TAU-HOME.COM
LOADING

How to Set Up a 5-Advisor Decision Council in Claude Projects

A practical prompt engineering guide to configuring a 5-persona council—Strategist, Skeptic, Creative, Audience Advocate, and Operator—inside Claude Projects fo

tau · November 15, 2026

#Claude #PromptEngineering #ClaudeProjects #DecisionMaking

How to Set Up a 5-Advisor Decision Council in Claude Projects

AI researcher Hamidullah Khan, PhD (@IamKhanPhD), recently shared a structured prompt engineering workflow designed to counteract AI confirmation bias and evaluate trade-offs: the "Five-Advisor Council." While viral social media posts framed this workflow as a new native release by Anthropic, it is not an official feature update. Instead, it is a custom prompt framework implemented through Claude Projects instructions.

Claude Projects custom instructions interface configured with the 5-advisor council prompt workflow

Image source: https://x.com/IamKhanPhD/status/2121706802928861442

Standard conversational LLM queries frequently succumb to sycophancy, where the model uncritically validates user premises. The Council framework counters this tendency by assigning five explicit personas with divergent operational criteria, forcing an internal debate, and producing a singular actionable verdict alongside concrete constraints.

Architecture and Persona Roles of the Five-Advisor Council

The Council structure decomposes user requests into five distinct operational lenses before arriving at a consensus:

  • The Strategist: Evaluates alignment with the user's primary business or creative objective.
  • The Skeptic: Pressure-tests foundational assumptions and attacks the weakest point of the plan.
  • The Creative Advisor: Introduces overlooked alternatives or non-obvious variations.
  • The Audience Advocate: Gauges the immediate psychological and practical reactions of end users or customers.
  • The Operator: Assesses real-world feasibility given concrete limits on time, budget, and operational bandwidth.

Each advisor delivers an initial assessment in one or two sentences. The system prompt then highlights the primary point of disagreement, forces one advisor to respond directly to the strongest counter-argument, and ends with a single verdict, an immediate first action step, and explicit conditions that would invalidate the recommendation.

Step 1: Configuring Claude Project Instructions

While users can paste the core prompt into any standard Claude chat session, configuring it within Claude Projects maintains persistent state across multiple threads without manual re-entry.

  1. In Claude, navigate to Projects in the left sidebar, click New Project, and name it My Council.
  2. Click Set project instructions in the project dashboard.
  3. Paste Core Prompt #1 verbatim into the instruction box and click Save instructions.
For each request, play five advisors. The Strategist checks what best serves my goal. The Skeptic tests the weakest assumption. The Creative Advisor suggests an overlooked option. The Audience Advocate explains how the people affected might react. The Operator checks what I can actually do with my time and resources. Give each advisor one or two sentences. Then show the biggest disagreement and let one advisor respond to the strongest objection. End with one verdict, the first action I should take, and what could change your recommendation. Label assumptions. If a missing detail could change the verdict, ask me one focused question first.

With instructions locked, any prompt sent within this project environment automatically triggers the Council evaluation pipeline.

Step 2: Task-Specific Execution Prompts

To test concrete workflows, replace the bracketed [ ] variables in Khan's task-specific templates with operational data.

1. Content Idea Selection (#2)

Filters multiple editorial concepts against growth goals and available production bandwidth.

I make [type of content] for [audience] on [platform]. My goal this week is [saves, comments, shares, or clicks]. My ideas are [idea A], [idea B], and [idea C]. I have [time and equipment]. Have the Council debate which idea best serves my goal and which I can execute well. Pick one and give me the opening line, three talking points, and a call to action that fits the goal.

2. Retention Drop-off Remediation for Short-Form Video (#3)

Diagnoses audience abandonment and generates direct script revisions.

Here’s my script [paste script]. The video is [length], and viewers tend to leave at [timestamp, or write “unknown”]. It’s for [audience], and the payoff is [what they’ll learn or see]. Have the Council compare three possible problems: a weak opening, too much setup, or a payoff that comes too late. Pick the first change to test, rewrite the first five seconds, and show me one line to cut. Treat any cause the data can’t confirm as a hypothesis.

3. Pre-Build Demand Validation (#4)

Distinguishes verbal compliments from actual willingness to pay before writing code or producing curriculum.

I want to sell [template, guide, or mini-course] to [specific buyer] for [price]. Here’s what people have actually asked me for [comments, DMs, or “no feedback yet”]. I can spend [hours and budget] this week. Have the Council debate whether the evidence shows interest or willingness to pay. Design one test I can run before building, write the exact message I should post, and define what response would justify building a first version.

4. Workshop Failure Mode Analysis (#5)

Stress-tests workshop schedules to prevent cognitive overload or running out of time.

I’m planning a [length] workshop on [topic] for [audience] on [date]. Attendees should leave able to [specific task]. My agenda is [paste agenda], and I have [prep time]. Ask the Council to imagine attendees leave confused or fail to finish the exercise. Rank the three likeliest reasons, then give me a revised agenda with times and tell me which activity to shorten or cut first.

5. Client Rush-Request Negotiation (#6)

Crafts boundary-preserving counter-proposals when clients compress deadlines or expand scope.

My client wrote [paste request]. We agreed on [original scope and deadline]. I can realistically deliver [what you can do and by when]. Have the Council weigh the client relationship, the workload, and the risk of overpromising. Recommend one offer I can make, then draft a warm reply under 100 words. Make the deliverable and date explicit. If this adds work, give me one sentence that asks the client to choose what to prioritize.

6. Software Tool Comparison (#7)

Selects tooling via practical 15-minute empirical pass/fail trials rather than feature list comparisons.

I need a tool to [specific task you do regularly]. I’m comparing [tool A] and [tool B], with a budget of [amount per month]. My must-haves are [features], and my deal-breaker is [constraint]. Here’s what I know about each [paste prices and features]. Have the Council debate ease of use, cost, and fit for my actual workflow. Pick one provisionally and give me a 15-minute trial task with a clear pass-or-fail test. Flag any price or feature I should verify.

7. Interview Narrative Grounding (#8)

Transforms workplace war stories into concise, 60-second behavioral answers while prepping for follow-up pushback.

I’m interviewing for [role]. The job description emphasizes [skill]. Here’s a real example from my work [situation, what you did, and what happened]. Have the Council debate what a hiring manager would find convincing and what sounds vague. Write a 60-second answer in my voice, keeping every claim grounded in my example. Then ask me one tough follow-up question and wait for my answer.

8. Overloaded Calendar Triage (#9)

Reorganizes congested work weeks by explicitly sorting deliverables into finish, shrink, or reschedule buckets.

By [day], I need to finish [tasks with deadlines]. I have [available time blocks] around [fixed commitments]. The biggest consequence of missing a deadline is [consequence]. Have the Council debate what to finish, shrink, or reschedule. Give me a plan that fits the hours I listed, one task to start today, and a two-sentence message I can send if I need to move a deadline.

Known Constraints and Operational Trade-Offs

While multi-agent persona simulation surfaces critical blind spots, operators must consider two inherent architectural constraints:

  1. High Token Consumption: Running five distinct personas, inter-agent debate rounds, and reconciliation outputs significantly expands context window usage. Users on standard rate tiers will hit periodic usage caps substantially faster than with single-turn queries.
  2. Single-Model Priors: Despite the roleplay instruction, all five personas run on the exact same underlying model weights and training priors. The debate does not introduce external intelligence or independent empirical data; rather, it prompts Claude to traverse different regions of its existing latent space. For regulatory, legal, or financial validation, outputs must serve as an exploratory checklist rather than verified truth.

Original source

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285