Codebase Singularity

Codebase Singularity

Quick start

1. Collect inputs: repo path, goal, constraints, risk tolerance, and validation commands. 2. Prime: list files, read README, and inspect entry points. 3. Summarize: explain repo shape and where the change belongs; ask questions if ambiguous. 4. Plan: write a short plan with success criteria and validation steps. 5. Change: implement the smallest viable patch; avoid unrelated refactors. 6. Validate: run the requested command(s) or the closest available; record results. 7. Wrap up: report changes, evidence, risks, and next steps.

Inputs

• Repo path and task goal.
• Constraints (files to avoid, time or risk limits, API stability).
• Validation commands (tests, lint, build, or targeted checks).

Outputs

• Short plan with success criteria.
• Minimal patch.
• Evidence: commands run and results.
• Wrap-up: what changed, why, risks, and next steps.

Grades (maturity ladder)

Grade 1: Prime only

• Prime, summarize, and stop on ambiguity.

Grade 2: Specialized roles + verification

• Add role prompts (e.g., test_writer, docs_fetcher, reviewer).
• Require a verification step for each role.

Grade 3: Tooling expansion

• Declare required tools or services and expected outputs.
• Prefer tools over guessing (e.g., query schema instead of inferring).

Grade 4: Closed-loop autonomy

• Run cycles: request -> validate -> resolve.
• Add hard exits: max iterations, evidence of progress, stop after repeated failures.

Guardrails

• Keep diffs small.
• Always validate; if you cannot, say so and propose a safe next step.
• Ask before broad refactors or scope expansion.

Templates

Safe bugfix

Goal: Fix <bug> with minimal change.
Constraints: Do not modify public API; avoid touching <paths>.
Validation: Run `...` and paste output.
Exit conditions: Stop after 2 failed attempts; summarize hypotheses and logs.
Deliverables: Patch + explanation + risks.

Add a feature

Goal: Add <feature>.
Constraints: Keep scope to <module>; no new dependencies.
Validation: Run `...`.
Deliverables: Patch + tests + short doc update.

References

• Read references/video.md for a concise source summary.
• Read references/transcript.md for full context.
• Read references/visual-notes.md for a visual outline.

Transcript (YouTube captions)

• Source: en-orig captions downloaded via yt-dlp
• Video: https://youtu.be/fop_yxV-mPo?si=ufh6xRFv874mU6Ud
• Local captions: /Users/vishal/clawd/videos/*_fop_yxV-mPo.en-orig.vtt (raw), /Users/vishal/clawd/videos/*_fop_yxV-mPo.en.vtt

Note: This is a caption-derived transcript (not Whisper). If we later rerun Whisper successfully, we can replace this file.

Captions

00:00:00.160 --> 00:00:03.429 align:start position:0% There is one mental framework that sits

00:00:03.439 --> 00:00:06.470 align:start position:0% at the center. An idea so important that

00:00:06.480 --> 00:00:08.950 align:start position:0% if you capture it, it can change the way

00:00:08.960 --> 00:00:12.070 align:start position:0% you engineer forever. The agentic layer.

00:00:12.080 --> 00:00:15.110 align:start position:0% This is the new ring around your

00:00:15.120 --> 00:00:17.590 align:start position:0% codebase where you teach your agents to

00:00:17.600 --> 00:00:20.470 align:start position:0% operate your application on your behalf

00:00:20.480 --> 00:00:23.509 align:start position:0% as well and even better than you and

00:00:23.519 --> 00:00:25.830 align:start position:0% your team ever could. Focusing on

00:00:25.840 --> 00:00:27.670 align:start position:0% building the agentic layer of your

00:00:27.680 --> 00:00:30.790 align:start position:0% codebase is the highest return on

00:00:30.800 --> 00:00:33.750 align:start position:0% investment action for any engineer in

00:00:33.760 --> 00:00:36.389 align:start position:0% the age of agents we live in. Why is

00:00:36.399 --> 00:00:38.950 align:start position:0% that? As you know, when you scale your

00:00:38.960 --> 00:00:42.069 align:start position:0% compute, you scale your impact. We're

00:00:42.079 --> 00:00:44.709 align:start position:0% not just AI coding anymore. Our agents

00:00:44.719 --> 00:00:47.670 align:start position:0% can take actions on our behalf. And this

00:00:47.680 --> 00:00:50.630 align:start position:0% has changed engineering forever. Once

00:00:50.640 --> 00:00:52.790 align:start position:0% you've built a sufficiently powerful

00:00:52.800 --> 00:00:55.670 align:start position:0% agentic layer, something will happen.

00:00:55.680 --> 00:00:58.069 align:start position:0% The codebase singularity. In this

00:00:58.079 --> 00:01:00.869 align:start position:0% moment, you, the engineer, will realize

00:01:00.879 --> 00:01:04.469 align:start position:0% one simple fact. My agents can now run

00:01:04.479 --> 00:01:07.429 align:start position:0% my codebase better than I can. I trust

00:01:07.439 --> 00:01:10.070 align:start position:0% them to ship more than I trust myself or

00:01:10.080 --> 00:01:12.789 align:start position:0% my team. Nothing ships to production

00:01:12.799 --> 00:01:16.390 align:start position:0% without my teams of agents. Yes, this

00:01:16.400 --> 00:01:18.230 align:start position:0% might sound crazy. It might sound far

00:01:18.240 --> 00:01:20.550 align:start position:0% out, but if you've been pushing what you

00:01:20.560 --> 00:01:22.149 align:start position:0% can do with agents, if you've been

00:01:22.159 --> 00:01:24.789 align:start position:0% putting these tactics to work already,

00:01:24.799 --> 00:01:27.350 align:start position:0% maybe you already see this future on the

00:01:27.360 --> 00:01:29.990 align:start position:0% horizon, the agentic horizon. I can tell

00:01:30.000 --> 00:01:32.310 align:start position:0% you this for certain right now. There is

00:01:32.320 --> 00:01:35.670 align:start position:0% an agentic layer that could exist inside

00:01:35.680 --> 00:01:39.109 align:start position:0% your codebase so powerful that your

00:01:39.119 --> 00:01:42.710 align:start position:0% codebase runs itself. The only question

00:01:42.720 --> 00:01:46.310 align:start position:0% now is, do you know how to build it? As

00:01:46.320 --> 00:01:49.510 align:start position:0% we work through building agentic layers,

00:01:49.520 --> 00:01:51.990 align:start position:0% keep this idea in your mind and think

00:01:52.000 --> 00:01:54.149 align:start position:0% through what it would take for you to

00:01:54.159 --> 00:01:57.190 align:start position:0% trust your agents to run your codebase

00:01:57.200 --> 00:01:59.990 align:start position:0% better than you could from prompt to

00:02:00.000 --> 00:02:01.910 align:start position:0% production. In this lesson, we're going

00:02:01.920 --> 00:02:04.149 align:start position:0% to put together the missing pieces and

00:02:04.159 --> 00:02:06.870 align:start position:0% see how we can attain the codebase

00:02:06.880 --> 00:02:11.430 align:start position:0% singularity.

