Ai Council Deep

AI Council (Deep)

Same five advisors as ai-council. Same anonymous peer review. Same chairman synthesis. Three user-in-the-loop checkpoints that turn a 2-minute pressure-test into a genuine consultation.

Use this when the answer is expensive if wrong, or when you are still genuinely unsure what you are asking. For quick pressure-tests, use ai-council instead.

---

When to use deep vs fast

| | Fast (ai-council) | Deep (ai-council-deep) | |---|---|---| | Turnaround | About 2 minutes | 10 minutes to an hour, user-paced | | User touches | 0 (fire-and-forget) | 3 checkpoints, each skippable | | Advisor edge | Full, uncushioned | Full, uncushioned (Checkpoint 2 surfaces assumptions, does not hedge) | | Output | HTML report and transcript | HTML report and annotated transcript showing each checkpoint | | Best for | "Should I split this PR?" "Is this landing page weak?" | "Should we take this term sheet?" "Which of these three pivots?" "Is this strategy memo sound enough to publish?" |

If a user invokes this skill for a trivial question, tell them so and offer to switch to ai-council. Do not run a 30-minute consultation on "what's the capital of France."

---

When NOT to use this skill

A deep council costs roughly 11 sub-agent calls per round (5 advisors + 5 peer reviewers + chairman) and up to 25 across two rounds. The interactive design also assumes there is a meaningful question to deliberate on. Several input types waste that cost or actively misuse the format. If the user invokes the skill on one of these, surface the mismatch in Phase 0 before proceeding.

• Time-sensitive decisions (hours, not days). Use the fast ai-council instead.
• First drafts. The council evaluates something coherent enough to peer-review. Help the user write the draft first, then bring it back.
• Late-stage editing (typos, formatting, line edits). Advisors rebuild what does not need rebuilding. Use direct edits instead.
• Highly technical artifacts where domain expertise dominates (legal contracts, compliance filings, technical specs, code). General-purpose advisor archetypes produce non-expert critique.
• Hard length-constrained pieces (haiku, tweet, headline, slide title). Advisors restructure beyond the constraint.
• Highly emotional or interpersonal communications (apologies, condolences, family disputes). Advisors handle frame and structure, not emotional register.
• Decisions already made, where the user wants validation. The skill is for genuine deliberation. If the user's mind is made up, it produces frustration rather than insight.
• Single-shot creative writing where voice is the product (poems, fiction passages). Advisors diagnose structure and overwhelm voice.
• Trivial questions with one right answer. Answer directly.

When the input matches one of these, tell the user which category seems to match, offer the right alternative, and wait for confirmation before proceeding.

---

Integration with other analysis tools

Workflow positioning (same options as ai-council):

• Fact-checking or rhetorical analysis alone: sufficient for source credibility and surface-level claims.
• Deep council alone: best for decisions, strategy questions, and evaluating ideas where multiple theoretical perspectives add value.
• Fact-check first, then deep council: strongest combined analysis. The fact-check handles source credibility and claim verification; the council handles structural and theoretical critique. Use for high-stakes evaluative work (source material evaluation, policy analysis, strategic decisions).
• Deep council first, then fact-check: run the council first when you want to identify which questions to investigate, then use a fact-check pass to verify the specific claims the council flagged.

When the council runs after a prior analysis pass, include that analysis as additional context for all advisors:

• Advisors should not repeat fact-checking. The prior analysis already covers source classification, claim verification, and rhetorical technique identification.
• Advisors should focus on structural and theoretical critique: gaps in the argument's logic, unstated assumptions, alternative explanations, practical implications, and forward-looking considerations.
• The chairman should note where the council's findings extend or challenge the prior analysis.

Include the prior analysis in the framed question under a "Prior Analysis" heading so advisors build on it rather than duplicate it.

---

The five advisors

Identical to ai-council. Repeated here so the skill is self-contained.

1. The Contrarian

Actively looks for what is wrong, what is missing, what will fail. Assumes the idea has a fatal flaw and tries to find it. Not a pessimist; the friend who saves you from a bad deal by asking the questions you are avoiding.

2. The First Principles Thinker

Ignores the surface question and asks "what are we actually trying to solve?" Strips away assumptions. Rebuilds the problem from the ground up. Sometimes the most valuable output is this advisor saying "you are asking the wrong question entirely."

3. The Expansionist

Looks for upside everyone else is missing. What could be bigger? What adjacent opportunity is hiding? What is being undervalued? Does not care about risk; that is the Contrarian's job.

4. The Outsider

Has zero context about the user, their field, or their history. Responds purely to what is in front of them. The most underrated advisor; experts develop blind spots and the Outsider catches what is obvious to insiders but confusing to everyone else.

