Repository navigation
overview glossary
Super Prompt — The Optimizer's default output: a six-section markdown prompt (ROLE, TASK, CONTEXT, APPROACH, CONSTRAINTS, OUTPUT FORMAT). Parsed by lib/optimizer/super-prompt-format.ts.
Coding Brief — A coding-agent handoff with eight fixed sections (GOAL, CONTEXT, SCOPE, STACK, OUT OF SCOPE, DONE WHEN, ASSUMPTIONS, REPORT), produced for CODING tasks and validated against V1–V14. Standard lives in docs/CODING_BRIEF_STANDARD.md.
Transform — The deterministic, no-network prompt scaffold. Detects intent, applies a persona, and wraps the input in a compiler profile. See lib/transform/engine.ts.
Optimizer — The LLM-backed prompt refiner that produces Super Prompts and Coding Briefs. See lib/optimizer/engine.ts.
Evaluator — The LLM-as-a-judge that scores a prompt across Structure, Clarity, Completeness, and Specificity. See lib/evaluator/engine.ts.
Compiler profile — A target-LLM-specific prompt skeleton. The Transform tool offers claude, codex, gemini, and grok (plus default). Implemented under lib/transform/compiler/.
Optimizer lane — INTERACTIVE (fast, low max-tokens) or DEEP (retrieval-aware, higher max-tokens and temperature). Controls the system prompt and token budget in lib/optimizer/engine.ts.
Effort level — A Transform-only setting (low/medium/high/xhigh/max) that maps to a max-token budget in the renderer.
Clarification round — After a delivered Coding Brief, the engine may offer up to three questions about CONTEXT, DONE WHEN, SCOPE, or ASSUMPTIONS lines. Defined by lib/prompt-quality/clarification.ts.
Thin brief — A valid but incomplete Coding Brief (V11) whose CONTEXT and DONE WHEN both defer to the user, so it is never marked complete.
Capsule — A standalone project directory that installs, builds, tests, runs, and deploys independently of the Monorepo. MyPrompt is the internal/prompt capsule.
SAFRS — "Human-Governed · Agent-Executed · Machine-Enforced", the repository operating model. Risk classes R0 (read-only) through R3 (high-impact) gate how work is executed.
R2 — The risk class recorded for MyPrompt: boundary-affecting changes (desktop IPC, provider keys, user data, payments, migrations).
CTE — "Compiler & Transformation Engine", the product's historical internal name (visible in lib/constants.ts and email branding).
Provider readiness — The resolved active LLM provider (and model chip) for a desktop session, computed from available env keys. See lib/llm/provider-readiness.ts.
User API key — A per-user, per-provider key stored AES-256-GCM encrypted with iv and authTag in the user_api_keys table.
Guest mode — The desktop path when no Supabase session exists. Transform and Optimizer run against environment-provided keys without a database user.
ModelId vs LLMProviderName — The Transform feature uses model IDs like claude-sonnet (lib/transform/schemas.ts), while the provider registry uses names like CLAUDE (types/index.ts). They are different namespaces and must not be mixed.
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