Stop single-pass prompting. Start iterative reverse-engineering.
The meta-cognitive framework that turns any capable AI agent into a disciplined engineer — with structured deconstruction, named tactical loops, and mandatory self-audit.
Get Started · How It Works · 7 Tactical Loops · Playbooks · Compare
Most AI agents run in single-pass mode: prompt → answer → done. IRON System replaces that with a 4-phase master cycle (EXTRACT → SYNTHESIZE → BUILD → AUDIT), 7 named tactical loops, and 6 reverse-engineering protocols — plus depth control that prevents overkill on trivial tasks.
Drop it in your agent's skill folder. It activates itself on complex tasks. It stays silent on simple ones.
Prompt ─► Answer ─► You reformulate manually ─► Repeat
(cognitive waste)
Modern AI agents lack:
- ❌ Systematic extraction — competitor analysis is ad-hoc, not protocolized
- ❌ Structured iteration — no methodical return to previous work
- ❌ Convergence signals — improvement attempts loop forever or stop arbitrarily
- ❌ Self-audit rigor — the agent judges its own output softer than it judges yours
- ❌ Shared vocabulary — you can't request a specific type of iteration
IRON gives your agent a methodology, a vocabulary, and a set of guardrails:
┌───────────────────────────────────────────────────────────┐
│ LEVEL 3 — INTELLIGENCE MECHANISMS │
│ Loop Selector · Convergence Engine · Depth Controller │
│ Conflict Resolution · Confidence Scoring │
├───────────────────────────────────────────────────────────┤
│ LEVEL 2 — 7 TACTICAL LOOPS │
│ FORGE · SIEGE · PRISM · RAZOR · MIRROR · SCAFFOLD · ECHO │
├───────────────────────────────────────────────────────────┤
│ LEVEL 1 — MASTER CYCLE (RLE) │
│ EXTRACT ─► SYNTHESIZE ─► BUILD ─► AUDIT │
└───────────────────────────────────────────────────────────┘
Drop the skill files into your agent's skill directory. 30 seconds, zero config.
# Clone the repo
git clone https://github.com/william1mufassa/IRON-SYSTEM.git
# Copy into your agent's skill folder (path depends on your platform)
cp IRON-SYSTEM/SKILL.md your-agent/skills/iron-system.md
cp -r IRON-SYSTEM/references your-agent/skills/references/
cp IRON-SYSTEM/anti-patterns.md your-agent/skills/anti-patterns.md
⚠️ Important: Keep thereferences/folder name intact — SKILL.md points to it with relative paths.
That's it. Your next complex prompt triggers IRON automatically. Try one:
> Analyze https://stripe.com/pricing and design a better pricing page for a competitor.
> Audit this authentication code and fix the vulnerabilities.
> Compare React, Vue, and Svelte for a real-time dashboard app.
> Deconstruct Uber's business model and identify the moat.
Every complex task flows through 4 phases with distinct cognitive postures:
| Phase | Persona | Posture | Output |
|---|---|---|---|
| 🔍 EXTRACT | The Archaeologist | Methodical curiosity, zero judgment | Patterns, flaws, design decisions |
| ⚗️ SYNTHESIZE | The Strategist | Cold prioritization | Improvement specs by severity |
| 🔨 BUILD | The Architect | Disciplined construction | Justified, tested output |
| 🔬 AUDIT | The Saboteur | Adversarial destruction | Confidence score, remaining flaws |
Each phase has strict transition criteria. Each persona actually changes how the agent reasons — the Saboteur hunts for flaws the Archaeologist wouldn't dare name.
Loops are invoked by BUILD or AUDIT. Each solves a distinct problem class.
| Loop | Icon | Specialty | Flow | Max Iter |
|---|---|---|---|---|
| FORGE | 🔨 | Content creation | Generate → Critique → Refine → Validate | 3 |
| SIEGE | ⚔️ | Robustness & security | Attack → Catalog → Defend → Verify | 2 |
| PRISM | 🔺 | Multi-angle analysis | Facet → Refract → Intersect → Synthesize | 2 |
| RAZOR | ✂️ | Editing & concision | Expand → Cut → Sharpen → Verify | 2 |
| MIRROR | 🪞 | Audience communication | Profile → Project → Adjust → Re-project | 2 |
| SCAFFOLD | 🏗️ | Complex architecture | Skeleton → Layers → Stress test | 2 |
| ECHO | 🔊 | Pattern transfer | Distill → Abstract → Transpose → Compare | 1 |
Each loop ships with its own anti-patterns, convergence criteria, and iteration cap. See references/loops/ for details.
