🛠️ Technique II — OS Capability
The action layers: once the foundation exists (Track 2), you verbs and arms to the OS. Skills, Tools/Connections, and Agents—in that order, without skipping steps.
How to read: the foundation (on the left) supports the three action layers (center, in cyan)—Skills, Tools, and Agents—which together become an OS that acts (on the right). This path builds these three in the right order: skills first, agents last.
Learning path map
Detailed content
🛠️ Skills — the Verbs You’ve Earned
The OS’s unit of capability. Start by hand, distill with reverse meta-prompting, iterate forever—and know when a skill should become a script.
Do the task manually first — sometimes spending an hour typing out all your reasoning — and only then capture the steps in a skill.
A skill for something you’ve never done by hand comes out generic and breaks. Doing it manually first ensures the recipe reflects reality.
Manual first · context priming · capture the process.
Ask the AI: "based on this conversation, distill the exact criteria and turn them into a slash command I can always reference".
It’s writing down the recipe after cooking: never redo the same one-hour briefing again.
Distill the conversation · slash command · recipe after the meal.
You release a V1 and keep refining it. The author has a skill that’s been fine-tuned almost every day for six months.
Treating a skill as "ready" lets it decay. The cycle is alive: use, observe, adjust.
Ship a V1 · continuous iteration · ∞.
At the end of the session, you ask "how should we adjust the skill?" and provide a rubric for when to change something (or not).
The rubric prevents impulsive adjustments: the skill changes only when the criteria justify it.
Post-mortem · rubric · change with criteria.
If a skill runs 100% autonomously (fixed input→output), it may be better as a deterministic script/cron job written by the AI.
A script uses fewer tokens and less context, and doesn’t need to be loaded into the window every session.
100% autonomous · cron · context savings.
People underuse it, accumulate it, or think they’re clever and create too many skills that go stale—and skills for things they’ve never done by hand.
Starting with ONE skill (the one you repeat most often) keeps your folder from filling up with dead recipes.
One skill first · no hoarding · no ghost skills.
A ready-to-use prompt that asks Claude Code to distill the current conversation into a /slash command that accepts an argument.
It’s the practical exercise: you leave the module with your first skill brought to life.
Copy-run · argument · SOP in sequence.
The parts of a SKILL.md: name, when to use it (trigger), the step-by-step procedure (SOP), and how to verify it.
Knowing the anatomy lets you read, debug, and improve any skill—yours or someone else’s.
Name · trigger · SOP · done-check.
🔌 Tools & Connections — the wires to the outside
Skill, CLI, API, or MCP alone? How to wrap an API in your own CLI — and the security hack that changes everything: a read-only CLI.
Before connecting any tool, audit it: does it offer a skill, CLI, API, or only MCP? The answer determines the path.
Choosing wrong costs maintenance and context. The question filters first, before you spend effort.
Audit the tool · 4 connectors · make an informed choice.
Over time, MCP tends to become less relevant. It’s only worthwhile when the tool offers MCP but not an API.
You stop installing MCPs by reflex and prefer a skill or CLI when there’s an API.
MCP in decline · MCP-only is the exception · prefer API.
You can ask Claude Code: "turn this API into a custom CLI with these main functions".
When the work involves a lot of command-line use, your own CLI is leaner than calling the raw API.
API wrapper · selected functions · lean remote control.
Design the CLI so it physically has no write access—it only reads data, never writes to the database.
Eliminates the risk of a POST accidentally destroy data when you're deep into the context window.
Read-only · no write access · physical tripwire.
Each connection or CLI you add is another piece to maintain. Add with intention, not by reflex.
Too much infrastructure breaks more than it helps. Sometimes a ready-made skill does the job without you building anything.
Maintenance cost · intentionality · less is more.
Supabase CLI (Health OS database), Obsidian CLI (second brain), and the Appify skill (scraping) — ready to use, with no CLI of its own.
Show when to build your own infrastructure and when a third-party skill already does the job.
CLI when needed · skill when enough · combine them.
A ready-to-use prompt that asks to turn an API into a CLI read-only, with the functions you list.
It’s the practical exercise: you leave with a safe CLI for your favorite tool.
Copy-run · read-only · explicit functions.
If a ready-made skill (e.g., Appify) already does the job, don’t break your back maintaining your own CLI just out of pride.
Recognizing when to stop is just as valuable as knowing how to build. Every extra piece is debt.
A ready-made skill is enough · don’t reinvent · infrastructure debt.
🤖 Agents — Roles with Judgment
The last layer, not the first. A skill is a verb; an agent is the chef who chooses skills, in order, with a review gate before anything goes out.
Skill = one repeatable verb. Agent = a specialist worker that orchestrates multiple skills with judgment.
Confusing the two leads you to create an agent when a skill would have been enough — expensive and fragile.
Verb vs. worker · judgment · orchestrates skills.
Every agent that produces something going outside passes through a review gate: a point where you (or another agent) checks it before sending.
It’s what separates a reliable agent from one that sends nonsense to a client/accountant.
Review gate · human in the loop · check before sending.
Only promote a routine to an agent if you already perform it manually today, with existing skills for it to orchestrate.
An agent without underlying skills or a real routine becomes theater: it looks powerful but doesn’t deliver.
Real routine · skills first · prerequisite.
A one-off task (auditing 100s of files) might call for a dynamic workflow, not a fixed agent. Need more power? Spin up disposable sub-agents.
Not every task deserves a permanent "employee." Choosing the right form saves.
Fixed agent · workflow · throwaway sub-agents.
In Freedom OS: a response agent (scans Gmail and drafts replies using the right documents) and a deep research agent (monitors changes to passport laws).
Real cases show agents as “employees” in a domain — each with a clear job.
Response agent · deep research · living OS.
A skeptical reviewer agent (devil's advocate) that "chews you out before the accountant/client does": checks the work before it goes out.
Reduces back-and-forth: your own OS roasts you before the mistake reaches anyone who matters.
Skeptical reviewer · devil’s advocate · adversarial gate.
As models get smarter, the same agent needs less instruction — or the agent is no longer needed because the harness already does it natively.
Makes you revisit agents periodically instead of accumulating them forever.
Less instruction · the agent disappears · periodic review.
A copyable skeleton of AGENT.md: role, the skills it orchestrates, order, review gate, and limits.
It’s the practical exercise: you leave with your first agent sketched out and ready to fill in.
Copy-run · role + skills + gate · limits.