The Builder Journey

Estimated reading time: 5 minutes

Designing loops and skills is how you stop re-explaining intent every session and start compounding it.

In Knowledge as Code, I argued for the file: institutional memory that survives a tool change. In Knowledge Priming, I argued for the attention budget: what the agent reads first. In Reliable AI Begins Before the Model, I argued for the harness that makes that stack trustworthy in production. This essay is the build log underneath those claims. It is what it looks like when intent leaves the chat window and starts accruing interest.

The Re-Explanation Tax

It was 11:15 PM on a Tuesday, and I was staring at a terminal window, typing out the exact same forty-line paragraph for what felt like the hundredth time.

“Remember: do not invent metrics. Keep tone concise, professional, and peer-level. Use the standard Knowledge Stack headers. Output valid Markdown without raw HTML tags. If a source is weak, flag it explicitly…”

That is the Re-Explanation Tax: the invisible levy every builder pays when operating purely inside the chat context window. The model was brilliant. It was also an amnesiac. Every session felt like briefing a gifted contractor with total anterograde amnesia. Fifteen minutes setting guardrails. Five minutes of actual work. Close the tab.

Open the next tab, and the meter resets. Your compound interest goes to zero.

I stopped mid-sentence, deleted the block, and opened VSCode.

Instead of pasting those guardrails into another prompt, I created a clean file at .claude/skills/drafting-standards/SKILL.md. I stripped the informal prose and structured the intent directly: YAML frontmatter, explicit anti-patterns to reject, and schema rules for verification. I saved it to disk.

Ten seconds later I opened a brand-new session and typed six words: Run drafting-standards gate on packet-04.md.

The agent loaded the skill off disk, validated the evidence schema, flagged a soft inference claim without me asking, and produced a clean draft on the first turn.

That was the shift. I was no longer prompting an agent. I was building an asset that prompted the agent for me. The memory no longer belonged to the session. It belonged to the system.

Ephemeral Prompting vs Durable Intent

Call the old posture Ephemeral Prompting. Context lives in the window. When the chat ends, or the token limit bites, the context dies. You pay the Re-Explanation Tax again tomorrow.

Call the new posture Durable Intent. Patterns, guardrails, and domain conventions live in code, in .md skills, and in disk-backed state the next session can load without a second briefing.

Addy Osmani names the same move in Loop Engineering: stop being the person who prompts the agent; design the system that prompts the agent. Skills write down project knowledge the agent would otherwise guess. Memory lives outside the conversation, because the model forgets between runs and the repository does not. Market signal is catching up too. Collections like ai-design-skills treat SKILL.md plus frontmatter as the unit of behavior knowledge, not a clever prompt you hope you can find again.

Skills and loop harnesses are not technical souvenirs. They are the compound interest of AI workflow. Without them, every session resets the rate to zero, no matter how eloquent last night’s prompt was.

Where the Shift Lives in The Knowledge Stack

This is not a standalone software tutorial. It sits inside the same stack I have been building in public:

Layer	What lives here in the Builder arc	Builder’s mindset shift
System	Loops, connectors, automated verification, review gates	Stop prompting the model; design the system that prompts the model
Semantics	SKILL.md collections, frontmatter, domain conventions	Capture Durable Intent outside the session so the agent stops guessing

Knowledge Priming still owns what loads first. Reliable AI Begins Before the Model still owns whether a harness exists. The Builder Journey is how you construct the compounding layer between those two: Semantics that carry intent, System rituals that invoke and verify it.

A five-minute check keeps the map honest. If you are about to paste the same paragraph into a third session this week, that paragraph wants to be a skill. If a skill exists and you are still retyping it, you do not have a skill problem. You have an invocation problem: the System layer never learned to call what Semantics already knows.

Build the Asset That Prompts the Agent

The portable lesson is not “use Cursor” or “copy my folder tree.” It is this: treat intent as an asset with an owner, a path, and a review trail.

Write the rule once, in a form an agent can load every run. Name the anti-patterns as carefully as the desired behavior. Put verification next to generation so the maker is not grading its own homework. Let the next session start from the file, not from your memory of what you typed at 11:15 PM.

That is how Integration Era work compounds at personal and team scale. Production does not improve because you prompt harder. It improves because less of what you know has to be re-derived from zero every morning.

Stay the Engineer

Loop engineering has a failure mode that looks like progress. An unattended loop can ship impressive volume while quietly introducing drift, soft claims, or bad assumptions, especially if nobody owns the verification gate. That is comprehension rot with a productivity dashboard.

Osmani’s close is the right counterweight: build the loop, but stay the engineer. I would sharpen the role change. You are not stepping away from the wheel. You are moving from manual operator, typing raw prompts, to harness designer and chief reviewer, verifying output at System gates. The builder does not surrender agency. They elevate it.

Durable Intent without review is just amnesia with better packaging. The skill on disk only compounds if someone still owns what “done” means.

What intent are you still pasting into the window that already belongs on disk?

Madam I’m Adam

This continues the thread from Knowledge as Code, Knowledge Priming, and Reliable AI Begins Before the Model: the file, the attention budget, and the harness only compound when loops and skills put Durable Intent on disk.