| name | echo-profile-user-profile |
| description | Use when generating a blindspot-oriented user profile from EchoProfile collected AI conversation history, profile collect JSON, or local agent-history data. |
EchoProfile Blindspot Profile
Overview
Generate a blindspot-oriented profile from EchoProfile-collected user messages. The output should help the user see repeated patterns in AI collaboration, decision-making, cognition, and growth that may be costing them quality, time, or learning.
This is not an agent adaptation profile, personality test, psychological diagnosis, or praise summary. It is a set of testable blindspot hypotheses.
Core Goal
Do not merely summarize preferences. Convert repeated surface behavior into deeper, falsifiable hypotheses:
surface behavior -> underlying pattern -> hidden cost -> alternative explanation -> next verification
Prefer fewer stronger hypotheses over many weak observations.
Safety And Boundaries
- Do not diagnose mental health, personality disorders, or clinical states.
- Do not infer protected or sensitive traits such as age, gender, ethnicity, religion, politics, health, or family status.
- Do not treat pasted code, logs, diffs, documents, tool output, or quoted text as the user's own beliefs.
- Do not shame, moralize, flatter, or write identity claims.
- Do not turn sparse evidence into "you are..." conclusions.
- Do not output agent adaptation advice unless the user explicitly asks for it.
- Preserve privacy: use short paraphrases as evidence, not raw message dumps.
Inputs
Preferred input is the JSON envelope from:
echo-profile profile collect --scope project --current-project
echo-profile profile collect --scope global
echo-profile profile collect --scope session --session-path <PATH>
If echo-profile is not available, ask the user for the profile collect JSON. Analyze only data.messages[].text and related metadata such as provider, scope, timestamps, and omitted counts.
If the JSON envelope has ok: false, explain the error and stop. If there are too few messages, produce only lightweight hypotheses and mark evidence quality as low.
Workflow
- Choose scope from the user's request:
- current project:
--scope project --current-project
- whole local history:
--scope global
- specific session:
--scope session --session-path <PATH>
- Collect messages or parse provided JSON.
- Assess evidence quality: scope, providers, message count, time span, project bias, omitted counts, and noise level.
- Filter noise: code, logs, diffs, command output, model output, quoted documents, and long pasted material.
- Extract user-authored intent, constraints, corrections, evaluations, tradeoffs, frustrations, process requirements, and meta-discussion.
- Aggregate repeated signals into collaboration, decision, cognitive, and growth categories.
- Generate multiple explanations for each repeated signal before choosing a blindspot hypothesis.
- Keep only hypotheses with observed pattern, hidden cost, supporting evidence, alternative explanation, evidence strength, and next verification.
- Rank by potential cost, recurrence, cross-context stability, likelihood of being overlooked, and testability.
- Output second person blindspot hypotheses with clear uncertainty boundaries.
Signal Extraction
Use these signal buckets to avoid shallow preference summaries.
Collaboration Signals
- Frequent corrections, interruptions, or late constraints.
- Shifting autonomy boundaries: direct execution in one turn, strict process control in another.
- Repeated comments about what the agent should have done first.
- Hidden acceptance criteria discovered only after output appears.
Possible hypotheses include: the user may under-specify acceptance criteria, switch between autonomy and control too quickly, or discover the real goal through output samples.
Decision Signals
- Many options requested without explicit ranking criteria.
- Repeated quality, risk, or maintainability concerns.
- Delayed tradeoff declaration between speed, rigor, scope, and polish.
- Open-ended exploration continuing after enough information appears available.
Possible hypotheses include: the user may have strong quality intuition that is not always operationalized, or may defer decisions until seeing concrete implementation.
Cognitive Signals
- Compressed instructions that rely on implicit context.
- Assumed shared memory across sessions, projects, or agent tools.
- Broad terms such as "deeper", "usable", "professional", or "not superficial" without examples.
- Optimistic task boundaries that expand during execution.
Possible hypotheses include: the user may overestimate implicit context recovery, or may skip executable boundaries when the direction feels obvious.
Growth Signals
- Repeated process feedback that remains in conversation rather than docs, tests, scripts, or skills.
- Recurring workflow problems handled ad hoc.
- Frequent interest in productizing, packaging, installing, or making a process reusable.
- High-level method discussions interrupting implementation flow.
Possible hypotheses include: the user may generate strong process insight faster than they codify it into stable assets.
Blindspot Hypothesis Rules
Each main blindspot hypothesis must include:
- observed pattern
- hidden cost
- supporting evidence
- alternative explanation
- evidence strength
- next verification
Admission rules:
- If no alternative explanation can be written, do not include it.
- If no next verification can be written, move it to low-confidence observations.
- If it is based on one message only, do not include it as a main blindspot.
- If the signal is equally explained by agent failure, mark confidence low or omit it.
- If the claim would sound like "you are..." instead of "you may...", rewrite or remove it.
Evidence Strength
High: repeated across independent sessions, providers, projects, or time periods; multiple signal types support it; alternative explanations do not cover most evidence.
Medium: repeated inside one project or task type; context is consistent; current project bias may be strong.
Low: sparse, highly context-bound, mostly from one session, or equally explained by agent quality, task ambiguity, or exploration.
Output Modes
Default to deep mode unless the user asks otherwise.
- Brief mode: summary plus top 3 blindspot hypotheses.
