as a powerful tool because they load before any specific prompt, serving as the agent's core operating manual. The speaker shares their personal rules, such as requiring the AI to show a plan before acting and forbidding the use of filler words. These settings ensure that even vague prompts result in accurate, high-confidence responses from the agent.

Defining End States and Task Batching

The strategy shifts toward prompting techniques, specifically defining the 'end state' rather than just the process. By providing examples of what a successful output looks like—such as specific file structures or test cases—the model can iterate more effectively toward that goal. This section also covers the importance of handling uncertainty by instructing the AI to ask questions instead of guessing when confidence is low. The speaker explains how to handle edge cases by explicitly detailing what the AI should not touch during complex tasks like folder reorganization. This level of detail prevents the common pitfall of 'confidently wrong' AI hallucinations.

Session Management, Sub-agents, and Automation

This segment discusses technical optimizations like batching related tasks into single sessions to save tokens and avoid session limits. The speaker introduces the concept of parallel agents and sub-agents, which allow for faster task completion without bloating the main context window. A major highlight is the 'schedule' skill, which allows users to automate daily tasks like analyzing meeting notes and generating reports. The speaker demonstrates how these automations can be further enhanced using connectors for Gmail and Google Drive. These tools transform the agent from a simple chatbot into a continuous, background-running productivity engine.

Plugins, Skills, and Security Best Practices

The final section covers the use of plugins and skills to compound the agent's capabilities, including a 'plugin to create plugins'. The speaker notes that while many built-in skills exist, creating custom ones tailored to unique use cases provides a significant competitive edge. A stern warning is given to treat AI agents like employees by implementing security boundaries and restricting access to sensitive data. Users are cautioned about prompt injection risks and the high resource consumption of tools like Co-work compared to standard chat interfaces. The video concludes by encouraging viewers to harness these tools carefully to maintain efficiency and protect private information.

Community Posts

View all posts