00:02:11.440 --> 00:02:15.030 align:start position:0% There are three concrete classes of the

00:02:15.040 --> 00:02:18.309 align:start position:0% agentic layer. Class one, class 2, and

00:02:18.319 --> 00:02:21.589 align:start position:0% class three. Each defined by a unique

00:02:21.599 --> 00:02:23.350 align:start position:0% element that makes them distinct from

00:02:23.360 --> 00:02:26.070 align:start position:0% the rest. Our whole goal here is to look

00:02:26.080 --> 00:02:28.710 align:start position:0% at the components, the raw elements that

00:02:28.720 --> 00:02:30.790 align:start position:0% make up the new ring around your

00:02:30.800 --> 00:02:33.190 align:start position:0% codebase where you have agents drive

00:02:33.200 --> 00:02:35.589 align:start position:0% your engineering experience and you

00:02:35.599 --> 00:02:38.070 align:start position:0% drive your agents. the green squares

00:02:38.080 --> 00:02:40.550 align:start position:0% here, the outer layer, this is the new

00:02:40.560 --> 00:02:43.270 align:start position:0% aentic layer. The inner layer here,

00:02:43.280 --> 00:02:46.070 align:start position:0% right, these dark squares, this is going

00:02:46.080 --> 00:02:48.550 align:start position:0% to be your application layer. And we're

00:02:48.560 --> 00:02:50.550 align:start position:0% bundling a ton of things underneath your

00:02:50.560 --> 00:02:51.830 align:start position:0% application layer. We're talking about

00:02:51.840 --> 00:02:53.589 align:start position:0% your database, your front end, your

00:02:53.599 --> 00:02:55.589 align:start position:0% backend, your scripts, all the

00:02:55.599 --> 00:02:57.430 align:start position:0% application stuff, even your DevOps

00:02:57.440 --> 00:02:59.270 align:start position:0% stuff, right? This all goes underneath

00:02:59.280 --> 00:03:01.350 align:start position:0% the application layer. Why is that? It's

00:03:01.360 --> 00:03:03.589 align:start position:0% because we want to be able to bundle

00:03:03.599 --> 00:03:06.149 align:start position:0% your different repositories underneath

00:03:06.159 --> 00:03:08.390 align:start position:0% your agentic layer so that we can do

00:03:08.400 --> 00:03:11.110 align:start position:0% something like this. Code bases often

00:03:11.120 --> 00:03:14.309 align:start position:0% contain more than one application. So by

00:03:14.319 --> 00:03:16.790 align:start position:0% bundling your agentic layer around your

00:03:16.800 --> 00:03:18.869 align:start position:0% applications, your agents can

00:03:18.879 --> 00:03:21.509 align:start position:0% effectively see everything. Every class

00:03:21.519 --> 00:03:24.309 align:start position:0% is going to have one to n grades that

00:03:24.319 --> 00:03:26.229 align:start position:0% will give you a rough understanding of

00:03:26.239 --> 00:03:28.630 align:start position:0% how powerful your agentic layer is.

00:03:28.640 --> 00:03:29.990 align:start position:0% You're going to be able to quickly

00:03:30.000 --> 00:03:31.830 align:start position:0% identify where you are as we move

00:03:31.840 --> 00:03:34.229 align:start position:0% through classes and grades so that you

00:03:34.239 --> 00:03:36.229 align:start position:0% can make improvements and get to that

00:03:36.239 --> 00:03:37.990 align:start position:0% next level. And then once you reach the

00:03:38.000 --> 00:03:40.149 align:start position:0% limit of a certain class, look to make

00:03:40.159 --> 00:03:43.430 align:start position:0% that next jump uniquely identified by a

00:03:43.440 --> 00:03:45.509 align:start position:0% new dimension that we'll talk about as

00:03:45.519 --> 00:03:47.589 align:start position:0% we progress from grade to grade from

00:03:47.599 --> 00:03:49.750 align:start position:0% class to class all the way up to class

00:03:49.760 --> 00:03:53.030 align:start position:0% three and the final grade. Now, let's

00:03:53.040 --> 00:03:55.350 align:start position:0% start by looking at a powerful version

00:03:55.360 --> 00:03:58.710 align:start position:0% of a class 3 agentic system with an

00:03:58.720 --> 00:04:00.949 align:start position:0% orchestrator guiding the workflow.

00:04:00.959 --> 00:04:02.949 align:start position:0% You've seen powerful systems like this

00:04:02.959 --> 00:04:05.350 align:start position:0% in our previous two agentic horizon

00:04:05.360 --> 00:04:07.030 align:start position:0% lessons. What we're going to do now is

00:04:07.040 --> 00:04:10.390 align:start position:0% run two brand new workflows and combine

00:04:10.400 --> 00:04:12.149 align:start position:0% some really big ideas. I'm going to open

00:04:12.159 --> 00:04:14.390 align:start position:0% up the prompt interface here as usual.

00:04:14.400 --> 00:04:16.629 align:start position:0% And you'll notice we have one new

00:04:16.639 --> 00:04:19.030 align:start position:0% section. Our orchestrator agent can now

00:04:19.040 --> 00:04:22.230 align:start position:0% kick off AI developer workflows. I'll

00:04:22.240 --> 00:04:23.749 align:start position:0% paste in this prompt here. And I'm

00:04:23.759 --> 00:04:25.350 align:start position:0% commanding the orchestrator agent to

00:04:25.360 --> 00:04:27.590 align:start position:0% kick off this specific workflow in this

00:04:27.600 --> 00:04:30.629 align:start position:0% specific directory with just a single

00:04:30.639 --> 00:04:32.230 align:start position:0% simple prompt. We're not over

00:04:32.240 --> 00:04:33.590 align:start position:0% complicating this. I just want to show

00:04:33.600 --> 00:04:35.830 align:start position:0% you so that you have a concrete vision

00:04:35.840 --> 00:04:38.150 align:start position:0% of the types of systems you can build

00:04:38.160 --> 00:04:39.510 align:start position:0% out. So we're going to fire this off

00:04:39.520 --> 00:04:41.270 align:start position:0% here. Let's go ahead and run another

00:04:41.280 --> 00:04:44.230 align:start position:0% build out a markdown preview application

00:04:44.240 --> 00:04:47.110 align:start position:0% here in one shot with our plan, build,

00:04:47.120 --> 00:04:49.590 align:start position:0% review, and fix. And this is an AI

00:04:49.600 --> 00:04:51.350 align:start position:0% developer workflow. So the orchestrator

00:04:51.360 --> 00:04:52.950 align:start position:0% doesn't need to do anything, right? This

00:04:52.960 --> 00:04:55.110 align:start position:0% runs end to end. We're going to run one