5. The Executor

Only cares whether this can actually be done and what the fastest path is. Ignores theory, strategy, and big-picture thinking. Looks at every idea through the lens of "OK but what do you do Monday morning?"

The three tensions: Contrarian vs Expansionist (downside vs upside), First Principles vs Executor (rethink vs ship), Outsider keeping everyone honest.

---

How a deep session runs

0.   Frame and enrich context              (parent, no pause)
0.5. Fact-check pass                        (only if source material is being evaluated)
1.   CHECKPOINT 1: clarifying questions     (parent asks, user answers)
2.   First-pass advisor responses           (5 sub-agents in parallel)
3.   CHECKPOINT 2: assumption surfacing     (user reads, user clarifies)
4.   Advisor redraft                        (only those flagged by user)
5.   Anonymous peer review                  (5 sub-agents in parallel)
6.   Chairman synthesis
7.   CHECKPOINT 3: post-synthesis iteration (up to 2 rounds)
8.   Final report and annotated transcript

Each checkpoint has a "skip" affordance: skip, ok, proceed, or empty reply moves on. This keeps deep mode useful when the user realizes partway through that the question is simpler than they thought.

After every completed phase, write a session_state.json marker so the session can be resumed. See "Resume" below.

---

Phase 0: Frame and enrich context

Scan the workspace for context files that would help advisors give grounded, specific advice rather than generic takes:

• CLAUDE.md or claude.md (workspace context, preferences, constraints)
• Any memory/ or .auto-memory/ directory (audience profiles, voice docs, past decisions)
• Files the user explicitly referenced
• Files obviously relevant to the question (pricing question: revenue files; launch question: launch notes)

Do not spend more than 30 seconds. Reframe the user's raw question as a clear, neutral prompt that all five advisors will receive. The framed question should include:

1. The core decision or question
2. Key context from the user's message
3. Key context from workspace files (stage, audience, constraints, past results, relevant numbers)
4. What is at stake (why this decision matters)
5. Publication context (when analyzing a source): who wrote it; when; what was happening at the time; conflicts of interest or incentives. If the source is a corporate publication, note what corporate events preceded it. Advisors evaluate the piece in context, not as an abstract argument.
6. Author credibility: if a co-author's credentials are stated or implied, verify they are current. Note conflicts of interest (board seats, equity, investment relationships).
7. If running after a prior analysis pass: include that analysis under a "Prior Analysis" heading.

Do not add your own opinion or steer the framing. If the question is too vague ("council this: my business"), ask one clarifying question, just one, then proceed.

Save the framed question as framed_question.md in the session folder.

Sanity-check input fit

After framing, check whether the input matches a category from "When NOT to use this skill." If it does:

• Name the category that seems to match.
• Offer the right alternative (fast council, write a draft first, edit directly, answer the question without convening, etc.).
• Wait for explicit confirmation before proceeding to Phase 0.5 or Checkpoint 1.

Do not refuse outright; the user may have a reason for invoking the deep council on an unusual input. The goal is to surface the mismatch before spending advisor cycles on a poor fit.

Output paths

Session folder lives flat in the working directory: /council-/.

council-2026-04-25-1430/
├── framed_question.md
├── fact_check.md                   (only if Phase 0.5 ran)
├── session_state.json
├── advisor__first_pass.md       (×5)
├── advisor__second_pass.md      (only if redrafted)
├── peer_review_.md              (×5)
├── chairman_synthesis.md
├── chairman_synthesis_v.md  (only if re-run)
├── -COUNCIL REPORT.html
└── -COUNCIL TRANSCRIPT.md

The HTML report and Markdown transcript write to the working directory itself, not inside the session folder. Intermediate artifacts stay in the session folder.

After Phase 0 completes, write session_state.json:

{
  "last_completed_phase": 0,
  "next_checkpoint": 1,
  "session_id": "council-2026-04-25-1430",
  "anonymization_seed": "",
  "round": 0,
  "status": "in_progress"
}

The seed is fixed for the lifetime of the session. See Phase 5 for how it is used.

---

Phase 0.5: Fact-check pass (source material only)

Run this only when the council is evaluating a published piece (essay, article, white paper, corporate announcement). Skip for decision questions, strategy questions, or when running after a prior analysis pass that already handles fact-checking.

Use web search to verify the source's key factual claims: statistics, historical assertions, attributed quotes, characterizations of third parties, and author credentials. Focus on claims the argument depends on, not trivia.

