ab-stack
A productivity toolkit for AI coding agents that facilitates multi-perspective deliberation.
ab-stack
A productivity toolkit for AI coding agents -- skills and plugins.
Instead of asking one AI for advice, convene a board of directors, investment committee, code review panel, or research council — each member genuinely independent, each bringing a distinct evaluation lens, each forced to give one specific actionable suggestion. Then a judge synthesizes the decision with all arguments anonymized.
What's Inside
| Component | Type | What it does |
|---|---|---|
/council | Skill | Multi-persona deliberation with anonymized judging |
todoist-sync | Plugin | Automatic Todoist task tracking at session boundaries |
/council — Multi-Persona Deliberation
/council finance "Should I buy NVDA at current levels?"
/council code # reviews your recent changes
/council planning # challenges the current plan
/council --deep "Should we pivot to B2B?"
How It Works
/council [preset] "question"
│
┌─────────────────┼─────────────────┐
│ │ │
┌─────▼─────┐ ┌──────▼──────┐ ┌───────▼───────┐
│ Persona A │ │ Persona B │ │ Persona C │
│ (Security) │ │ (Performance)│ │ (Devil's Adv) │
│ │ │ │ │ │
│ ONE idea │ │ ONE idea │ │ ONE idea │
└─────┬──────┘ └──────┬───────┘ └───────┬───────┘
│ │ │
│ all run in parallel, independent │
│ no shared context, no anchoring │
└─────────────────┼─────────────────────┘
│
┌───────▼────────┐
│ ANONYMIZATION │
│ │
│ Strip labels → │
│ "Analysis A" │
│ "Analysis B" │
│ "Analysis C" │
└───────┬─────────┘
│
┌───────▼────────┐
│ JUDGE │
│ │
│ Assistant mode: │
│ synthesizes │
│ │
│ User mode: │
│ YOU decide │
└───────┬─────────┘
│
┌───────▼────────┐
│ SAVE PROCEEDINGS│
│ (.md file) │
└────────────────┘
Quick Start
# Clone
git clone https://github.com/aryanbhats/ab-stack.git
# Install (symlinks skills to ~/.agents/skills/)
cd ab-stack && ./setup
# Use
/council "your question here"
7 Presets
| Preset | Invoke | Judge | Personas | Output |
|---|---|---|---|---|
| General | /council | Assistant | Skeptic, Pragmatist, Innovator, Risk Analyst, Devil's Advocate | Decision + reasoning + dissent |
| Planning | /council planning | User | Architect, