High Signal Prompting

A compact prompt method where each single-word header defines one of four sections with a specific meaning for the model to interpret. This optimizes model reasoning to both drastically reduce token usage per prompt and constrain the model to execute tasks from a place of high specificity.

The goal is maximizing signal density per token which results in:

• Optimized token consumption / reduced agentic costs
• Explicit interpretation / reduction of implicit inference guess
• Output consistency / mistake reduction and overstepping scope
• Maximize window reasoning retention / reduced drift and entropy
• Increased prompt response speed / overall increased workflow

Section Types

Retrieval – Input Grounding

Represents the tangible data and area of responsibility by which the model is to target.

Framing – Interpretation Layer

A set of instructions that shapes how the model is meant to reason about the prompt.

Refinement – Output Control (optional)

Applies special instructions to the model that don’t fit in the other sections.

Execution – Operator Selection

Defines the actions for the model to perform based on the prompt.

Section Ordering

1. Retrieval
2. Framing and Refinement
3. Execution

Contract

• Ideally, use one Retrieval, Framing, and Execution section at minimum
• Additional Framing sections further specify interpretation
• Any number of Retrieval, Framing, or Refinement sections can be used
• Recommend to use only one Execution section per prompt

---

Syntax Structure

Sectioning

key: value structure (as an unordered list)

Key:
- Value one
- Value two

Unordered List

Using - dashes for unordered sequencing

Task:
- Complete task value XYZ
- Complete task value ABC

Ordered List

Using 1. numbering for stepwise sequencing

Task:
1. Complete task value 123
2. Complete task value 456

Priority List

Using P0. for priority sequencing

Constraint:
P0. Do not change public API
P1. Minimize file edits
P2. Improve naming where obvious

Grouping

Nested as whitespace-sensitive aligned columns

Retrieval:
- Value one
- Value two
- Backend Services:
  1. Auth service is at website.com/auth
  2. User service is located at C:\Services
- Value four

Wildcarding

Using * asterisk to represent anything

Scope:
- C:\Project
- Data_*.txt

Symbolic Referencing

Using [ ] brackets to anchor and name a context as a reusable reference

[Context]:
- Color red
- Color green

Task:
1. Mix [Context] values AS [Combined]
2. Observe [Combined]
3. Subtract red from [Combined] and observe

Inline Aliasing

Using AS to instantiate a new symbol based on the immediate prior context

Observation:
- [Files] contains [Property] AS [FileProperties]

---

Section Verbiage

These section headers are examples, not reserved terms. Any single-word header is valid if its role is clear.

Retrieval

Scope Focus Target Directory URL Attachment

Framing

Context Observation Behavior Example Specification State Fact

Refinement (optional)

Consideration Constraint Ambiguity Unknown Assumption Note Mode Limit Caution

Execution

Task Action Question Evaluate Compare Explain Generate Simulate Modify Diagnose

---

Symbolic Referencing

Symbolic referencing uses bracket-wrapped identifiers ([Target], [Files], [Property]) to anchor named contexts as reusable references.

Each symbol represents a single explicitly defined context.

Purpose

• Eliminate repetition
• Preserve semantic consistency
• Reduce token usage
• Constrain interpretation

Definition

[Symbol]:
- Definition or list of definitions
- Conditional consideration

Referenced as:

Task:
- Evaluate [Symbol]

Symbol Composition

[Files]:
- ALL component.json files

[Property]:
- "inspector"

[Target]:
- [Files] + [Property]

Rules

• Single Responsibility — one concept per symbol
• Explicit Definition — no vague or inferred meaning
• Consistent Usage — no redefinition or overloading
• No Implicit Expansion — treat as fixed reference
• Reusability — usable across all sections

Constraints

• Use only when reused or critical – overdoing it increases tokens
• Avoid trivial symbols
• Keep names clear and meaningful
• Symbol names can also be PascalCase
• More than one word can be used, but can be more costly and prone to drift

Symbolic Referencing defines reusable semantic anchors, improving precision, consistency, and efficiency.

---

Usage Examples

Simple junk file removal

Directory:
- C:\Windows

State:
- Running low on disk space

Task:
1. Remove all cache files
2. Remove all temp files

Authorization error explanation with stepwise reasoning

Scope:
- Python API authentication
- https://www.webpage.com/authentication/documentation

Directory:
- C:\Python\auth.py

Context:
 - Authorization token refresh issue
 - Error upon user refresh
 - Error: "error 123"
 - Python API Authentication version: 4.5.6
 - Environment: Windows 10 22H2

Constraint:
- Be concise
- Specific explanation

Mode:
- Stepwise reasoning

Task:
1. Locate refresh logic
2. Identify failure condition
3. Determine root cause

Component data shape transference from one to another

Scope:
- C:\Programming\ShareBuilders\Skills\sharebuilders-skills\connectors\crm\core
- Trigger_ActivityCreated
- Trigger_ActivityUpdated
- Attached CSV file

Context:
- Trigger_ContactCreated.js maps output records with the mapContactCreatedRecords() function

Note:
- Do not retain backward compatibility

Task:
- In the same context as Trigger_ContactCreated, add an output record mapping for each Trigger_ActivityCreated and Trigger_ActivityUpdated
- The mapping will be based on the attached CSV file
- Update the component.json file to match the remapped output shape

Component property exposure realignment to match the API

Directory:
- C:\Programming\ShareBuilders\CRM\Efficio\Crm.UtilityScheduled
- C:\Programming\ShareBuilders\CRM\Efficio\EfficioSolution-Development-Feature
- C:\Programming\ShareBuilders\Skills\sharebuilders-skills\connectors\crm\core

Scope:
- The action connector "Action_SearchOpportunity" calls an API
- API Input: Array of GetPendingRequest
- Response Output: Array of PendingResponse
- Behavior: GET https://crm-api.share-builders.com/api/v1/Pending

Modify:
- All inputs matching the 'API Input' exposed as individual properties on the action, mapped flatly
- All outputs matching the 'Response Output' exposed as individual properties on the action, mapped flatly
- Include a separate raw output propery in the output
- Do not retain any unnecessary existing legacy properties or logic
- Use friendly names for properties
- Ensure value type is indicated on the property
- Add defaults where defined by 'API Input'
- Note 'optional' and 'default' in property tooltip as defined by 'API Input'

When complete:
- Update the related agents files to point to the API Input and Response Output references

Scan all user.json files to find invalid or missing email fields, and group the affected files by name

Scope:
- /app/users

[Files]:
- All user.json files

[Field]:
- "email"

[Invalid]:
- Missing [Field]
- [Field] does not contain "@"

[Affected]:
- [Files] where [Field] matches [Invalid]

Task:
1. Scan [Files]
2. Identify [Affected] based on [Field] and [Invalid]
3. List [Affected] grouped by file name