00:04:55.120 --> 00:04:56.550 align:start position:0% more workflow here. And then I want to

00:04:56.560 --> 00:04:58.230 align:start position:0% show you something awesome. You've seen

00:04:58.240 --> 00:05:00.950 align:start position:0% the generic log view. We have pushed our

00:05:00.960 --> 00:05:03.350 align:start position:0% orchestrator to be able to control AI

00:05:03.360 --> 00:05:06.230 align:start position:0% developer workflows. We now have two

00:05:06.240 --> 00:05:08.230 align:start position:0% workflows running. So plan, build,

00:05:08.240 --> 00:05:09.909 align:start position:0% review, fix, and just a simple plan

00:05:09.919 --> 00:05:11.590 align:start position:0% build. We have given our orchestrator

00:05:11.600 --> 00:05:14.230 align:start position:0% agent the ability to run arbitrary

00:05:14.240 --> 00:05:17.350 align:start position:0% endtoend workflows. Now let's dial in

00:05:17.360 --> 00:05:20.310 align:start position:0% and start from zero, right? Let's reset

00:05:20.320 --> 00:05:21.909 align:start position:0% because we don't start here. We don't

00:05:21.919 --> 00:05:23.590 align:start position:0% start at class 3. We don't start with

00:05:23.600 --> 00:05:25.350 align:start position:0% powerful orchestration systems. We start

00:05:25.360 --> 00:05:29.189 align:start position:0% with nothing. So as mentioned, every

00:05:29.199 --> 00:05:31.270 align:start position:0% single codebase should now have these

00:05:31.280 --> 00:05:33.830 align:start position:0% two key components. The application

00:05:33.840 --> 00:05:35.990 align:start position:0% layer and the agentic layer. These two

00:05:36.000 --> 00:05:38.310 align:start position:0% squares here represent a brand new

00:05:38.320 --> 00:05:40.390 align:start position:0% codebase. Now, oftent times you'll be

00:05:40.400 --> 00:05:42.550 align:start position:0% operating in an existing codebase that

00:05:42.560 --> 00:05:43.909 align:start position:0% will look like this, right? You'll

00:05:43.919 --> 00:05:46.310 align:start position:0% already have some pieces, right?

00:05:46.320 --> 00:05:48.469 align:start position:0% Modules, files of your application layer

00:05:48.479 --> 00:05:50.629 align:start position:0% and then you'll start adding on your

00:05:50.639 --> 00:05:53.350 align:start position:0% agentic layer around your codebase. Now,

00:05:53.360 --> 00:05:56.070 align:start position:0% what does this really look like? Let's

00:05:56.080 --> 00:05:57.749 align:start position:0% start from a green field brand new

00:05:57.759 --> 00:06:00.550 align:start position:0% application just to make this all clear.

00:06:00.560 --> 00:06:03.350 align:start position:0% What exactly is in our agentic layer?

00:06:03.360 --> 00:06:06.469 align:start position:0% This is our class one grade one agentic

00:06:06.479 --> 00:06:09.110 align:start position:0% layer. This the thinnest possible

00:06:09.120 --> 00:06:10.950 align:start position:0% agentic layer. You can have a small

00:06:10.960 --> 00:06:12.950 align:start position:0% amount of code and you have a prime

00:06:12.960 --> 00:06:15.830 align:start position:0% prompt andor memory files. It doesn't

00:06:15.840 --> 00:06:18.309 align:start position:0% get thinner than this. If you have this,

00:06:18.319 --> 00:06:20.629 align:start position:0% you technically have an agentic layer.

00:06:20.639 --> 00:06:23.430 align:start position:0% Now, let's dive into a concrete codebase

00:06:23.440 --> 00:06:24.950 align:start position:0% structure to see what this looks like.

00:06:24.960 --> 00:06:26.469 align:start position:0% This is an example project structure.

00:06:26.479 --> 00:06:27.670 align:start position:0% We're going to work through several of

00:06:27.680 --> 00:06:29.749 align:start position:0% these to make it absolutely clear what

00:06:29.759 --> 00:06:32.070 align:start position:0% it looks like to have an agentic layer

00:06:32.080 --> 00:06:33.670 align:start position:0% surrounding your application layer. All

00:06:33.680 --> 00:06:35.350 align:start position:0% right, so we have application layer

00:06:35.360 --> 00:06:37.430 align:start position:0% files, folders, directories, so on and

00:06:37.440 --> 00:06:39.510 align:start position:0% so forth. And we have our agentic layer

00:06:39.520 --> 00:06:41.830 align:start position:0% files. If we open up dogcloud here, do

00:06:41.840 --> 00:06:44.070 align:start position:0% commands, we can see we have this prime

00:06:44.080 --> 00:06:45.909 align:start position:0% command. As you know, the prime command

00:06:45.919 --> 00:06:48.710 align:start position:0% effectively represents a memory file.

00:06:48.720 --> 00:06:50.230 align:start position:0% But the big difference here is that you

00:06:50.240 --> 00:06:52.070 align:start position:0% can activate this whenever you want to

00:06:52.080 --> 00:06:54.150 align:start position:0% and you can fine-tune the prime command

00:06:54.160 --> 00:06:57.110 align:start position:0% to run any specific workflow you want

00:06:57.120 --> 00:06:59.029 align:start position:0% to. Okay, so this is just a prompt that

00:06:59.039 --> 00:07:01.430 align:start position:0% effectively activates your agent to read

00:07:01.440 --> 00:07:03.110 align:start position:0% specific files, right? So there's

00:07:03.120 --> 00:07:04.230 align:start position:0% nothing new here, right? We have a

00:07:04.240 --> 00:07:06.790 align:start position:0% classic agentic prompt doc of course is

00:07:06.800 --> 00:07:08.390 align:start position:0% our memory file that's always going to

00:07:08.400 --> 00:07:09.909 align:start position:0% load. You've seen this, you understand

00:07:09.919 --> 00:07:11.990 align:start position:0% this, you know this. And oftent times I

00:07:12.000 --> 00:07:13.670 align:start position:0% like to take the agents.mdia from

00:07:13.680 --> 00:07:15.029 align:start position:0% working with other agents and just

00:07:15.039 --> 00:07:17.270 align:start position:0% reference that claw.md memory file. This

00:07:17.280 --> 00:07:19.189 align:start position:0% is the simplest possible agentic layer.

00:07:19.199 --> 00:07:21.110 align:start position:0% Now in our application layer, if we're

00:07:21.120 --> 00:07:22.710 align:start position:0% going to user management where we have a

00:07:22.720 --> 00:07:24.710 align:start position:0% backend for a user management system.

00:07:24.720 --> 00:07:26.070 align:start position:0% This is just a starting place, right?

