Builder's Field Notes: 28 Moments from Inside the IDE
I use the Logic API every day. Not to demo it. To build with it.
These field notes are 28 screenshots from real work sessions captured over weeks of building Ejentum: backend infrastructure, benchmark design, website copy, security auditing, blog writing, ability dataset purification, and research. Different tasks. Different days. Same tool.
Three moments that define the guide
The agent checking its own bias. I'm writing about benchmark results and the Haki scaffold forces Claude to ask: "Am I presenting the correctness flat result as 'not a problem' because I want the product to look good?" The blog post that came out of that session includes the correctness dip front and center. That's the scaffold working.
The wrong domain that still worked. I asked for an ability to help with naming inconsistency. The API returned failure mode classification. Wrong domain. But Claude read the scaffold and said: "Not a perfect match, but the core principle is potent." The agent transferred the suppression signal across domains. This is the 62% finding in action.
17 rounds, 101 fixes. The security ritual escalated from a single Ki to a Haki with four mathematical paradigm shifts (set theory, state machines, cryptography, game theory). The compound topology found an SSRF that individual rounds missed. The longitudinal Ki caught a fix-propagation bug across round 12 and round 17.
Patterns that emerged
Ki for execution, Haki for planning. More abilities doesn't mean better. It means different.
Suppression matters more than domain match. A compression ability worked for editorial precision. Suppression is domain-agnostic.
The scaffold persists. Half-life of 24 steps. 17 security rounds later, the agent still referenced abilities injected at the start.
Read the full guide
Read the full guide with all 28 screenshots on GitHub
100 free calls. No card required. Get started.