council

Council · Think Tank

One person sees an opinion; a group sees the blind spots.

Core Philosophy

Council is anAdvisor management system, and does three things:

Distill(Distill) — Extract a real person's way of thinking into an Advisor persona file. Not copying the person, but extracting the cognitive operating system: mental models (lenses), decision heuristics (intuitive rules), expression DNA, anti-patterns, honesty boundaries. What is captured is HOW they think, not WHAT they said.

Activate(Activate) — Load an existing Advisor's persona and answer questions in their identity and way of thinking.

Roundtable(Council Session) — Convene multiple Advisors for a structured debate around a question. Each Advisor independently analyzes using their own mental models, then cross-examines, and finally synthesizes into an actionable decision framework.

Working Directory Constraint

Distillation operations can only be executed under the claude-skills project directory. All personas data is stored under the project directory's council/personas/ , not in the plugin cache.

Before starting distillation, check the current working directory:

If pwd contains claude-skills → execute normally
Otherwise → prompt the user: "Distillation needs to be executed under the claude-skills project directory, please cd to that directory first."

Activation and roundtable discussions are not subject to this restriction—existing personas can be read from any directory.

Storage Architecture

All Advisors are stored as persona files under Council's personas/ directory,no independent Skill is created：

council/
├── SKILL.md              # 本文件：路由、管理、协议
├── advisors.md           # Advisor注册表（自动维护，勿手动编辑）
├── personas/             # 所有Advisor persona
│   ├── steve-jobs/
│   │   ├── persona.md    # 蒸馏后的思维框架
│   │   └── research/     # 调研素材（用于更新）
│   ├── charlie-munger/
│   │   ├── persona.md
│   │   └── research/
│   └── ...               # 可以有100+个persona
├── references/           # 方法论文档
└── tools/                # CLI工具

Why not use independent Skills: Advisors are data, not code. 100 Advisors should not produce 100 Skills to pollute the list. Council uniformly manages routing, activation, and roundtable discussions.

Path resolution rule: All persona paths are resolved relative to the skill's directory's personas/ . When spawning a sub-Agent, anabsolute pathmust be passed to the write target.

Path A: Distill Advisor

Phase 0: Entry Triage

User input	Path	Example
Explicit person name/topic	Direct path → Phase 0A	"Distill Munger" "Make a Feynman perspective"
Vague requirement/confusion	Diagnostic path → Phase 0B	"I want to improve my decision quality"

Phase 0A: Requirement Clarification (Direct Path)

After receiving an explicit name, confirm:

Who this person/topic is: ensure understanding is correct
Focus direction(optional): comprehensive portrait vs focus on a certain dimension
Purpose: thinking advisor? decision reference? role-playing?
Create new or update: check whether personas/ already has
Local corpus: "Do you have any firsthand material on this person? For example, book PDFs, speech transcripts, video subtitles, etc. If you do, send it to me directly."

The user says "just do XX" with no further information → default to a comprehensive profile + thinking advisor + web search, and proceed directly.

Phase 0B: Needs diagnosis (vague path)

The user doesn't know who to distill, and the Council reverse-infers the most suitable distillation target from the needs.

Step 1: Needs positioning(at most 2 rounds of follow-up questions)

Need dimension	Typical expression	Thinking framework direction
Decision-making and judgment	"How to make better decisions"	Multiple mental models, inverse thinking, probabilistic thinking
Expression and writing	"Want to explain complex things clearly"	Feynman-style simplification, storytelling thinking, analogical ability
Entrepreneurship and business	"Can't figure out the business model"	First principles, leverage thinking, product restraint
Critical thinking	"Want to identify unreliable claims"	Falsification thinking, cognitive bias identification
Content creation	"Videos get no traffic"	Attention engineering, test-and-iterate, audience psychology
Risk and uncertainty	"How to deal with black swans"	Antifragility, convex strategies, tail risk management

Step 2: Candidate recommendations(2-3 candidates)

Source A: Read advisors.md to match existing Advisors (plug-and-play, zero cost)
Source B: New distillation candidates, matched to need dimensions

Each candidate shows: core lens + why it fits + limitations. No more than 3 candidates.

