AgentIME

Inspiration

AI agents are powerful, but writing precise instructions from scratch is slow and inconsistent. AgentIME treats instructions like an input method: instead of typing a long prompt word by word, users select reusable directive modules and compose a clear instruction quickly.

What it does

AgentIME is a bilingual command input method for AI agents. It helps users build structured instructions through a tree of directive modules such as planning, execution, verification, recovery, and completion.

The app compiles selected modules into a deterministic, readable prompt. It works without connecting to an AI model or reading conversation content: users can copy the result or paste it into any text field.

How we built it

We built AgentIME as a native Swift application for macOS, with a reusable directive core shared by the interface. The core includes:

A local bilingual directive pack
Tree-based candidate selection
Deterministic recommendation rules
Conflict detection between incompatible instructions
A prompt compiler that orders modules by semantic role
A macOS menu-bar workflow with global shortcut support

The project is designed to extend to an iOS keyboard extension while keeping the instruction library local and privacy-preserving.

Challenges we ran into

The main challenge was avoiding a simple prompt-snippet library. We needed instructions to remain composable, consistent, and predictable. We solved this with typed directive slots, conflict rules, and a deterministic compiler, so the same selection always produces the same result.

We also designed the interaction for speed: a user should be able to move from intent to a usable agent instruction in a few selections rather than writing a large prompt manually.

Accomplishments that we're proud of

Built a working bilingual AI-agent instruction composer
Created a structured local directive database rather than static prompt templates
Implemented deterministic compilation and conflict handling
Produced a native macOS release and demonstration video
Kept the initial product private by design: no cloud account, API key, or conversation reading is required

What we learned

Prompting becomes much more approachable when it is represented as selectable cognitive actions rather than a blank text box. We also learned that predictability matters: users should be able to understand exactly what instruction will be sent.

What's next for AgentIME

Next, we plan to add editable directive packs, personal recipes, local usage-based recommendations, and an iOS keyboard version that inserts compiled instructions directly into supported text fields.