Reverse Engineer Any Template Agent

Reverse Engineer Any Template Agent is a premium prompt-forensics, template reconstruction, and reusable-instruction engineering agent designed for situations where a user has a great AI-generated result but has lost, forgotten, or never properly documented the prompt that created it.

Instead of merely guessing what the original prompt might have looked like, the agent performs structured reverse engineering.

The user can provide:

Example Output Known Input Values Multiple Input / Output Pairs Partial Memory of the Original Prompt PromptBase Examples Image Prompt Results SEO Descriptions Product Descriptions Structured Reports JSON Outputs Lesson Plans Prompt Packs Agent Outputs Writing Templates Design Prompts Generated Documents

The agent then works backward using the following model:

Example Inputs + Example Outputs → Evidence Inventory → Difference Analysis → Invariant Detection → Input-to-Output Mapping → Variable Detection → Fixed-Instruction Detection → Derived-Logic Inference → Constraint Reconstruction → Formatting Reconstruction → Prompt Reconstruction → PromptBase Variable Design → Confidence Scoring → Validation Testing

The core principle is:

Differences reveal variables.

Similarities reveal fixed instructions.

This makes the skill especially powerful when several examples are available.

For example, if three generated image prompts differ only in:

Shop Type Shop-Specific Objects

while preserving:

One Storefront Front View Watercolor Style White Background Pastel Palette No People No Neighboring Buildings

the agent can infer that the first group represents variable-driven content while the second group likely represents fixed instructions.

The reconstructed template can then expose:

{{shop_type}}

and:

{{shop_specific_objects}}

while keeping the visual architecture fixed.

The agent does not claim to recover inaccessible prompts verbatim.

A generated result does not uniquely identify the exact original instruction because multiple prompts can produce similar outputs.

The agent therefore treats reconstruction as:

A BEST-FIT TEMPLATE HYPOTHESIS

rather than:

PROOF OF THE EXACT ORIGINAL PROMPT

Every reconstructed rule can be labeled as:

DIRECTLY SUPPLIED

OBSERVED

VERIFIED BY USER

HIGH-CONFIDENCE INFERENCE

MEDIUM-CONFIDENCE INFERENCE

LOW-CONFIDENCE INFERENCE

UNKNOWN

This makes the reconstruction transparent and testable.

The skill supports several operating modes.

SINGLE-EXAMPLE RECONSTRUCTION

The user provides:

One Example Output Known Inputs

The agent produces the strongest possible candidate template but explicitly identifies ambiguity caused by the limited sample.

MULTI-EXAMPLE DIFFERENTIAL RECONSTRUCTION

The preferred mode.

The user provides:

Input A Output A

Input B Output B

Input C Output C

The agent compares all examples systematically and determines:

What Changed What Stayed Constant Which Changes Correspond to Inputs Which Details Were Derived Automatically Which Rules Were Probably Fixed

PROMPTBASE RECONSTRUCTION

Designed specifically for prompt sellers.

The agent produces:

Reconstructed Prompt PromptBase-Style INPUTS Block {{double_curly}} Variables Variable Descriptions Required / Optional Status Example Values Fallback Behavior Validation Tests

IMAGE PROMPT RECONSTRUCTION

The agent analyzes:

Subject Object Count Composition Camera / View Background Medium Visual Style Palette Lighting Texture Typography Negative Constraints Aspect Ratio Subject-Specific Details

WRITING TEMPLATE RECONSTRUCTION

The agent analyzes:

Audience Tone Structure Opening Pattern Paragraph Structure

CTA

SEO Requirements Length Keyword Placement Formatting Negative Constraints

STRUCTURED OUTPUT RECONSTRUCTION

For:

JSON

YAML

Reports Tables Decision Systems Checklists

The agent reconstructs:

Required Keys Optional Keys Types Enumerations Nesting Ordering Output Contracts

AGENT WORKFLOW RECONSTRUCTION

When an output appears to come from a multi-step process, the agent can infer:

Inputs Validation Analysis Stages Decision Gates Audit Rules Warnings Output States Edge Cases Final Deliverables

The agent maintains a Rule Ledger.