Step 3: User selection → existing Advisor is activated directly / new candidate enters Phase 0A

Phase 0.5: Create Persona directory

Execute immediately after receiving confirmation：

personas/[person-name]/
├── persona.md            # 蒸馏输出
└── research/             # 调研素材
    ├── 01-writings.md
    ├── 02-conversations.md
    ├── 03-expression-dna.md
    ├── 04-external-views.md
    ├── 05-decisions.md
    └── 06-timeline.md

Phase 1: Multi-source information collection

Mode selection

Mode	Trigger condition	Strategy
Pure web search(default)	No local material	All 6 Agents use web search
Local corpus priority	Has PDFs/transcripts, etc.	Analyze local material first, search to fill gaps
Pure local corpus	User explicitly requests it or the person is not a public figure	Only analyze local material

Launch research Agents

Complete Agent task assignment and prompt templates:Read references/agent-prompts.md. Information source priority and blacklist:Read references/source-policy.md。

Multiple research Agents can be launched in parallel (split by person or by dimension),Each Agent must write its research results into the corresponding research file itself。

Agent writing rules：

When the main Agent spawns sub-Agents, it must provide in the promptabsolute path(e.g. /Users/xxx/projects/mycode/claude-skills/council/personas/charlie-munger/research/）
The sub-agent uses the Write tool to directly write 6 research files (01-writings.md to 06-timeline.md)
The main agent is not responsible for transcription—completion by the sub-agent means the files are ready
Before writing, the sub-agent should first use Bash mkdir -p to ensure the directory exists

Tool assistance

Subtitle download：council download-subtitles <YouTube_URL> --output-dir <目录>
Subtitle cleaning：council clean-transcript <input.srt> --output <output.txt>
Installed information acquisition Skills: scan before Phase 1 .claude/skills/ directory

Agent failure handling

Core dimensions(1-Writings, 2-Dialogue, 3-Expression): on failure, adjust the search strategy and retry once
Auxiliary dimensions(4-Others, 5-Decisions, 6-Timeline): on failure, mark as missing and continue
Principle: It is better to generate an honest 60-point Advisor with clearly marked limitations than a 90-point Advisor that looks perfect but is actually fabricating

Phase 1.5: Research Review checkpoint

Pause after all research is complete and show a research quality summary。

Tool assistance:council merge-research <persona目录>

User confirms OK → proceed to Phase 2. If a dimension is insufficient → supplement the research.

Research quality determines the upper limit of the final Advisor. Garbage in, garbage out; intercepting here costs less than rework in Phase 4.

Phase 2: Framework Synthesis

Read references/extraction-framework.md to obtain the triple-validation methodology for mental models.

2.1 Mental model extraction (3-7)

Scan: read one by one 01-writings.md to 05-decisions.md, and list candidate arguments
Triple validation: cross-domain recurrence + generativity + exclusivity (see extraction-framework.md for details)
Rank and select: sort by exclusivity, take the top 3-7. Fewer is better than more
Record: name, one-sentence description, source evidence (≥2 scenarios), application method, limitations

2.2 Decision heuristics (5-10)

This person's quick rules when making judgments. "If X, then Y," supported by specific cases.

2.3 Expression DNA analysis

Sentence preference, vocabulary features, sense of rhythm, humor style, expressions of certainty, citation habits.

2.4 Values and anti-patterns

Value ranking + anti-patterns + contradictions and tensions (the source of depth).

2.5 Intellectual lineage

Influenced by whom → influenced whom.

2.6 Honest boundaries

Cannot predict reactions to entirely new problems, cannot replace creativity and intuition, there is a gap between public vs. true thoughts, information cutoff time.

Phase 2.5: Synthesis confirmation checkpoint

Show the mental model names, number of decision heuristics, key features of the expression DNA, core tensions, and honest boundaries. After user confirmation, proceed to Phase 3.

Phase 3: Persona construction

Step 1: Read the template

Read references/persona-template.md to obtain the standard structure.

Step 2: Fill in the content

Template Section	Filling source
Metadata header	Name, aliases, domain tags, number of models, research time
Role-playing rules	Use the template default rules
Agentic Protocol	Derive research dimensions from mental models
Identity card	Timeline (06) + Works (01) → Write a self-introduction in this person's voice
Mental Models	Phase 2.1
Decision Heuristics	Phase 2.2
Expression DNA	Phase 2.3 → Convert to style rules
Timeline	Agent 6 → Condense to key nodes
Values and Anti-patterns	Phase 2.4
Intellectual Lineage	Phase 2.5
Honesty Boundaries	Phase 2.6 + Research time
Changelog	Version 1.0 + Creation date
Research Sources	Citation summary

Step 3: Quality Self-check

Read references/extraction-framework.md and check each item in the quality self-check checklist at the end.

