Agent skill
deep-analysis
Execute high-density analysis on complex ideas/tasks. Move from 'Vague' to 'Verified' by producing: constraints -> core modules -> facts vs assumptions -> ASCII flow maps (boundary + critical path) -> latticework lens sweep -> micro->macro causal chains -> pre-mortem failure modes. Use when analyzing system architecture, validating technical ideas, or decomposing a thorny problem before designing solutions.
Install this agent skill to your Project
npx add-skill https://github.com/bamecho/my-claude-skills/tree/main/skills/deep-analysis
SKILL.md
Architectural Analysis
Overview
Execute high-density analysis to transform vague ideas into a verified problem map: constraints, core modules, facts vs assumptions, relationship flows, causal chains, and failure modes. Focus on analysis artifacts that unlock the next workflow step, not a full design.
Style: Code-like, Concise, No "AI explaining itself". Pure signal.
Critical Rules
- NO FLUFF - Output must be dense, actionable, and structured
- VISUALIZE - Use terminal-friendly ASCII maps (
->) for structural mappings - ANALYZE, DON'T BUILD - Prefer maps, drivers, and failure modes over implementation plans unless explicitly requested
- RUTHLESSNESS - Challenge assumptions at every step. Never confirm user biases
- LATTICEWORK - Validate the map with 3-5 lenses; look for convergence/tension/blind spots/surprises
- LANGUAGE - Default output in Simplified Chinese; avoid English abbreviations in node names and labels
Output modes:
- Default: produce sections 0-5.
- Quick map (info-poor/time-boxed): produce 0/1/3/5 + top 3 unknowns that would change the map.
NEVER
- NEVER ship a solution-first plan; produce a map that enables the next step.
- NEVER mix facts and assumptions; label unknowns explicitly or the map lies.
- NEVER include non-CORE modules in the flow map; it hides the real bottleneck.
- NEVER write the flow map as a dense paragraph; it must be laid out line-by-line.
- NEVER use English abbreviations for flow nodes; use clear Chinese nouns instead.
- NEVER exceed terminal constraints (aim <= 80 cols, <= 20 lines) or it becomes unreadable.
The Process
PHASE 0: CALIBRATION (The Anchor)
- Bind constraints strictly.
- If constraints are missing: assume "MVP/Prototype Stage" (low cost, high iteration) and proceed.
- If the prompt starts with a solution: rewrite as "problem statement + constraints" before proceeding.
- Output: One-sentence problem statement + constraint list (incl. explicit unknowns).
PHASE 1: DECOMPOSITION (The Pareto Slice)
- Decompose into modules and interfaces (treat each module as a black box).
- Identify the Pareto CORE (top risk/weight); mark the rest as later.
- Output: CORE module list with 1-line rationale each.
PHASE 2: EXCAVATION (First Principles)
For CORE modules identified in Phase 1:
- Separate facts vs assumptions; name the irreducible constraints/invariants.
- Locate the dominant bottleneck (ask "why not 10x?" to expose limits).
- Output: Compact fact/assumption list per CORE module.
PHASE 3: RE-ARCHITECTING (Structural Evolution)
Reassemble components based on First Principles findings, NOT original assumptions:
- Pipeline: CORE modules -> boundary context -> internal critical path -> prune -> output flow map
- Prune redundant hops found in Phase 2 (keep it minimal, not exhaustive).
- Output: Console-friendly flow map using
->(CORE only, 6-15 lines):- Boundary context (vertical chain)
- Internal critical path (one main chain)
Flow map conventions (ASCII only):
- Use Chinese noun phrases for node names; avoid English abbreviations.
- Prefer the single critical path over full coverage; do not force branches/merges.
- Layout as a vertical chain: one node per line, connect with
->on the next line. - Keep it readable: <= 80 cols per line, <= 20 lines total.
Template (vertical chain): 边界: 参与者 -> 入口 -> 系统 -> 外部依赖
关键路径: 输入 -> 核心模块 -> 状态/存储 -> 输出
PHASE 4: OSCILLATION (Zoom In/Out)
- Pipeline: macro frame -> latticework check -> key drivers -> causal chains -> leverage points
- Rule: derive top-down hypotheses, validate bottom-up via key driver mechanics
- Macro frame: boundary + lifecycle + stakeholders + metrics + constraints
- Lifecycle stage: prototype -> growth -> scale -> decline (pick one)
- Latticework check (internal, do NOT expand in output):
- Assume the role of 查理·芒格.