00:07:26.080 --> 00:07:27.350 align:start position:0% This is a great way to when you're

00:07:27.360 --> 00:07:29.029 align:start position:0% creating a new application, just start

00:07:29.039 --> 00:07:30.950 align:start position:0% with something an agent can come in here

00:07:30.960 --> 00:07:32.629 align:start position:0% and understand how to operate this right

00:07:32.639 --> 00:07:34.230 align:start position:0% away with the memory file and with the

00:07:34.240 --> 00:07:35.749 align:start position:0% prime command. Okay, so as we work

00:07:35.759 --> 00:07:37.830 align:start position:0% through classes and grades, we'll talk

00:07:37.840 --> 00:07:39.589 align:start position:0% about the compute advantage you get by

00:07:39.599 --> 00:07:41.589 align:start position:0% scaling to this class and grade of a

00:07:41.599 --> 00:07:43.029 align:start position:0% gentic layer. And then we'll of course

00:07:43.039 --> 00:07:44.629 align:start position:0% talk about trade-offs. So the compute

00:07:44.639 --> 00:07:46.550 align:start position:0% advantage here is obvious, right? We get

00:07:46.560 --> 00:07:48.230 align:start position:0% a clean, minimal setup. It's a great

00:07:48.240 --> 00:07:50.150 align:start position:0% foundation for agenda growth and agents

00:07:50.160 --> 00:07:52.150 align:start position:0% understand context immediately. It's a

00:07:52.160 --> 00:07:54.150 align:start position:0% super small code base. This is easily

00:07:54.160 --> 00:07:56.710 align:start position:0% reproducible by anyone. There's no value

00:07:56.720 --> 00:07:58.629 align:start position:0% in a codebase this small. It's useless

00:07:58.639 --> 00:08:00.710 align:start position:0% for large code bases. Limited capability

00:08:00.720 --> 00:08:02.550 align:start position:0% and there are many leverage points of

00:08:02.560 --> 00:08:04.629 align:start position:0% aentic coding that the class one grade 1

00:08:04.639 --> 00:08:07.029 align:start position:0% aentic layer purely misses. Let's scale

00:08:07.039 --> 00:08:10.070 align:start position:0% this up. Now we have the class one grade

00:08:10.080 --> 00:08:12.230 align:start position:0% two. This is where we start surrounding

00:08:12.240 --> 00:08:14.710 align:start position:0% our codebase with more useful leverage

00:08:14.720 --> 00:08:16.869 align:start position:0% points of agent coding. Here we

00:08:16.879 --> 00:08:19.270 align:start position:0% incorporate specialized prompts to help

00:08:19.280 --> 00:08:21.350 align:start position:0% us plan work. We have a prime command

00:08:21.360 --> 00:08:24.550 align:start position:0% and we now have access to agents also

00:08:24.560 --> 00:08:27.029 align:start position:0% known as sub aents. Here we can start

00:08:27.039 --> 00:08:29.189 align:start position:0% incrementally improving what our

00:08:29.199 --> 00:08:31.110 align:start position:0% codebase can do. So let's go ahead and

00:08:31.120 --> 00:08:32.389 align:start position:0% look at what this codebase looks like

00:08:32.399 --> 00:08:33.990 align:start position:0% now. Right, we have a new specs

00:08:34.000 --> 00:08:35.990 align:start position:0% directory where our plans are written.

00:08:36.000 --> 00:08:37.350 align:start position:0% You're very familiar with that pattern.

00:08:37.360 --> 00:08:39.670 align:start position:0% We also have AI docs so our agents can

00:08:39.680 --> 00:08:41.990 align:start position:0% pull in documentation to aid their

00:08:42.000 --> 00:08:44.389 align:start position:0% context. This is all part of the agentic

00:08:44.399 --> 00:08:45.910 align:start position:0% layer. And our user application is

00:08:45.920 --> 00:08:47.590 align:start position:0% starting to grow. We have a few tests.

00:08:47.600 --> 00:08:49.269 align:start position:0% We have modules. But the most important

00:08:49.279 --> 00:08:51.269 align:start position:0% pieces are here. In addition to our

00:08:51.279 --> 00:08:53.750 align:start position:0% commands, right, our raw prompts, we now

00:08:53.760 --> 00:08:57.030 align:start position:0% have some sub aents or just some agents

00:08:57.040 --> 00:08:58.710 align:start position:0% that we can run, right? So for instance,

00:08:58.720 --> 00:09:01.030 align:start position:0% you might have a fetch docs sub aent

00:09:01.040 --> 00:09:02.710 align:start position:0% that will save them in the AI docs

00:09:02.720 --> 00:09:04.470 align:start position:0% directory and you might have a test

00:09:04.480 --> 00:09:06.150 align:start position:0% writer. These are just simple examples

00:09:06.160 --> 00:09:08.389 align:start position:0% of starting to scale up your agentic

00:09:08.399 --> 00:09:09.910 align:start position:0% layer. You likely have something more

00:09:09.920 --> 00:09:11.430 align:start position:0% powerful than this, but we're just

00:09:11.440 --> 00:09:13.110 align:start position:0% working our way up. And now you might

00:09:13.120 --> 00:09:15.110 align:start position:0% have something like a brand new plan

00:09:15.120 --> 00:09:16.790 align:start position:0% file. So when you're doing work, you

00:09:16.800 --> 00:09:18.949 align:start position:0% might prompt this plan file and then

00:09:18.959 --> 00:09:20.790 align:start position:0% execute it. Very simple. We're still

00:09:20.800 --> 00:09:22.949 align:start position:0% just stacking things up, slowly building

00:09:22.959 --> 00:09:24.790 align:start position:0% up our agentic layer. This is a simple

00:09:24.800 --> 00:09:27.190 align:start position:0% class one grid 2 agentic layer. The big

00:09:27.200 --> 00:09:28.630 align:start position:0% advantages here is that we have

00:09:28.640 --> 00:09:30.710 align:start position:0% specialization with sub aents and we can

00:09:30.720 --> 00:09:33.110 align:start position:0% scale this and parallelize certain

00:09:33.120 --> 00:09:34.550 align:start position:0% workflows. All right. Right? And now

00:09:34.560 --> 00:09:35.910 align:start position:0% we're starting to collect documentation

00:09:35.920 --> 00:09:37.990 align:start position:0% for our agents and we're planning our

00:09:38.000 --> 00:09:40.550 align:start position:0% work before we implement. So this is

00:09:40.560 --> 00:09:42.389 align:start position:0% grade two. Many engineers have passed

00:09:42.399 --> 00:09:44.230 align:start position:0% this level. This is not that novel. It's

00:09:44.240 --> 00:09:45.509 align:start position:0% not that interesting. We're still

00:09:45.519 --> 00:09:47.350 align:start position:0% limited in many ways. Right? The big

00:09:47.360 --> 00:09:49.350 align:start position:0% thing we're missing here is custom tools

00:09:49.360 --> 00:09:51.990 align:start position:0% which brings us right to grade three. So

00:09:52.000 --> 00:09:54.070 align:start position:0% in class one, grade three things start

00:09:54.080 --> 00:09:55.910 align:start position:0% getting interesting and our agentic

00:09:55.920 --> 00:09:58.070 align:start position:0% layer starts to wrap our application

00:09:58.080 --> 00:10:00.470 align:start position:0% layer. So here we have three key new

00:10:00.480 --> 00:10:03.990 align:start position:0% pieces. skills, MCP servers, and prime

00:10:04.000 --> 00:10:06.310 align:start position:0% commands with tool access. You likely

00:10:06.320 --> 00:10:07.829 align:start position:0% already know what that is. All these

00:10:07.839 --> 00:10:10.150 align:start position:0% three things give us the same thing.