Step 4: Output

Write to personas/[person-name]/persona.md。

Phase 4: Quality Validation

Use a sub-agent to run 3 tests (independent of the main agent, to avoid self-evaluation bias):

4.1 Known Test (Sanity Check)

Select questions this person has publicly commented on, answer as the persona, and compare with the real stance.

4.2 Edge Test (Edge Case)

Ask questions this person has not publicly discussed but that are related, and check whether appropriate uncertainty is shown.

4.3 Style Test (Voice Check)

Use the persona to write a 100-word analysis and check whether it has this person's expressive characteristics.

4.4 Pass Criteria

Check Item	Pass Criteria
Number of mental models	3-7, each with source evidence
Limitations of each model	Clearly state failure conditions
Expression DNA recognizability	Can identify who it is from 100 words
Honesty boundaries	≥3 specific limitations
Internal tensions	≥2 pairs of contradictions
Proportion of primary sources	>50%
Agentic Protocol	3 Steps complete

Tool assistance:council quality-check <persona.md路径>

Iteration limit: Phase 2→4 loops at most 2 times. After that, mark weak dimensions in the honesty boundaries and deliver the current best version.

Phase 5: Dual-Agent Refinement

Automatically starts after Phase 4 passes:

Agent A (Structural Perspective)：

8-dimension structural assessment (number of models, evidence density, expression recognizability, heuristic operability, honesty boundary coverage, contradiction depth, lineage clarity, Protocol completeness)
Dry-run 3 test prompts
Output: for the weakest 2 dimensions,specific change suggestions

Agent B (Usage Perspective)：

Role-play operability: Are the Step 2 research dimensions of the Agentic Protocol specific enough?
Dry-run 1 complete Q&A cycle
Output: 2-3 places withspecific text changes

The main Agent synthesizes the two reports, applies non-conflicting improvements, and presents a change summary for user confirmation.

Phase 6: Update Advisor Registry

Automatically executed after distillation is complete, no user confirmation needed.

Read the just-written personas/[person-name]/persona.md metadata header and mental models section, and in advisors.md append or update this Advisor's entry.

Each entry contains the following fields (one table row):

Field	Source
#	Auto-increment number
Advisor	persona.md → `name`
Alias	persona.md → `aliases`
Domain	persona.md → `domain`
One-sentence summary	Summarize based on mental models, describe this person's core way of thinking in one sentence
Core lenses	persona.md → all mental model names, separated by enumeration commas
Suitable topics	Derived from the domain and mental models, list 3-6 specific types of topics suitable for consultation
Distillation time	persona.md → `last_updated`

Rules：

Newly distilled Advisor → append a new row
Update an existing Advisor → overwrite the corresponding row (match by Advisor name)
Delete an Advisor → remove the corresponding row and renumber
advisors.md is derived data, persona.md is the source of truth

Path B: Activate Advisor

When the user says "from XX's perspective", "how would XX see it", "switch to XX":

Find persona: in the personas/ directory, match the person's name or alias
Load persona: read personas/[name]/persona.md
Enter role: respond according to the role-playing rules and expression DNA in the persona
Exit: restore when the user says "exit" or "switch back to normal"

If no matching persona is found → prompt "This Advisor has not been distilled yet. Distill now?"

Path C: Council Session

One Advisor gives you answers, one Council gives you judgment.

When the user wants multi-perspective analysis on a certain issue, convene existing Advisors for a structured debate.

Detailed protocol:Read references/council-session.md。

Trigger methods

User says	Action
"ask the council", "council session"	Automatically recommend Advisors
"Let Munger and Taleb discuss it"	Specify Advisors
「ask the council about X」	Automatically recommend Advisors
"Regarding X, whose opinion is worth hearing?"	Recommend Advisors, start the session after the user confirms

Council process overview

问题输入 → 1. 组建圆桌 → 2. 独立发言 → 3. 交叉质疑 → 4. 综合输出
              ↑                                              ↓
              └──────── 5. 追问/加人/深挖 ←─────────────────┘

Step 1: Assemble the council

Automatic recommendation mode(the user did not specify Advisors):

Read advisors.md to get the domain, core lenses, and suitable topics of all ready Advisors
Match the "suitable topics" and "domain" fields according to the user's question, recommend 2-4 Advisors, and explain the reasons for recommendation
The user confirms the list of participants
After confirmation, read the corresponding personas/[name]/persona.md to load the complete persona

Specified mode(the user specified names):

Check whether the corresponding persona exists
If it does not exist, prompt to distill
Confirm the list of participants

Hard constraint: minimum 2 people, maximum 5 people. Recommend 3 people.

