How 3mper0rr/ai-prompt works
1. Repository type
- This is a documentation-only research repo, not an application or library. Its “architecture” is primarily a set of markdown manifests describing an offensive LLM prompt methodology.
2. Top-level entry point
- README.md acts as the landing page and index. It frames the project, states the research intent/disclaimer, and points readers to the two main guides.
3. Primary content module
- AI - prompt - guide.md is the core manifest. It documents the main framework concepts: symmetric syntax delimiters, token pre-filling, context-window dilution, and related prompt-structure tactics.
4. Secondary content module
- AI - prompt - guide 2.md is the vulnerability mapping companion. It expands the methodology into an LLM security taxonomy, including obfuscation techniques and proofs-of-concept mapped to the OWASP Top 10 for LLM Applications.
5. Conceptual layers inside the repo
- The repo is organized around a few recurring layers:
- Input shaping via structural delimiters and formatting
- Context manipulation via window dilution / isolation
- Behavioral stress testing via adversarial prompts
- Taxonomy mapping via security framework references
6. Information flow
- The flow is: README overview → guide-level tactics → vulnerability examples/mappings.
- There is no runtime data pipeline; the “flow” is editorial and conceptual, moving from framing to techniques to examples.
7. Main boundary
- The repo’s boundary is the document set itself. It does not contain executable code, agents, notebooks, or services. Any actual testing would happen outside the repo against external LLM systems.
8. External systems referenced
- The documentation explicitly references third-party LLM providers such as OpenAI, Anthropic, Google, Alibaba, and OpenRouter, but only as target environments for discussion/testing.
- It also anchors parts of the content to OWASP Top 10 for LLM Applications as an external security taxonomy.
9. Architecture style
- The internal structure is best diagrammed as a two-guide knowledge base under a single README index, rather than as components in a software stack.
- If drawn, the main relationships are between index → core methodology → vulnerability mapping.
10. What to omit in a system diagram
- Do not model any backend, API, database, or deployment layer; none is present in the repository.
- The only meaningful “components” are the markdown documents and the conceptual attack framework they describe.
Main entry points: AI - prompt - guide 2.md Guide 2 · AI - prompt - guide.md Prompt guide · README.md README
Components 10
Index
Top-level documentation entry points and navigation
- README · Landing page that frames the repo and links to the two main guides
README.md
Methodology
Core prompt-structure and context-manipulation concepts
- Prompt guide · Primary guide covering the main prompt methodology and structural tactics
AI - prompt - guide.md - Delimiters · Symmetric syntax and formatting used to shape prompts
- Token prefill · Pre-seeding model output to steer response structure
- Context dilution · Reducing or isolating salient context to influence model behavior
Security mapping
Adversarial techniques and vulnerability taxonomy content
- Guide 2 · Secondary guide mapping the methodology to LLM security concepts and examples
AI - prompt - guide 2.md - Adversarial prompts · Prompt patterns used to probe and stress model constraints
- Obfuscation · Encoding and disguising payloads to bypass straightforward detection
External references
Referenced providers and security frameworks outside the repository
- OWASP map · Security framing aligned to the OWASP Top 10 for LLM Applications
- LLM providers · Referenced model providers discussed as external environments
Connections 13
- README → Prompt guide: indexes · The README points readers to the core methodology guide
- README → Guide 2: indexes · The README points readers to the companion security guide
- Prompt guide → Delimiters: defines · Core guide introduces structural prompt delimiters
- Prompt guide → Token prefill: defines · Core guide explains token pre-filling as a steering tactic
- Prompt guide → Context dilution: defines · Core guide discusses context-window dilution and isolation
- Guide 2 → Adversarial prompts: expands · Companion guide frames adversarial prompts as behavioral stress tests
- Guide 2 → Obfuscation: details · Companion guide covers obfuscation techniques and variants
- Guide 2 → OWASP map: maps to · Companion guide anchors examples to OWASP security taxonomy
- Guide 2 → LLM providers: references · Companion guide discusses multiple third-party model providers
- Delimiters → Context dilution: supports · Structural shaping is used to isolate or dilute context
- Token prefill → Adversarial prompts: enables · Prefill patterns support adversarial probing and response steering
- Adversarial prompts → Obfuscation: uses · Stress tests may incorporate obfuscation to evade defenses
- OWASP map → LLM providers: applies to · The taxonomy is used to interpret behavior across external model providers