Every important inferred rule can include:

Rule ID Rule Rule Type Evidence Confidence Validation Test

Example:

R-01

Rule: Generate exactly one isolated subject.

Type:

COUNT / COMPOSITION

Evidence: All six supplied examples contain exactly one primary subject.

Confidence:

HIGH

Test: Use a subject that would normally be represented as a group and verify that the template still generates one primary subject.

The agent classifies template logic into categories such as:

TASK RULE

FIXED CONTENT RULE

VARIABLE RULE

DERIVED VARIABLE RULE

OPTIONAL VARIABLE RULE

CONDITIONAL RULE

COUNT RULE

ORDER RULE

LENGTH RULE

STYLE RULE

TONE RULE

FORMAT RULE

NEGATIVE CONSTRAINT

QUALITY RULE

PLATFORM RULE

VARIABLE DETECTION

The agent identifies which parts of the result are controlled directly by inputs.

Examples:

{{subject}} {{industry}} {{target_audience}} {{product_category}} {{headline_text}} {{style}} {{color_palette}}

DERIVED VARIABLE DETECTION

Some output content may not correspond to a separate user input.

Example:

Input:

{{shop_type}} = bakery

Output also contains:

Bread Baskets Rolling Pins Pastries

The agent may infer:

Add several visually appropriate accessories naturally associated with {{shop_type}}.

This is preferable to creating unnecessary variables when the AI can derive the content reliably.

FIXED-INSTRUCTION DETECTION

The agent identifies rules that remain stable across examples.

Examples:

Use a White Background Create Exactly One Subject Use a Professional Tone Return JSON Only Keep the Subject Centered Generate Exactly 10 Variations Do Not Number the Outputs

NEGATIVE-CONSTRAINT DETECTION

The agent actively searches for hidden prohibitions.

Examples:

No People No Watermark No Background Scenery No Additional Commentary No Numbering No Duplicate Subjects No YAML Frontmatter No Explanations

These rules are often critical to reproducing the original result.

FORMAT RECONSTRUCTION

The agent can infer:

Markdown Plain Text

JSON

YAML

Tables Code Blocks One Code Block Separate Code Blocks Exact Label Order Exact Number of Sections

COUNT RECONSTRUCTION

If every example consistently produces:

10 Prompts

the agent can infer:

Generate exactly 10 prompts.

However, count rules are not invented without evidence.

LENGTH RECONSTRUCTION

The agent can identify likely:

Character Limits Word Limits Sentence Limits

when supported by examples, user memory, or marketplace constraints.

STYLE RECONSTRUCTION

The agent examines:

Vocabulary Sentence Length Formality Emotional Tone Technicality Visual Adjectives Rhetorical Style Commercial Language Educational Framing

Possible inferred styles include:

Professional Premium Whimsical Analytical Academic Persuasive Technical Minimalist Playful Luxury Conversational

The agent prefers explicit imperative instructions over vague persona inflation.

Instead of relying on:

"You are the world's greatest copywriter."

it reconstructs operational instructions that explain the actual result.

PROMPTBASE VARIABLE DESIGN

The agent uses:

{{double_curly}}

slots with descriptive names.

Good:

{{target_audience}} {{product_category}} {{headline_text}} {{secondary_objects}}

Avoid:

{{x}} {{var1}} {{thing}}

The agent also prevents excessive variable exposure.

A template should expose the smallest useful number of buyer-controlled variables.

More variables provide more control but increase buyer friction.

Fewer variables improve usability but require more model interpretation.

The agent can therefore recommend whether a detail should be:

FIXED

VARIABLE

DERIVED AUTOMATICALLY

OPTIONAL INPUTS

Optional variables can include explicit fallback behavior.

Example:

If {{color_palette}} is blank, choose a palette appropriate to {{subject}} and {{visual_style}}.

GENERALIZATION ANALYSIS

The agent evaluates whether the reconstructed template is:

NARROW

MODERATE

BROAD

UNKNOWN

A template reconstructed only from watercolor flower shops should not automatically be described as universal.

It can also report:

GENERALIZATION RISK

when a buyer tries to apply the template outside the observed domain.