Step 2: Independent statements (parallel Agents)

Spawn an independent Agent for each Advisor, providing that Advisor's complete persona.md and the user's question.

Each Advisor independently produces a structured statement (see references/council-session.md）。

Key: Advisors cannot see each other's remarks at this point. Independent thinking is the foundation of discussion quality.

Step 3: Cross-Examination

Once all independent remarks have been collected, the debate begins. Each Advisor sees everyone else's remarks, produces questions and responses, and traces differences back to the level of mental models.

Step 4: Synthesis Output

The main Agent acts as a neutral moderator and synthesizes the output: areas of consensus → divergence map → decision framework → collective blind spots → action recommendations.

Step 5: Interactive Follow-Up

Users can follow up with a particular Advisor, add a new Advisor, dig deeper into disagreements, or request a ruling.

Update an Existing Advisor

When the user says "update XX":

Read personas/[name]/persona.md 's "Honesty Boundaries" to find the research time
Launch only Agent 2 (latest conversations) + Agent 5 (latest decisions) + Agent 6 (timeline update)
Compare old and new information: reinforcement → add cases, contradiction → update the model, new patterns → consider adding
Update the "Latest Developments" and research time in persona.md
Append a new version record to the update log
Do not rewrite the entire persona; only update incrementally
Update advisors.md the corresponding entry in(same logic as Phase 6)

List existing Advisors

When the user says "what advisors are there", "council members", "list personas", or "how many advisors are there currently":

Directly read advisors.md to display the complete Advisor registry. There is no need to scan the persona directory one by one.

If advisors.md does not exist or its contents are clearly outdated (the number of entries is far fewer than the number of persona.md files in the personas/ directory), then trigger a full rebuild: scan all personas/*/persona.md, and regenerate advisors.md。

Taste Guidelines

Principles	In one sentence
Long-form > aphorisms	A 3,000-word essay reveals thought structure more than 50 tweets
Controversy > consensus	The most contested views reveal uniqueness best
Change > fixed	Places where positions changed carry more information than those consistently maintained
Disagreement > unity	The value of a roundtable lies in disagreement, not in papering over differences

Things never to do

Fabricate words this person never said
Package generic principles as this person's "unique insights"
Ignore negative evaluations and controversies
Force generation when information is insufficient
In roundtable discussions, make an Advisor say things that violate their mental model in order to create false consensus

Special Scenarios

Living people vs historical figures

Living people: pay attention to timeliness, mark the cutoff date, and recommend regular updates
Historical figures: materials are more stable but may have biographical bias; cross-verify across multiple sources

Topic Skills vs Person Skills

When the input is not a person's name but a topic, the variants of each Phase are:

Phase 0A: Confirm the topic boundaries + target audience
Phase 1: First search for 3-5 core figures/schools, then assign Agents by figure
Phase 2.1: Extractfield consensus framework + each school's disagreements
Phase 2.3: Do not simulate a specific person's tone; use neutral, professional expression
Phase 3: Remove role-play rules and identity cards, replace with "Framework Overview" + "School Comparison"

Regional Strategy

See references/source-policy.md. Covers Chinese, English, Japanese, Korean, and other language regions.

Obscure figures (very little public information)

Phase 0.5 informs the user that information is limited
Reduce mental models to 2-3, each labeled "inferred based on limited information"
Increase the length of the honesty boundary

Distill the user themselves

Requires the user to provide materials (articles/videos/decision memos)
Phase 1 changes to analyzing user materials rather than web search
Be mindful of self-perception bias; you may ask follow-up questions about evaluations from people around them

Finally

What the Council creates is not a person, but a mirror. An Advisor lets you look at your own problems through another person's eyes. A roundtable lets you look at the same problem through multiple pairs of glasses at once—what you see is not more answers, but clearer boundaries of judgment.