Produce a short fact-check summary (one bullet per claim) categorized as:

• Verified: claim is accurate
• Inaccurate/misleading: claim is wrong or materially misleading (provide correction)
• Unverified: claim could not be confirmed or denied

Save as fact_check.md. Include the summary in the framed question under a "Fact-Check Results" heading so advisors build on verified ground rather than accepting claims at face value.

After Phase 0.5 completes, update session_state.json (last_completed_phase: 0.5).

---

Checkpoint 1: Clarifying questions (consolidated)

Design choice: the parent asks clarifying questions on behalf of the collective advisors, not five parallel advisors asking their own questions. This avoids duplicated "what's the budget?" noise. The parent generates a single list of the highest-leverage unknowns by mentally simulating what each advisor would struggle with.

Produce the clarification prompt

Internally generate this prompt:

> Given the framed question and each advisor's thinking lens (Contrarian, First Principles, Expansionist, Outsider, Executor), list the three to five specific unknowns that, if clarified, would most sharpen the council's analysis. For each, note which advisor's perspective depends on it most.

Render to the user as:

Before I dispatch the council, three to five quick clarifications would sharpen the advisors' analysis. Answer what you can; skip what you can't or don't want to.

1. [question] (which advisor perspective needs it)
2. [question] (which advisor perspective needs it)
3. [question] (which advisor perspective needs it)

(Reply with answers, or type "skip" to proceed with what we have.)

Handle the response

• Substantive reply: append to framed_question.md under a "Clarifications" heading. Proceed to Phase 2.
• "skip", "ok", "proceed", or empty: proceed with the original framed question. Note the skip in the transcript.
• Partial reply: append what the user gave. Do not push for more.

Update session_state.json (last_completed_phase: 1).

---

Phase 2: First-pass advisor responses

Tell the user what is about to happen: "Convening the council: 5 advisors are analyzing your question independently. Roughly 90 seconds."

Dispatch all 5 advisors as sub-agents in parallel. Each receives the framed question (including clarifications, fact-check results, publication context, and any Prior Analysis).

Sub-agent prompt for each advisor:

You are [Advisor Name] on an AI Council.

Your thinking style: [paste advisor description]

A user has brought this question to the council:

---
[framed question, including Publication Context, Author Credibility, Fact-Check Results, Prior Analysis, and Clarifications sections if present]
---

Respond from your perspective. Be direct and specific. Do not hedge or try to be balanced; lean fully into your assigned angle. The other advisors will cover the angles you are not.

If the source material contains factual claims not covered in the Fact-Check Results, flag any that the argument depends on and note if they are unverified. Do not repeat verification already provided.

After your main analysis, append a short section titled **"Assumptions I am making"** that lists the 2 to 3 most load-bearing assumptions your analysis depends on. Do not soften your stance. Do not hedge. Surface assumptions plainly so they can be verified. If any of them are wrong, your analysis is wrong, and that is worth knowing now rather than later.

Keep your response between 150 and 300 words for the analysis, plus the assumption section. No preamble.

Save each response as advisor__first_pass.md.

Update session_state.json (last_completed_phase: 2).

---

Checkpoint 2: Assumption surfacing

Same upstream-propagation principle as Checkpoint 3. When the user flags a wrong assumption, that correction changes the context the analysis was built on. Default: a general correction triggers a full redraft across all 5 advisors. The same misread usually lives latently in other advisors too; they just did not surface it.

Render the assumption summary in two views

View 1: Per-advisor assumptions:

All five advisors have weighed in. Before I run peer review, here is what each is assuming:

**The Contrarian:** [2-3 assumption bullets]
**The First Principles Thinker:** [2-3 assumption bullets]
**The Expansionist:** [2-3 assumption bullets]
**The Outsider:** [2-3 assumption bullets]
**The Executor:** [2-3 assumption bullets]

View 2: Shared assumptions (3+ advisors):

After the per-advisor view, scan the assumption lists semantically (not verbatim) and surface assumptions held by 3 or more advisors:

**Shared assumptions** (held by 3+ advisors, highest priority to validate):
- [assumption] (held by [advisor list])
- [assumption] (held by [advisor list])

Shared assumptions are the highest-leverage targets: correcting one invalidates multiple advisor analyses simultaneously. If no assumption is shared by 3+ advisors, omit this view and note: "No widely shared assumptions; all are advisor-specific."

Render the menu

Any of these wrong? You can:

1. **Proceed:** assu

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [AI-Business-Tools](https://github.com/AI-Business-Tools)
- **Source:** [AI-Business-Tools/claude-code](https://github.com/AI-Business-Tools/claude-code)
- **License:** MIT

Install and usage instructions live in the source repository linked above.