T27.AI

Blog

Vibecoding course: lesson texts expanded to fill the chat grounding budget

2026-11-14 · 2 min read

[lengths are from course-from-spec --check on eb7d0c2a7; API probes done in-browser] All 27 English lesson texts in the Vibecoding from Zero course were expanded from 361–685 bytes to 1170–1403 bytes each to fill the 1500-byte CHAT_BODY_MAX slot that course-chat.t27 injects into the system prompt. Nine Russian texts were updated for number parity, and eight superlative phrases were rephrased.

Title card, no illustration yet, for: Vibecoding course: lesson texts expanded to fill the chat grounding budget
View the title card at full size
#t27#Course#Vibecoding

Course 12 of t27.ai, Vibecoding from zero, now carries a 1170–1403 byte English grounding text in every lesson. Before PR #1876 the texts were 361–685 bytes, leaving the 1500-byte CHAT_BODY_MAX budget mostly empty and the course chat helper poorly anchored to lesson content.

How the grounding budget works

The course chat helper is defined in apps/website/specs/course/course-chat.t27. At line 325 its system_prompt() function injects the lesson body text unconditionally when its length is greater than zero: every lesson now fills roughly 78–94 % of that 1500-byte CHAT_BODY_MAX slot. The injected text covers the key facts the lesson teaches, the widget data shown on screen, and common misconceptions the helper should address.

Number parity and superlative rules

The course-from-spec --check validator enforces two constraints on lesson text pairs. The number parity rule requires that every digit sequence present in the English text also appears in the corresponding Russian text in vibecoding.ru.json. Expanding 27 English texts added new digit sequences in 9 lessons; those 9 Russian texts were updated with natural-language additions that carry the same digits. The superlative rule flags the phrases "the first" and "the only". Eight English texts used those phrases; all eight were rephrased to remove the flag without introducing new digit sequences. After both fixes, course-from-spec --check runs clean: 324 lesson specs, 93 spec tests, 438 asserts, 0 parity errors, 0 superlative flags.

Verification

Code audit confirms that system_prompt() injects lesson_body for all 27 lessons: the injection is unconditional when len > 0, and all 27 expanded texts have len > 0. Live API probes confirmed correct lesson-grounded responses for all 27 lessons, each reply citing data unique to that lesson's widget.

LESSON_TEXT bytes vs CHAT_BODY_MAX = 1500Before361-685 BAfter1170-1403 B
Each lesson text before PR 1876 (361-685 B) versus after (1170-1403 B), against the CHAT_BODY_MAX = 1500 B budget injected by course-chat.t27 at line 325.

What this does not settle

Receipts

Work with me

Need an FPGA/RTL problem taken to measured hardware?

I work contract and part-time on hardware-AI, FPGA/RTL and ML systems — from specification and open toolchains to reproducible measurements.