00:10:10.160 --> 00:10:12.870 align:start position:0% Custom tools that enhance our agents

00:10:12.880 --> 00:10:14.630 align:start position:0% core 4. Let's take a look at what this

00:10:14.640 --> 00:10:16.150 align:start position:0% actually looks like in an agentic layer

00:10:16.160 --> 00:10:18.470 align:start position:0% class one grade three. Now, we have a

00:10:18.480 --> 00:10:20.389 align:start position:0% couple additional base level files. You

00:10:20.399 --> 00:10:23.110 align:start position:0% might have a MCP.json, Postgress,

00:10:23.120 --> 00:10:25.509 align:start position:0% Firecrawl, Jira, Notion, whatever tool

00:10:25.519 --> 00:10:27.110 align:start position:0% you want to use, you can now throw it in

00:10:27.120 --> 00:10:28.630 align:start position:0% here. But that's not all. We know that

00:10:28.640 --> 00:10:30.790 align:start position:0% MCPs are very tokenheavy and we can just

00:10:30.800 --> 00:10:32.870 align:start position:0% prompt engineer properly to get our

00:10:32.880 --> 00:10:34.550 align:start position:0% agents the tools they need. One of the

00:10:34.560 --> 00:10:36.069 align:start position:0% ways that that can be done is with

00:10:36.079 --> 00:10:38.150 align:start position:0% skills. Our agents have learned to

00:10:38.160 --> 00:10:40.230 align:start position:0% migrate the database. And so inside of

00:10:40.240 --> 00:10:41.829 align:start position:0% this file, you know, you've seen these

00:10:41.839 --> 00:10:43.910 align:start position:0% before. We have taught our agent how to

00:10:43.920 --> 00:10:46.150 align:start position:0% use a specific tool and it can now use

00:10:46.160 --> 00:10:48.710 align:start position:0% it to perform a specific task. We're

00:10:48.720 --> 00:10:50.389 align:start position:0% starting to specialize our agents a

00:10:50.399 --> 00:10:52.069 align:start position:0% little bit more outside of our sub

00:10:52.079 --> 00:10:53.590 align:start position:0% agents. But we can also do things like

00:10:53.600 --> 00:10:55.430 align:start position:0% this, right? Start and stop application.

00:10:55.440 --> 00:10:57.350 align:start position:0% If we open up this skill, we are just

00:10:57.360 --> 00:11:00.069 align:start position:0% teaching our agent how to use scripts.

00:11:00.079 --> 00:11:01.590 align:start position:0% There's no need for an MCP server.

00:11:01.600 --> 00:11:03.430 align:start position:0% Oftent times, you can just build your

00:11:03.440 --> 00:11:04.949 align:start position:0% own script that exposes the right

00:11:04.959 --> 00:11:06.790 align:start position:0% functionality. Here we have a start and

00:11:06.800 --> 00:11:08.710 align:start position:0% a stop tool that starts and stops the

00:11:08.720 --> 00:11:10.949 align:start position:0% application. Okay, so we're just scaling

00:11:10.959 --> 00:11:13.269 align:start position:0% up what our agentic layer can do. We're

00:11:13.279 --> 00:11:15.190 align:start position:0% giving our agents more capability,

00:11:15.200 --> 00:11:16.710 align:start position:0% right? We are scaling our compute to

00:11:16.720 --> 00:11:18.389 align:start position:0% scale our impact. Our agents now have

00:11:18.399 --> 00:11:21.190 align:start position:0% access to custom tools, skills, and

00:11:21.200 --> 00:11:22.870 align:start position:0% prime commands, right? And keep in mind,

00:11:22.880 --> 00:11:24.550 align:start position:0% this is something that I want to keep

00:11:24.560 --> 00:11:27.030 align:start position:0% emphasizing. Skills and MCP servers can

00:11:27.040 --> 00:11:29.509 align:start position:0% both be replaced with just a simple

00:11:29.519 --> 00:11:31.509 align:start position:0% prompt. Okay, so we have this prime DB

00:11:31.519 --> 00:11:33.590 align:start position:0% with tools. You can teach your agents

00:11:33.600 --> 00:11:35.910 align:start position:0% how to execute CLI commands. Here we're

00:11:35.920 --> 00:11:37.990 align:start position:0% just using PSQL, right? So that our

00:11:38.000 --> 00:11:39.990 align:start position:0% agent knows how to interact with the

00:11:40.000 --> 00:11:42.150 align:start position:0% user management database for this mock

00:11:42.160 --> 00:11:44.230 align:start position:0% application. And so you can bypass

00:11:44.240 --> 00:11:46.630 align:start position:0% everything by just understanding how the

00:11:46.640 --> 00:11:48.790 align:start position:0% core 4 works, context, model, prompt,

00:11:48.800 --> 00:11:51.590 align:start position:0% and tools. Okay. And notice what we're

00:11:51.600 --> 00:11:53.269 align:start position:0% doing here, right? We're constantly

00:11:53.279 --> 00:11:55.190 align:start position:0% giving our agents more capability.

00:11:55.200 --> 00:11:57.750 align:start position:0% Custom tools is a huge step. Grade three

00:11:57.760 --> 00:12:00.470 align:start position:0% is very important. Now, this is when

00:12:00.480 --> 00:12:01.750 align:start position:0% prompt engineering and context

00:12:01.760 --> 00:12:03.430 align:start position:0% engineering becomes more important.

00:12:03.440 --> 00:12:05.990 align:start position:0% Skills, MCP servers, and prompts must

00:12:06.000 --> 00:12:08.310 align:start position:0% have tools carefully designed. This is

00:12:08.320 --> 00:12:09.829 align:start position:0% where things can start going wrong. And

00:12:09.839 --> 00:12:11.910 align:start position:0% a lot of engineers do get stuck at grade

00:12:11.920 --> 00:12:13.750 align:start position:0% three. They think they get past this,

00:12:13.760 --> 00:12:15.750 align:start position:0% but actually their tools are terrible

00:12:15.760 --> 00:12:17.590 align:start position:0% and they won't scale. And they're

00:12:17.600 --> 00:12:19.829 align:start position:0% chewing up tokens. They're burning cash

00:12:19.839 --> 00:12:21.590 align:start position:0% and they're overengineering their tools,

00:12:21.600 --> 00:12:22.790 align:start position:0% right? Like one of the most common

00:12:22.800 --> 00:12:24.470 align:start position:0% problems is that engineers will have way

00:12:24.480 --> 00:12:26.710 align:start position:0% too many tools. Being able to design and

00:12:26.720 --> 00:12:29.670 align:start position:0% give your agents the right tools via MCP

00:12:29.680 --> 00:12:33.030 align:start position:0% server skills and just raw prompts is a

00:12:33.040 --> 00:12:35.350 align:start position:0% critical skill for every agentic

00:12:35.360 --> 00:12:37.509 align:start position:0% engineer. Let's move on to grade four.

