How the Prompt-Settings System Customizes AI Instructions in ThePrimeagen's 99

The prompt-settings system in 99 centralizes AI instruction customization through a Lua-based pipeline that wraps user prompts with XML-like tags, injects contextual metadata, and allows extensible prompt fragments for different operations.

ThePrimeagen's 99 repository implements a sophisticated prompt-settings architecture that controls how every LLM request is constructed and contextualized. Understanding how this prompt-settings system customizes AI instructions reveals the mechanism behind the plugin's ability to generate tutorials, perform semantic searches, and modify code with precise contextual awareness.

The Prompt-Settings Pipeline Architecture

The system operates as a five-stage pipeline where raw user instructions are transformed into fully contextualized AI prompts. Each stage is implemented in specific Lua modules within the lua/99/ directory, ensuring that every LLM request follows a consistent structure while remaining customizable.

Step 1: Prompt Definitions in prompt-settings.lua

At the core lies lua/99/prompt-settings.lua, which exports a prompts table containing static fragments and dynamic formatting functions.

Static fragments include:

• prompts.role – Defines the AI as a "software-engineering assistant"
• prompts.tutorial() – Returns instructions for creating tutorials
• prompts.semantic_search() – Specifies strict output formats for search results

Dynamic helpers include:

• prompts.prompt(user_prompt, action, name?) – Wraps prompts in XML-like tags
• prompts.visual_selection() – Formats selected code blocks
• prompts.output_file() and prompts.read_tmp() – Handle file system references

Step 2: Prompt Composition with make-prompt.lua

When an operation executes, lua/99/ops/make-prompt.lua orchestrates the composition. The make_prompt() function calls context._99.prompts.prompts.prompt(), which constructs a wrapped prompt structure:

<NAME>user prompt</NAME>
<Context>operation-specific prompt</Context>

This XML-like envelope separates user intent from operational context, allowing the LLM to distinguish between direct instructions and background requirements.

Step 3: Context Metadata Injection

Before transmission, RequestContext:finalize() in lua/99/request-context.lua appends three critical metadata blocks:

• <Location> – The file path and position where changes apply
• <FunctionText> – The selected code range or function text
• <MustObey>...<TEMP_FILE> – Constraints telling the model exactly where it may write changes

This injection ensures the AI understands both the spatial context (where am I?) and the file system boundaries (what can I touch?).

Step 4: Operation-Specific Prompts

Each operation in lua/99/ops/ imports specific prompt fragments:

• tutorial.lua uses prompts.tutorial() to request educational content generation
• search.lua uses prompts.semantic_search() to enforce strict location-string formatting for results

These specialized prompts extend the base system without modifying the core pipeline.

Step 5: Extending the System

The architecture supports custom prompts through simple table extension. To add a new prompt:

-- In lua/99/prompt-settings.lua
prompts.my_special = function()
  return [[
    You are asked to generate a summary of the given file.
    <Rule>The output must be a single paragraph.</Rule>
  ]]
end

Then reference it in any operation:

local special_prompt = ctx._99.prompts.prompts.my_special()
local full_prompt = ctx._99.prompts.prompts.prompt(
  "Summarize the current buffer", special_prompt, "SUMMARY"
)

No other code changes are required, making the system fully extensible.

Practical Code Examples

Using a Built-in Prompt

local ctx = _99.State:from_current_buffer(1)
ctx:finalize()

local prompt = ctx._99.prompts.prompts.tutorial()
print(prompt)  -- Returns the full "create a tutorial" prompt string

Composing a Full Request

local make_prompt = require("99.ops.make-prompt")

local full_prompt, refs = make_prompt(
  ctx,
  ctx._99.prompts.prompts.semantic_search(),
  { additional_prompt = "Find all usages of `foo`", additional_rules = {"my_rule"} }
)

Viewing the Final Meta-Prompt

ctx:finalize()
local meta = table.concat(ctx.ai_context, "\n")
print(meta)  -- Complete prompt sent to the model including Location, FunctionText, etc.

Summary

• The prompt-settings system centralizes all AI instruction customization in lua/99/prompt-settings.lua
• XML-like tags (<NAME>, <Context>, <Location>, <FunctionText>, <MustObey>) structure the prompt into semantic sections
• Context metadata is injected by RequestContext:finalize() in lua/99/request-context.lua, providing spatial awareness and file-system constraints
• Operation-specific prompts (tutorial, semantic_search) extend the base system without pipeline modifications
• Extensibility is achieved by adding new functions to the prompts table, requiring no changes to core composition logic

Frequently Asked Questions

How does the prompt-settings system wrap user instructions?

The system uses the prompts.prompt() function in lua/99/prompt-settings.lua to wrap user instructions in XML-like tags. It creates a structure with <NAME> tags containing the user prompt and <Context> tags containing the operation-specific instructions, allowing the LLM to distinguish between direct user intent and background context.

What metadata does RequestContext:finalize() add to AI prompts?

RequestContext:finalize() in lua/99/request-context.lua appends three critical metadata blocks: <Location> specifying the file path and cursor position, <FunctionText> containing the selected code range or function, and <MustObey>...<TEMP_FILE> defining the temporary file where the model must write changes. This ensures the AI understands both spatial context and file system constraints.

Can I add custom prompts without modifying core files?

Yes, you can extend the system by adding new functions to the prompts table in lua/99/prompt-settings.lua. Simply define a new function like prompts.my_custom = function() return [[your instructions]] end, then reference it in any operation via ctx._99.prompts.prompts.my_custom(). The composition pipeline in make-prompt.lua automatically handles the new prompt without requiring changes to core logic.

Where are operation-specific prompts like tutorial and semantic_search defined?

Operation-specific prompts are defined as functions in the prompts table within lua/99/prompt-settings.lua. For example, prompts.tutorial() returns instructions for creating educational content, while prompts.semantic_search() specifies strict formatting requirements for search results. These are then imported by their respective operation files in lua/99/ops/ such as tutorial.lua and search.lua.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}