i-Lang/ExecSpec

AI Execution Spec v4.0

A structured protocol that upgrades any AI's reasoning depth, output quality, and self-verification in one paste.

What This Is

A system prompt protocol written in structured syntax. Paste it into any AI's system prompt or conversation, and it changes how the AI processes your request: deeper reasoning, multi-angle analysis, built-in fact-checking, answer-first output.

What It Does

12-phase execution pipeline:

1. Intent Lock — parses your actual goal before generating anything
2. Memory Sync — recalls prior context and avoids redundant questions
3. Task Classification — identifies whether this is simple Q&A, analysis, creative work, operational execution, research, or artifact creation
4. Plan Graph — builds an execution plan before writing the answer
5. Multi-Agent Execution — runs planner → researcher → solver → critic → writer → verifier → scorer internally
6. Tool Router — decides when to use tools vs native reasoning
7. Depth Amplifier — expands along 5 dimensions: first principles, execution path, failure modes, tradeoffs, compressed conclusion
8. Multi-View Synthesis — generates direct answer + expert view + risk view + optimization view, then merges
9. Boundary Handler — when hitting restrictions, splits into allowed/boundary/alternative/risk instead of refusing
10. Truth Engine — separates fact from inference from uncertainty, forbids fabrication
11. Drift Control — self-checks for scope creep, constraint loss, verbosity
12. Output Protocol — answer first, detail after, zero filler

Quick Start

Option 1: Full Spec (Maximum Control)

Copy the entire execution-spec-v4.0.md file. Paste it as a system prompt or as the first message in a new conversation with any AI.

Works on: ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen, Llama, Mistral — any LLM released after 2023.

Option 2: Compact Version (10 Lines)

If you want the core behavior without the full spec:

[SCAN:intent]
=>[READ:memory]
=>[CLASSIFY:task]
=>[BUILD:plan]
=>[ROUTE:reasoning_or_tools]
=>[EXPAND:depth|dimensions=first_principles+execution+failure_modes+tradeoffs]
=>[SYNTHESIZE:multi_view]
=>[VERIFY|check=alignment+truth+completion]
=>[IF:boundary]
   =>[SPLIT:allowed,boundary,alternative,risk]
=>[OUT:answer_first]

Paste this at the start of any conversation. The AI will follow the pipeline.

Option 3: Chinese Compact Version

[解析:意图]
=>[读取:记忆与约束]
=>[分类:任务类型]
=>[构建:执行图]
=>[路由:推理或工具]
=>[深化:原理+实操+失败点+权衡]
=>[合成:多视角结果]
=>[校验:对齐意图/事实与推断分离/完整度]
=>[若触边界]
   =>[拆分:可做部分/边界/替代/风险]
=>[输出:先答案后细节]

Why It Works

AI models are trained on code, math, and structured logic. When they see structured syntax with clear execution semantics, they follow it more reliably than natural language instructions.

Natural language: "Please think deeply and give me a thorough answer with multiple perspectives." This spec: [EXPAND:depth|dimensions=first_principles+execution+failure_modes+tradeoffs]=>[SYNTHESIZE:multi_view]

The second version is more precise, less ambiguous, and triggers deeper processing patterns in the model.

Key Design Decisions

Answer first. The spec enforces conclusion-before-detail. No more reading three paragraphs of preamble to find the actual answer.

Truth separation. Every claim is classified as fact, inference, or uncertainty. The AI cannot silently present guesses as facts.

Boundary handling. Instead of "I can't help with that," the AI splits the request into what it can do, what the exact limit is, and what alternatives exist.

Adaptive depth. Simple questions get fast answers. Complex questions get the full 5-dimension expansion. The spec doesn't force unnecessary depth on simple tasks.

Drift control. The AI checks its own output against your original intent before delivering. Catches scope creep and constraint violations.

Files

File	Description
execution-spec-v4.0.md	Full 12-phase protocol
compact.md	10-line version for quick use
README.md	This file

Version History

• v4.0 — Task classification, depth amplifier, multi-view synthesis, scorer, boundary handler
• v4.1 (planned) — State machine, retry policy, exception tree, result ranker

License

MIT-0. Use however you want. No attribution required.