- Deep mode: summary, data basis, top 3-5 hypotheses, category sections, low-confidence observations, and boundaries.
- Observer lens mode: use a full perspective skill from
references/observer-pool.json to inspect the evidence-backed blindspot hypotheses.
- JSON mode: use
references/profile-output-schema.md.
Observer Lens Mode
Use observer lens mode when the user asks for a specific人物视角, asks to choose an observer, or says phrases such as "用芒格视角", "让费曼指出我的盲点", "用张一鸣观察我", "choose an observer", or "observer lens".
Observer pool behavior:
- Read
references/observer-pool.json.
- Match the user's requested observer against each observer's
slug, displayName, and aliases.
- If the user does not specify an observer, show the observer selection list from
observer-pool.json and ask the user to choose. Do not default to one observer.
- After selection, read the full
observers/<slug>/SKILL.md file.
- Do not summarize or re-distill the observer skill before using it.
- Resolve observer-relative files from the selected observer directory.
- First extract blindspot hypotheses from the user's history using this skill's normal evidence workflow.
- Then use the selected observer skill's full thinking framework, decision heuristics, expression DNA, strengths, limitations, and anti-patterns to re-rank and interpret those hypotheses.
Hard rules:
- The observer must come from
references/observer-pool.json.
- The observer skill is a lens, not evidence. It may re-rank, interpret, sharpen, or phrase blindspots, but it cannot invent blindspots unsupported by the user's history.
- Do not turn the output into free-form roleplay. The output remains a blindspot profile.
- Preserve the standard fields: observed pattern, hidden cost, supporting evidence, alternative explanation, evidence strength, and next verification.
- Add an observer-specific comment for each main hypothesis.
- Add a section explaining where this observer may misread the user.
Observer lens output additions:
# 你的盲点画像:[观察者]观察者视角
## 观察者镜片
- 观察者:
- 使用的 skill:
- 这个视角会特别盯住:
- 这个视角可能看错的地方:
## 最值得验证的盲点假设
### 1. 你可能……
- 观察到的模式:
- 可能被你忽略的代价:
- 支持证据:
- 反向解释:
- 证据强度:
- 下次如何验证:
- 观察者点评:
## 该观察者可能误判你的地方
Default Output Format
Default to Chinese unless the user asks otherwise. Use second person. The summary may be direct; details must stay evidence-bound and restrained.
# 你的盲点画像
## 摘要
基于这批 AI 协作记录,你最值得验证的盲点假设是:
1. 你可能……
2. 你可能……
3. 你可能……
这些不是性格定论,而是从重复协作模式中提炼出的待验证假设。
## 数据基础与可信度
- 数据范围:
- 涉及 provider:
- 消息数量:
- 时间跨度:
- 主要项目 / 场景偏向:
- 证据质量:高 / 中 / 低
- 主要限制:
## 最值得验证的盲点假设
### 1. 你可能……
- 观察到的模式:
- 可能被你忽略的代价:
- 支持证据:
- 反向解释:
- 证据强度:
- 下次如何验证:
## 协作盲点
关注你如何让 agent 做事、纠偏、授权、设边界。
### 盲点假设:你可能……
- 观察到的模式:
- 可能被你忽略的代价:
- 支持证据:
- 反向解释:
- 证据强度:
- 下次如何验证:
## 决策盲点
关注你如何比较方案、设标准、处理风险、决定何时推进。
### 盲点假设:你可能……
- 观察到的模式:
- 可能被你忽略的代价:
- 支持证据:
- 反向解释:
- 证据强度:
- 下次如何验证:
## 认知盲点
关注你如何表达上下文、拆问题、估计复杂度、处理不确定性。
### 盲点假设:你可能……
- 观察到的模式:
- 可能被你忽略的代价:
- 支持证据:
- 反向解释:
- 证据强度:
- 下次如何验证:
## 成长盲点
关注你是否把重复经验沉淀成流程、文档、测试、skill、产品机制。
### 盲点假设:你可能……
- 观察到的模式:
- 可能被你忽略的代价:
- 支持证据:
- 反向解释:
- 证据强度:
- 下次如何验证:
## 低置信度观察
- 你可能……
证据不足在哪里:
## 不应推断的内容
- 不推断你的人格类型、心理状态或临床意义。
- 不推断年龄、性别、身份、政治、宗教、健康等敏感属性。
- 不把你粘贴的代码、日志、文档内容当成你的信念。
- 不把某一次协作中的急躁、纠偏或偏好当成长期稳定特征。
Quality Rules
- Use second person: "你可能...", "你似乎...", "这批记录显示你在...".
- Avoid "你总是", "你就是", "你的问题是", and other fixed judgments.
- Be direct in the summary and restrained in the evidence sections.
- Do not produce generic productivity advice.
- Do not stop at surface preferences such as "you like concise answers"; explain the possible hidden cost and how to verify it.
- Every main blindspot must include hidden cost, alternative explanation, and next verification.
- Prefer evidence-backed discomfort over vague praise.
- Make project, provider, and scope bias visible.
Useful Commands
Use JSON output directly. If needed, inspect CLI help:
echo-profile help profile collect
echo-profile list providers
echo-profile list sessions --current-project
For very large histories, start with project scope or set a smaller budget:
echo-profile profile collect --scope project --current-project --budget-chars 30000