Repository navigation
by the numbers
Data collected on 2026-09-23, from the standalone repository root. Counts exclude node_modules, dist-electron, and Prisma generated output.
The capsule is a single TypeScript project: 108 TypeScript source files (plus 6 .mjs scripts and 1 .js), 16 of which are test files. TypeScript accounts for roughly 23,800 of the tracked source lines.
xychart-beta horizontal
title "Lines of code by language"
x-axis ["TypeScript", "YAML", "HTML", "CSS", "Prisma", "SQL", "PowerShell", "Markdown", "JSON", "MP3 (binary)"]
y-axis "lines" 0 --> 24000
bar [23816, 7999, 444, 328, 296, 258, 204, 255, 195, 967]
TypeScript dominates at ~23,800 lines. The YAML figure is largely the pnpm-lock.yaml lockfile. There is one workspace: pnpm-workspace.yaml declares zero packages, so the capsule itself is the sole build unit.
112 commits total. The first commit landed 2026-06-28. The project is young and the pace is uneven:
| Month | Commits |
|---|---|
| 2026-06 | 1 |
| 2026-07 | 12 |
| 2026-08 | 2 |
| 2026-09 | 97 |
September dominates by a wide margin, driven by the Coding Brief feature, the terminal-transcript shell, and the repository migration and security-hardening work.
Git author attribution splits into one human and several agent/bot lanes:
| Author | Commits |
|---|---|
| Codex | 42 |
| Gaffer.dev | 35 |
| drferdii | 10 |
| drferdi | 7 |
| dependabot[bot] | 7 |
| Claude | 4 |
| github-actions[bot] | 4 |
| ferdiiskandar | 2 |
| dr Ferdi Iskandar | 1 |
The Codex, Gaffer.dev, and Claude lanes are agent attributions, and the dependabot[bot] / github-actions[bot] rows are bots. Treating those as non-human gives an AI/bot-assisted share of roughly 78% by commit count. This is a lower bound: inline assistants that do not touch git author metadata leave no trace.
The clearest signal is file size. The desktop renderer is an outlier:
| File | Size |
|---|---|
desktop/renderer/renderer.ts |
98.6 KB |
data/templates/business.ts |
62.5 KB |
data/templates/coding.ts |
51.7 KB |
data/templates/creative.ts |
42.7 KB |
data/templates/marketing.ts |
42.3 KB |
lib/llm/prompt-builder.ts |
21.9 KB |
desktop/main.ts |
18.8 KB |
desktop/ipc/core.ts |
16.8 KB |
types/index.ts |
16.4 KB |
lib/optimizer/engine.ts |
15.6 KB |
The renderer at ~98.6 KB (about 3,000 lines) concentrates the terminal console in one module: command parsing, rendering, streaming, and the clarification round. The template data files are large because they hold long prompt templates, not logic. See Complexity hotspots.
Open-source prompt engineering and multi-LLM tooling by Sentra Artificial Intelligence.
MyPrompt develops practical approaches to prompt engineering, multi-LLM optimisation, reusable prompt systems, and AI-native workflows — with an emphasis on structured, interoperable, and real-world AI use.
Built in Indonesia as part of the Sentra Artificial Intelligence ecosystem.
Sentra Artificial Intelligence · Source Repository · Official Website
Dr Ferdi Iskandar — Creator & Maintainer
LinkedIn ·
ORCID ·
Hugging Face ·
Kaggle ·
Medium ·
Substack ·
X ·
Threads
MyPrompt · Sentra Artificial Intelligence · Indonesia