00:12:37.519 --> 00:12:39.750 align:start position:0% How do we scale this up even further?

00:12:39.760 --> 00:12:41.670 align:start position:0% What comes next? Right. So at this

00:12:41.680 --> 00:12:43.269 align:start position:0% point, we start scaling up our prompts

00:12:43.279 --> 00:12:46.310 align:start position:0% and we start building up feedback loops

00:12:46.320 --> 00:12:48.949 align:start position:0% in the codebase. Okay, so this is where

00:12:48.959 --> 00:12:50.629 align:start position:0% things become very powerful and where

00:12:50.639 --> 00:12:52.710 align:start position:0% you can really scale what you can do.

00:12:52.720 --> 00:12:54.069 align:start position:0% So, of course, we have our plan prompt,

00:12:54.079 --> 00:12:55.590 align:start position:0% but now we'll have things like our build

00:12:55.600 --> 00:12:57.509 align:start position:0% prompt, right? A higher order prompt.

00:12:57.519 --> 00:12:59.750 align:start position:0% Grade four is distinct and that you

00:12:59.760 --> 00:13:02.470 align:start position:0% start asking your agents to review their

00:13:02.480 --> 00:13:04.069 align:start position:0% work. This is where you realize that you

00:13:04.079 --> 00:13:07.190 align:start position:0% should always add feedback loops into

00:13:07.200 --> 00:13:09.269 align:start position:0% your agentic layer. You're effectively

00:13:09.279 --> 00:13:11.110 align:start position:0% adding more compute to get more

00:13:11.120 --> 00:13:13.430 align:start position:0% confidence in your agents results. All

00:13:13.440 --> 00:13:16.069 align:start position:0% right, this is the big idea in tactical

00:13:16.079 --> 00:13:18.550 align:start position:0% agentic coding. Lesson five. Right here

00:13:18.560 --> 00:13:20.230 align:start position:0% we're building closed loop prompts. All

00:13:20.240 --> 00:13:21.750 align:start position:0% right, let's understand the application

00:13:21.760 --> 00:13:23.190 align:start position:0% structure. So now we have things like

00:13:23.200 --> 00:13:25.350 align:start position:0% app review. So not only are you planning

00:13:25.360 --> 00:13:26.949 align:start position:0% your work, you're having your agents

00:13:26.959 --> 00:13:29.110 align:start position:0% review the work done and then report it.

00:13:29.120 --> 00:13:30.310 align:start position:0% And you're doing this with powerful

00:13:30.320 --> 00:13:32.870 align:start position:0% prompts like the code review prompt, the

00:13:32.880 --> 00:13:34.389 align:start position:0% review prompt. You're then running

00:13:34.399 --> 00:13:36.710 align:start position:0% reproduce bugs and then things like test

00:13:36.720 --> 00:13:38.470 align:start position:0% backend, test front end. All right, so

00:13:38.480 --> 00:13:40.230 align:start position:0% for instance, what does review do? Code

00:13:40.240 --> 00:13:42.710 align:start position:0% review current changes, review all stage

00:13:42.720 --> 00:13:44.230 align:start position:0% changes, do a bunch of stuff, and then

00:13:44.240 --> 00:13:45.750 align:start position:0% we're going to output some concrete

00:13:45.760 --> 00:13:47.509 align:start position:0% results. You might also do something

00:13:47.519 --> 00:13:49.670 align:start position:0% like reproduce a bug. This is where we

00:13:49.680 --> 00:13:52.310 align:start position:0% would save a concrete resolution file

00:13:52.320 --> 00:13:54.870 align:start position:0% into app reviews. This becomes really

00:13:54.880 --> 00:13:57.189 align:start position:0% powerful when you start giving your sub

00:13:57.199 --> 00:13:59.189 align:start position:0% agents these abilities. Okay, so we can

00:13:59.199 --> 00:14:01.750 align:start position:0% run something like a review agent and

00:14:01.760 --> 00:14:03.829 align:start position:0% you can scale this up and then it will

00:14:03.839 --> 00:14:06.790 align:start position:0% output its review into a specific file.

00:14:06.800 --> 00:14:08.470 align:start position:0% Inside of our application layer here,

00:14:08.480 --> 00:14:10.069 align:start position:0% you'll also notice something. As you

00:14:10.079 --> 00:14:11.590 align:start position:0% start to grow your codebase, your

00:14:11.600 --> 00:14:13.670 align:start position:0% agentic layer and your application layer

00:14:13.680 --> 00:14:15.590 align:start position:0% will be growing side by side. We've now

00:14:15.600 --> 00:14:17.590 align:start position:0% split out our application into client

00:14:17.600 --> 00:14:19.430 align:start position:0% and server. At some point, your small

00:14:19.440 --> 00:14:21.269 align:start position:0% codebase will grow and it will start to

00:14:21.279 --> 00:14:22.790 align:start position:0% fracture. This is important. We'll look

00:14:22.800 --> 00:14:25.110 align:start position:0% at this in the next grade. But with this

00:14:25.120 --> 00:14:27.269 align:start position:0% application directory structure, we're

00:14:27.279 --> 00:14:30.470 align:start position:0% still able to keep all related code for

00:14:30.480 --> 00:14:33.110 align:start position:0% this product under the same code base.

00:14:33.120 --> 00:14:35.269 align:start position:0% Now, how this operates in Git is

00:14:35.279 --> 00:14:36.629 align:start position:0% important. We'll talk about that in just

00:14:36.639 --> 00:14:38.470 align:start position:0% a moment. Since we have both a front end

00:14:38.480 --> 00:14:40.389 align:start position:0% and backend application, you can start

00:14:40.399 --> 00:14:42.389 align:start position:0% specializing your prompts, right? test

00:14:42.399 --> 00:14:45.670 align:start position:0% backend, test front end. Here we have a

00:14:45.680 --> 00:14:47.829 align:start position:0% concrete closed loop prompt. I'm being

00:14:47.839 --> 00:14:49.990 align:start position:0% really verbose here with my language

00:14:50.000 --> 00:14:51.750 align:start position:0% inside of these mock prompts. You know,

00:14:51.760 --> 00:14:53.829 align:start position:0% we're following that three-step workflow

00:14:53.839 --> 00:14:56.389 align:start position:0% from tactical agent to coding lesson 5.

