Prompt Architecture (Top-Down)

This is the runtime order used in current code.

Top-Down Flow Diagram

PromptSpecWithSkills(...)
        |
        v
ApplyPersonaIdentity(...)
        |
        v
AppendLocalToolNotesBlock(...)
        |
        v
AppendPlanCreateGuidanceBlock(...)
        |
        v
AppendTodoWorkflowBlock(...)
        |
        v
Channel PromptAugment(...)
        |
        v
AppendMemorySummariesBlock(...)
        |
        v
BuildSystemPrompt(...) -> system prompt text

1) Identity Layer

Base identity comes from agent.DefaultPromptSpec().

Then runtime may replace it via local persona files:

• file_state_dir/IDENTITY.md
• file_state_dir/SOUL.md

Applied by promptprofile.ApplyPersonaIdentity(...).

2) Skills Layer

skillsutil.PromptSpecWithSkills(...) discovers selected skills and injects only metadata into the system prompt:

• skill name
• SKILL.md path
• short description
• required auth_profiles (if provided)

The model reads the actual skill file through read_file when needed.

3) Core Policy Blocks

Common blocks are appended in task runtime order:

1. local tool notes (SCRIPTS.md) via AppendLocalToolNotesBlock
2. plan_create guidance via AppendPlanCreateGuidanceBlock
3. TODO workflow policy via AppendTodoWorkflowBlock

4) Channel Blocks

Channel runtime then appends channel-specific blocks.

Examples:

• Telegram: AppendTelegramRuntimeBlocks(...)
• Slack: AppendSlackRuntimeBlocks(...)
• LINE: AppendLineRuntimeBlocks(...)
• Lark: AppendLarkRuntimeBlocks(...)

5) Memory Injection Block

If enabled, memory snapshot text is appended as another block through AppendMemorySummariesBlock(...).

6) System Prompt Rendering

agent.BuildSystemPrompt(...) renders agent/prompts/system.md with:

• identity
• skills metadata
• appended blocks
• tool summary from registry
• rules list

7) Message Stack Sent to LLM

In engine.Run(...), message order is:

[system] rendered system prompt
   ->
[user] mister_morph_meta (optional)
   ->
[history] non-system messages
   ->
[user] current message or raw task

1. system prompt
2. injected runtime metadata message (mister_morph_meta, if present)
3. history messages (non-system)
4. current message (if provided) or raw task text

Practical Rule

If you need to customize prompt behavior, prefer these extension points in order:

1. IDENTITY.md / SOUL.md
2. SKILL.md + skills.load
3. SCRIPTS.md
4. runtime-specific prompt augment
5. low-level agent.WithPromptBuilder(...) (only for full custom wiring)