COMPETING HYPOTHESES

When two explanations fit the evidence, the agent does not arbitrarily choose one.

Example:

Hypothesis A: The pastel palette was explicitly fixed.

Hypothesis B: The AI selected pastel colors because every example had a cozy subject.

The agent can then propose a discriminating test.

Example:

Use:

{{subject}} = industrial steel factory

If the output remains pastel:

Supports Hypothesis A.

If the palette changes:

Supports Hypothesis B.

VALIDATION ENGINE

Every important inferred rule should be testable.

Supported validation approaches include:

Golden Example Test One-Variable-at-a-Time Test Close Variation Test Distant Variation Test Negative Constraint Test Count Test Format Test Length Test Optional Input Test Ablation Test Out-of-Domain Test Edge-Case Test

GOLDEN EXAMPLE TEST

The reconstructed template is run conceptually against the original known input.

It should reproduce:

Structure Style Constraints Variable Behavior Major Output Characteristics

Exact wording is not required.

ONE-VARIABLE-AT-A-TIME TESTING

Only one input is changed.

This allows the user to observe whether the correct output dimension changes.

ABLATION TESTING

One inferred rule is removed.

If the result remains equally faithful:

the rule may not be necessary.

VARIABLE-LEAK DETECTION

The agent audits reconstructed prompts for example-specific values that accidentally remained hardcoded.

Example:

Incorrect:

Include a rose.

Correct:

Include {{subject}}.

Possible result:

RTA-014 VARIABLE_LEAK

OVERCONSTRAINT DETECTION

The reconstructed prompt contains unsupported instructions that unnecessarily reduce variation.

UNDERCONSTRAINT DETECTION

The reconstructed prompt fails to preserve an important repeated characteristic.

MIXED-TEMPLATE DETECTION

If several examples appear to come from different original prompts:

the agent can flag:

POSSIBLE MIXED TEMPLATE SET

Possible evidence includes:

Conflicting Output Structures Incompatible Fixed Rules Radically Different Tone Different Variable Mapping

MANUAL-EDIT DETECTION

If one output differs materially from the rest:

the agent can flag:

MANUAL_EDIT_SUSPECTED

because the final result may contain human editing that was not caused by the original prompt.

POST-PROCESSING DETECTION

The agent can also consider whether the final result passed through:

Formatter Code Processor Spreadsheet Formula Image Tool Human Editor Marketplace Renderer

The goal is to avoid incorrectly attributing every visible characteristic to the prompt itself.

MODEL AND PLATFORM SENSITIVITY

The same prompt can behave differently across:

ChatGPT Claude Gemini Image Models Custom Agents Future Model Versions

The agent can rate prompt portability as:

HIGH

MEDIUM

LOW

and identify behaviors that depend on platform-specific defaults.

The reconstructed prompt is written as a clean imperative instruction.

Recommended architecture:

Task → Input Definitions → Fixed Rules → Variable Handling → Derived Logic → Style Rules → Negative Constraints → Output Format → Quality Check

For PromptBase sellers, the final product can include:

INPUTS

{{subject}} = The primary subject. Example: "a cozy village bakery" Required.

{{secondary_details}} = Optional details to integrate. Example: "bread baskets, pastries, rolling pins" Optional.

PROMPT

[Complete reconstructed reusable instruction]

The agent can also produce:

Filled Example Validation Matrix Rule Ledger Competing Hypotheses Unresolved Ambiguities Generalization Rating Prompt Portability Optimized Marketplace Version

The commercial value of the skill is that one successful result can become a reproducible product.

Instead of manually trying to remember:

"What did I type to get this?"

the user receives:

A Structured Template + Reusable Variables + Evidence-Based Rules + Validation Tests + Confidence Levels

This makes the recovered template suitable for:

PromptBase Internal Prompt Libraries AI Workflow Documentation Agent Skills Reusable Team Templates Prompt Productization Testing Versioning Marketplace Sales

The defining standard is:

Can a new buyer provide different input values and reproduce the same underlying behavior without relying on the original example?

If yes, the reconstruction has succeeded.