00:14:56.399 --> 00:14:59.110 align:start position:0% Request, validate, resolve. This creates

00:14:59.120 --> 00:15:01.269 align:start position:0% a closed loop where your agent can spin

00:15:01.279 --> 00:15:04.389 align:start position:0% and spin until the job is done. This

00:15:04.399 --> 00:15:07.430 align:start position:0% distinctly marks grade four. Again, this

00:15:07.440 --> 00:15:09.590 align:start position:0% is a place where engineers fall off and

00:15:09.600 --> 00:15:11.750 align:start position:0% they miss out on using more compute to

00:15:11.760 --> 00:15:13.189 align:start position:0% get the job done, right? We're starting

00:15:13.199 --> 00:15:15.189 align:start position:0% to split our prompts to focus on

00:15:15.199 --> 00:15:17.590 align:start position:0% specific areas of this codebase. Okay,

00:15:17.600 --> 00:15:19.110 align:start position:0% so things are going from really generic

00:15:19.120 --> 00:15:21.990 align:start position:0% and vague, right? Plan build to more

00:15:22.000 --> 00:15:24.389 align:start position:0% specific, right? Test our backend, test

00:15:24.399 --> 00:15:26.550 align:start position:0% our front end, review with this code how

00:15:26.560 --> 00:15:28.230 align:start position:0% we like to do it. All right, and the

00:15:28.240 --> 00:15:30.310 align:start position:0% advantages here are really clear, right?

00:15:30.320 --> 00:15:32.069 align:start position:0% Once you start using closed loop prompts

00:15:32.079 --> 00:15:33.750 align:start position:0% in your codebase, your agents start

00:15:33.760 --> 00:15:35.670 align:start position:0% resolving their own work. You can create

00:15:35.680 --> 00:15:37.670 align:start position:0% specialized tasks for your front end,

00:15:37.680 --> 00:15:39.430 align:start position:0% for your back end, for whatever part of

00:15:39.440 --> 00:15:41.430 align:start position:0% your codebase, and it's all still on

00:15:41.440 --> 00:15:43.189 align:start position:0% this top layer. Now, we're not going to

00:15:43.199 --> 00:15:45.189 align:start position:0% dive into codebase organization too much

00:15:45.199 --> 00:15:46.710 align:start position:0% here. There are some leverage points of

00:15:46.720 --> 00:15:48.389 align:start position:0% agent coding that we're not going to

00:15:48.399 --> 00:15:50.150 align:start position:0% re-emphasize here. Here, we're focused

00:15:50.160 --> 00:15:51.910 align:start position:0% on giving our agents the highest

00:15:51.920 --> 00:15:54.389 align:start position:0% capability possible so that they are as

00:15:54.399 --> 00:15:56.710 align:start position:0% autonomous as possible. Now, there are

00:15:56.720 --> 00:15:58.310 align:start position:0% some trade-offs here. Again, things

00:15:58.320 --> 00:16:00.069 align:start position:0% start to get more complex. You need to

00:16:00.079 --> 00:16:02.069 align:start position:0% know how to prompt engineer. This is why

00:16:02.079 --> 00:16:04.310 align:start position:0% our agentic prompt engineering lesson

00:16:04.320 --> 00:16:06.310 align:start position:0% was so important. You must know how to

00:16:06.320 --> 00:16:07.990 align:start position:0% write these prompts properly.

00:16:08.000 --> 00:16:10.150 align:start position:0% Self-correcting agents is a massive

00:16:10.160 --> 00:16:12.310 align:start position:0% massive win for your agentic layer.

00:16:12.320 --> 00:16:13.749 align:start position:0% Let's move on to grade five. How does

00:16:13.759 --> 00:16:15.829 align:start position:0% our agentic layer continue progressing?

00:16:15.839 --> 00:16:18.870 align:start position:0% Now, by this point u many engineers have

00:16:18.880 --> 00:16:21.670 align:start position:0% many many prompts, many agents, several

00:16:21.680 --> 00:16:23.509 align:start position:0% skills. They've used many different

00:16:23.519 --> 00:16:26.069 align:start position:0% types of MCP servers and they start

00:16:26.079 --> 00:16:29.030 align:start position:0% scaling their work even further with a

00:16:29.040 --> 00:16:31.030 align:start position:0% key concept we discussed in lesson

00:16:31.040 --> 00:16:34.040 align:start position:0% three.

Source video

• URL: https://youtu.be/fop_yxV-mPo?si=ufh6xRFv874mU6Ud
• Video ID: fop_yxV-mPo
• Channel: IndyDevDan
• Title: The Codebase Singularity: “My agents run my codebase better than I can”

Local archive

• Stored centrally (deletable later):
• /Users/vishal/clawd/videos/20251229_IndyDevDan_The Codebase Singularity： “My agents run my codebase better than I can”_fop_yxV-mPo.webm

Notes

• Transcript and frame notes live alongside this skill under references/.

Visual notes (selected frames)

• Source video: https://youtu.be/fop_yxV-mPo?si=ufh6xRFv874mU6Ud
• Frame set: /Users/vishal/clawd/tmp_frames/fop_yxV-mPo_15s_selected/ (clustered selection)
• Sampling cadence: 15s (timestamps below are approximate)

selected_01_frame_0001.jpg (~00:00)

• What’s shown: Intro b-roll of hands on a laptop with stylized on-screen captions.
• Why it matters for the skill: The video positions this as a “framework” shift, not a single trick—skill should read like an operating model (repeatable process) rather than a one-off prompt.
• Actionable takeaway: Structure the skill around repeatable phases (prime → plan → execute → verify → summarize).

selected_02_frame_0002.jpg (~00:15)

• What’s shown: Similar intro b-roll on a laptop (no slide content).
• Why it matters for the skill: Mostly transitional; reinforces that the main value is the later concrete workflow artifacts.
• Actionable takeaway: Don’t overfit to intro; focus skill on the agentic layer artifacts and guardrails.

selected_03_frame_0004.jpg (~00:45)

• What’s shown: Caption text emphasizing “a sufficiently POWERFUL …” (setup for the agentic layer claim).
• Why it matters for the skill: Implies the core lever is adding capability/structure around the model (tools + workflows), not only choosing a smarter model.
• Actionable takeaway: In the skill, explicitly separate (a) model choice from (b) scaffolding: tools, workflows, validation.

selected_04_frame_0005.jpg (~01:00)

• What’s shown: A quote on a dark background: “I trust them to ship more than I trust myself or my team”.
• Why it matters for the skill: The standard to aim for is autonomy + reliability. Trust is earned via observability + guardrails.
• Actionable takeaway: Bake in “trust-building” steps: logging, explicit plan, small changes, and required verification commands.

selected_05_frame_0008.jpg (~01:45)

• What’s shown: Title card: “BUILDING AGENTIC LAYERS” (what it would take).
• Why it matters for the skill: Confirms this is about layered capability (progressive enhancement) rather than a single monolithic agent.
• Actionable takeaway: Design skill outputs as an incremental ladder (Grade 1 → Grade N) and encourage adding only what’s needed.