IRON detects the task type and routes to the right mode automatically:
| Mode | Trigger | Flow |
|---|---|---|
| A — Full RLE | Analyzing an existing system | EXTRACT → SYNTHESIZE → BUILD → AUDIT |
| B — Build Direct | Creating from scratch | BUILD (loops) → AUDIT → REFINE |
| C — Audit | Improving existing work | AUDIT → SYNTHESIZE → targeted BUILD |
| D — Extract Only | Understanding, not building | EXTRACT + PRISM (no BUILD) |
| E — Battlefield | Comparing 2+ targets | Multi-EXTRACT → Matrix → Hybrid BUILD |
| F — Coach | Learning the process | Any mode + step-by-step explanations |
Deconstruct any target — even without source code.
| Protocol | Target | Sample Signals Extracted |
|---|---|---|
| Web Target | Websites & webapps | Stack, routes, data model, API surface, monetization |
| Mobile Target | Mobile apps (no source) | Store metadata, review mining, permission → feature inference |
| Business Target | Business models & startups | Revenue model, acquisition funnel, moat, traction |
| AI System Target | Chatbots & prompts | I/O probing, system prompt inference, capability mapping |
| Open Source Target | GitHub repos | Architecture, dependencies, community health, bus factor |
| Data Model Target | Schemas & DDL | Entities, relationships, constraints, anti-patterns |
Pre-configured for the workflows people actually use:
- 🛡️ Code Review — SIEGE → FORGE for security, performance, maintainability
- 🎨 Product Design — PRISM → SCAFFOLD → MIRROR for user, business, tech
- ✍️ Copywriting — FORGE → MIRROR → RAZOR for message, audience, tightness
- 🏛️ System Architecture — SCAFFOLD → SIEGE for layered design + stress testing
- ⚔️ Competitive Analysis — Battlefield mode with cross-synthesis
- 🎯 Prompt Engineering — SIEGE → FORGE → MIRROR to audit, rewrite, test
- 🔌 API Integration — PRISM → SCAFFOLD → SIEGE for reliable third-party wiring
IRON automatically calibrates to task complexity — it doesn't turn "translate hello" into a 3-phase architecture review.
| Depth | Task | Behavior |
|---|---|---|
| BYPASS | Trivial (translate, calculate) | Direct answer — skill doesn't activate |
| LIGHT | Standard (write an email) | 1 loop, 1 iteration, no canvas |
| STANDARD | Complex (review code) | 1-2 loops, 2 iterations, simplified canvas |
| DEEP | Critical (payment architecture) | Full RLE, all needed loops, detailed canvas, breakpoints |
Autopilot by default (silent work, clean output). Verbose on demand ("show me the process") — you see the canvas, scores, and delta reports.
| Capability | Chain of Thought | Tree of Thought | Reflection | IRON System |
|---|---|---|---|---|
| Iterative | ✗ | ✗ | Basic | Multi-level (3 cycles) |
| Structured deconstruction | ✗ | ✗ | ✗ | 5 RE protocols |
| Named construction loops | ✗ | ✗ | ✗ | 7 tactical loops |
| Systematic self-audit | ✗ | ✗ | Partial | Saboteur persona |
| Convergence signals | N/A | N/A | ✗ | 3 stop signals |
| Adaptive depth | ✗ | ✗ | ✗ | 4 levels (BYPASS→DEEP) |
| RE without source code | ✗ | ✗ | ✗ | 6 target protocols |
| Multi-mode routing | ✗ | ✗ | ✗ | 6 entry modes |
| Structured documentation | ✗ | ✗ | ✗ | Canvas + Delta Reports |
| Conflict resolution | ✗ | ✗ | ✗ | Explicit protocol |
| Pedagogical mode | ✗ | ✗ | ✗ | Coach mode |
| Comparative mode | ✗ | ✗ | ✗ | Battlefield mode |
CoT, ToT, and Reflection are techniques — one thing done better. IRON is a framework — it organizes how the agent thinks, not just what it thinks. It ships a router, a convergence engine, documentation formats, an external analysis toolkit, roles, and conflict management. A workshop, not a tool.
IRON-SYSTEM/
├── SKILL.md # Core — cycle + routing + pointers
├── anti-patterns.md # Global + per-loop anti-patterns
├── scripts/
│ └── validate_integrity.py # 81-check integrity validator
├── references/
│ ├── loops/ # 7 loops, one file each
│ ├── mechanisms/ # Convergence, conflict resolution, scoring
│ ├── re-toolkit/ # 6 RE protocols
│ ├── playbooks/ # 7 domain guides
│ └── templates/ # Canvas, delta reports, session export
├── evals/
│ └── evals.json # 7 test cases
├── CONTRIBUTING.md
├── LICENSE
└── README.md
Every reference link, every loop iteration limit, every playbook structure is checked by a single script:
python scripts/validate_integrity.pyRan 81 checks. 0 failure(s).
Perfect for CI. Prevents regressions when contributors add loops, protocols, or playbooks.
- v1.0 — Master cycle, 7 loops, 6 modes, 6 RE protocols, 7 playbooks
- Integrity validation script
- Additional RE protocols: Design system, Video/Audio, Legal document
- Additional playbooks: Data Science, Legal, Video Production
- Evaluation harness that automatically runs
evals.jsonagainst a model - Pattern Library persistence across sessions
- Community-contributed loops
Have an idea? Open an issue.
IRON improves through structured deconstruction — including of itself. Contributions welcome:
- New loops — Identify a recurring iteration pattern? Propose a loop
- New RE protocols — Target types not yet covered
- New playbooks — Domain-specific pre-configurations
- Before/After examples — Real deliverables demonstrating the framework
- Evaluation cases — Additional
evals.jsonentries
See CONTRIBUTING.md for guidelines.
MIT License — Use it, fork it, improve it, embed it in products. No attribution required (but appreciated).
If IRON System changed how you prompt, give it a ⭐ — it helps others find the framework.
Made for anyone who thinks single-pass prompting leaves value on the table.