A note about this build log entry: I published this after the week it covers. DevNitro reconstructed a draft from work-related entries in my journal, Git commit history, and time tracking. I reviewed, edited, and approved the final post. The story reflects what I knew and built at the time.
From the Backlog to the Build Log
Week of May 29 - June 4, 2026
What I Worked On
This was a lighter, consolidation week. A lot of the hours went to testing and content, and the standout new capability was small in code but big in effectiveness: a built-in Planning skill. Making the agent plan before it builds is a quiet but central move in Intelligent Engineering.
The idea is that the agent should not always go straight from a request to writing code. It should plan first. So I built planning in as a first-class skill: the agent turns a feature request into a structured plan, and only then starts building. I gave the plan agent its own memory with a defined knowledge scope so it works from the right context, a secrets scan so a plan can never accidentally capture a credential, and telemetry so I can see how planning actually performs. Then I made the blueprints rely on the built-in planning skill instead of each carrying its own duplicated planning docs. One source of truth, inherited everywhere.
Why bother? Because the gap between a vibe-coding assistant and something a team can trust is mostly the plan. An agent that plans, shows you the plan, and then executes is one you can supervise. An agent that just starts typing is one you have to babysit. This is the same instinct behind turning the backlog into a plan, pushed one level earlier into how a single change gets made.
What the AI Did Well
The agent handled the supporting tool work cleanly: an install_playwright infrastructure tool with browser pre-checks so visual verification sets itself up, a requirement that the app-start tool be given an explicit project path instead of guessing, and blockquote support in the markdown formatter. It also refactored the payment configuration to use provider-specific keys, which is the kind of tidy, correctness-focused change I am happy to delegate.
What Needed Correction
A familiar gremlin came back: a pipe deadlock, this time fixed for good by draining a process's standard output in the background so a full buffer can never freeze the whole thing. I had patched a version of this before, and this week I made it a rule rather than a one-off.
I also cleaned up the rough edges around identity and startup: standardized the authentication messages behind constants so they are consistent, switched to an interactive CLI login prompt so signing in is obvious, and tightened project-scoped session reuse so a conversation reconnects to the right project first. And I added edit and delete UI for personas, because a governance feature you cannot manage from the screen is only half a feature.
What Shipped
DevNitro has a built-in Planning skill with a scoped plan-agent memory, a secrets scan, and telemetry, and the blueprints now use it directly. There is an install_playwright infra tool with browser pre-checks, a safer app-start tool, a real fix for git pipe deadlocks, blockquote markdown, standardized auth messages, an interactive login, project-scoped session reuse, provider-specific payment config, and persona edit and delete UI.
The Numbers
| Category | Hours |
|---|---|
| Product / Features | 11h 36m |
| Content | 10h 39m |
| Testing | 9h 11m |
| Business | 2h 36m |
| Community | 1h 8m |
| Deploy / Infra | 0h 31m |
| Total | 35h 41m |
Thirty-five hours, split fairly evenly across product, content, and testing. Not a headline week, but that is ok. Nine hours of testing on a light week is the tax you pay to keep moving fast without breaking what already works.
Where This Fits in Intelligent Engineering
A vibe-coding agent goes straight from a request to creating code, and then you babysit it. Intelligent Engineering puts a plan in between. The agent turns a request into a structured plan, shows it to you, and only then builds, so you supervise the change instead of watching every keystroke. A plan you can see is the smallest unit of trust. I also made the blueprints rely on the one built-in planning skill instead of each duplicating its own, so the practice is inherited everywhere, not reinvented per project. Plan first, then build, is easy for a person to say and easy for an agent to skip. That is exactly why it is now built in.
A plan you can see is the smallest unit of trust. This week I made the agent produce one every time.
Curious how this site was built?
Watch the 5-minute walkthrough on YouTube →
0 Comments
No comments yet. Be the first to share your thoughts!
Leave a comment