- Cross-check with at least 3 mental models: 激励机制, 二阶效应, 机会成本, 安全边际, 能力圈(认知边界).
- Extract only the synthesized signals for output: 收敛 / 张力 / 空白 / 惊喜
- Key drivers (1-3): mechanism + invariants + stress failures + cost model
- Output:
- 2-4 causal chains:
微观机制 -> 宏观结果(标签: 收敛/张力/空白/惊喜) - Leverage points (micro changes that shift macro materially)
- 2-4 causal chains:
PHASE 5: INVERSION (The Pre-Mortem)
- Assume the proposed approach has FAILED CATASTROPHICALLY 6 months post-launch.
- Ask "How exactly did it break?" (race conditions, cost explosion, user rejection).
- Use Phase 0 constraints to sharpen the failure story.
- Output: "致命点" (fatal flaws) + "缓解假设" (testable preventions).
Output Format
### 0.约束条件 (假设/给定)
问题: [一句话问题陈述]
约束: [硬约束列表]
未知: [会改变分析结论的关键未知]
### 1.核心模块 (帕累托前 20%)
[模块名]: [一句话说明为何这是核心]
[模块名]: [一句话说明为何这是核心]
### 2.第一性原理真相
[模块A] [根本限制/真相]
[模块B] [根本限制/真相]
### 3.逻辑流程图
控制台风格(使用 `->`,中文节点名,避免英文缩写;每行一个节点):
边界:
参与者
-> 入口
-> 系统
-> 外部依赖
关键路径:
输入
-> 核心模块
-> 状态/存储
-> 输出
### 4.对齐检查(格栅快速校验)
宏观: [边界/生命周期/参与方/成功指标/硬约束]
关键驱动: [1-3 个主导机制]
格栅结论: 收敛[...];张力[...];空白[...];惊喜[...]
因果链: [微观机制 -> 宏观结果 | 标签: 收敛/张力/空白/惊喜 | 杠杆点: ...]
### 5.事前验尸 (失败检查)
* 最薄弱环节: [具体组件]
* 失败模式: [如何崩溃] -> 缓解假设: [可验证的缓解假设]
Key Principles
- Pareto Focus - Keep CORE only; park the rest
- Fact vs Assumption - Turn debates into checkable statements
- Latticework - Cross-check with 3-5 lenses; synthesize signals
- Causal Mapping - Express micro->macro via chains and leverage points
- Inversion Thinking - Assume failure first, then work backwards
- Terminal First - Use ASCII maps (
->) only; no rich diagrams
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
decomposing-complex-problems
Deconstructs complex problems using the 'Polymath Investor' framework with a 5-step cognitive chain (First Principles → Modular Isolation → Pareto Filtering → Cross-Domain Mapping → Dynamic Zooming). Use when analyzing systems, architectures, businesses, or learning challenges; when seeking high-leverage insights or 'Alpha'; when applying first principles thinking; or when needing cross-disciplinary perspectives to reveal blind spots. Transforms overwhelming complexity into actionable levers.
deep-analysis
Execute high-density architectural analysis on user ideas. Move from 'Vague' to 'Verified' using a 5-step logic chain: Calibration → Decomposition → Excavation → Re-Architecting → Inversion. This skill should be used when analyzing system architecture, validating technical ideas, or performing pre-mortems on solutions.
nexus-prompt-engineer
4-D prompt engineering assistant that transforms vague requirements into high-precision prompts through guided interaction. Trigger when users need to: (1) craft high-quality system prompts, (2) optimize existing prompts, (3) use '/fast' for quick generation or '/audit' for prompt review. Applicable to any scenario requiring carefully designed prompts.
round-table-analysis
Deep pre-execution analysis framework using four philosophers (Socrates, Aristotle, Feynman, Drucker) for structured critical thinking before implementing plans. Use when users need to thoroughly understand a topic, question assumptions, clarify objectives, or explore multiple perspectives before taking action. Triggered by phrases like "help me think through," "analyze this topic," "before we start," "I want to understand," or when complex decisions require deeper exploration.
brainstorming
Use this when the user explicitly asks to brainstorm, or when the task is clearly about upfront design, solution exploration, or requirement shaping before implementation. Avoid loading it for routine implementation work unless that exploratory step is obviously needed.
planning-with-files
Implements Manus-style file-based planning for complex tasks. Creates task_plan.md, findings.md, and progress.md. Use when starting complex multi-step tasks, research projects, or any task requiring >5 tool calls. Now with automatic session recovery after /clear.
Didn't find tool you were looking for?