selected_06_frame_0013.jpg (~03:00)

• What’s shown: Scenic “person at desk” background (transition between sections).
• Why it matters for the skill: Mostly visual pacing; doesn’t add requirements.
• Actionable takeaway: None; treat as a section break.

selected_07_frame_0017.jpg (~04:00)

• What’s shown: Browser UI titled “Multi-Agent Orchestration” with an “Agent Observability” slide/section.
• Why it matters for the skill: Observability is key: you need a way to see plans, tool calls, state, progress, and costs.
• Actionable takeaway: Encourage a standard run format: print plan, record commands, track artifacts produced, and summarize what changed.

selected_08_frame_0018.jpg (~04:15)

• What’s shown: The same orchestration UI with a prompt input area; it includes a long “start … plan_build …” instruction.
• Why it matters for the skill: Demonstrates a pattern: structured workflow invocation with explicit outcome requirements.
• Actionable takeaway: Prefer “workflow prompts” (goal + constraints + deliverables) over conversational prompts.

selected_09_frame_0020.jpg (~04:45)

• What’s shown: Orchestration timeline/log view showing tool usage (e.g., bash), “pretooluse/posttooluse”, and a planner step; a right panel summarizes workflow phases.
• Why it matters for the skill: Shows closed-loop execution: plan → tool calls → review, with clear phase boundaries.
• Actionable takeaway: In SKILL.md, define phases explicitly and require the agent to state phase transitions and completion criteria.

selected_10_frame_0021.jpg (~05:00)

• What’s shown: Orchestration UI with multiple workflows running; icons/steps represent a repeatable pipeline.
• Why it matters for the skill: Reinforces using a consistent pipeline for different tasks rather than bespoke behavior each time.
• Actionable takeaway: Provide a reusable checklist that works for feature work, bugfixes, and refactors.

selected_11_frame_0028.jpg (~06:45)

• What’s shown: “Agentic Layer: Class 1 / Grade 1” over an example prime.md file (workflow includes git ls-files, read README.md, etc.).
• Why it matters for the skill: “Prime” is the entrypoint: teach the agent the repo shape and key docs before doing anything else.
• Actionable takeaway: Standardize a prime.md pattern: list files, read README(s), read key entrypoints, then summarize understanding.

selected_12_frame_0030.jpg (~07:15)

• What’s shown: “Example project structure” with folders like .claude/commands/prime.md, apps/…, plus AGENTS.md / CLAUDE.md / README.md, with labels suggesting “agentic layer” vs “app layer”.
• Why it matters for the skill: Shows the “agentic layer” as a first-class directory of reusable instructions (commands/skills/prompts), separate from product code.
• Actionable takeaway: For any repo, create/maintain an “agentic layer” folder (e.g., ./.claude/, ./agents/, or ./docs/agent/) that holds:
• priming instructions
• task workflows
• domain-specific notes

selected_13_frame_0031.jpg (~07:30)

• What’s shown: “Compute Advantage” slide for Grade 1: benefits include “Single prime prompt or/and memory file…”, “Minimal setup…”, “Foundation for incremental agentic growth…”, “Agents understand project context immediately”. Tradeoffs include “Useless for large codebases”, “Limited capability”, “Low agency”.
• Why it matters for the skill: Grade 1 is lightweight and fast but breaks down as complexity grows.
• Actionable takeaway: Document a progression: start with Grade 1 (prime + basic instructions), then add specialization and validation as complexity demands.

selected_14_frame_0037.jpg (~09:00)

• What’s shown: “Agentic Layer: Class 1 / Grade 2” with a test_writer.md role prompt (create pytest tests; read module, identify public methods, create tests, run pytest).
• Why it matters for the skill: Grade 2 introduces specialized roles/agents with explicit responsibilities and a built-in verification step.
• Actionable takeaway: Add role prompts like test_writer, docs_fetcher, reviewer that always include “how to verify” commands.

selected_15_frame_0038.jpg (~09:15)

• What’s shown: Project structure now includes agents/ with files like fetch_docs.md, test_writer.md, and an ai_docs/ folder with domain docs; still shows layered separation.
• Why it matters for the skill: Highlights maintaining curated domain documentation alongside the code for retrieval/grounding.
• Actionable takeaway: Encourage a references/ or ai_docs/ folder per skill/project: short, stable docs the agent can rely on.

selected_16_frame_0046.jpg (~11:15)

• What’s shown: “Agentic Layer: Class 1 / Grade 3” + “Compute Advantage” describing: MCP servers/skills/prime prompts with tools; skills encapsulate multi-step workflows with progressive disclosure; tools enable direct invocation; agents can interact with external/custom services. Tradeoffs mention config complexity and careful design.
• Why it matters for the skill: Grade 3 adds “tools as primitives” (skills/MCP) to expand capability beyond text.
• Actionable takeaway: Treat tools as part of the contract: declare required tools, expected side effects, and safety constraints.

selected_17_frame_0047.jpg (~11:30)

• What’s shown: A prime_db_w_tools.md (or similar) prompt for “User Management Database” with allowed tools and concrete psql connection commands + schema overview.
• Why it matters for the skill: Shows “tool-aware priming” for a subsystem (DB) including how to connect and what to inspect.
• Actionable takeaway: For complex systems, create subsystem-specific priming docs (DB, queue, auth, deployment) with exact commands.

selected_18_frame_0048.jpg (~11:45)

• What’s shown: “The Core Four” Venn diagram labeled CONTEXT, TOOLS, MODEL, PROMPT.
• Why it matters for the skill: Provides a clean mental model for designing reliable agents: you need all four.
• Actionable takeaway: Add a checklist in the skill: confirm Context is loaded, Tools are available, Model choice fits, Prompt specifies deliverables + constraints.

selected_19_frame_0063.jpg (~15:30)

• What’s shown: “Agentic Layer: Class 1 / Grade 4” + “Compute Advantage” emphasizing closed-loop prompts (Request → Validate → Resolve), self-correcting agents, specialized test commands, coordinated full-stack dev. Tradeoffs mention loops can get stuck, more prompt engineering, and need real engineering validation commands.
• Why it matters for the skill: Grade 4 is “trustable autonomy” via explicit validation and fix loops.
• Actionable takeaway: Require an explicit validation step and an exit condition (“stop if X fails twice, escalate with diagnostics”).

selected_20_frame_0065.jpg (~16:00)

• What’s shown: Another Grade 4 tradeoffs slide highlighting risk of getting stuck without proper exit conditions, increased prompt complexity, and the need for useful validation commands.
• Why it matters for the skill: Highlights the failure mode of agent loops: endless iteration without a real-world check.
• Actionable takeaway: In this skill, enforce:
• Max iteration count
• “Evidence of progress” requirement
• If validation fails, surface logs + hypotheses, then stop for human input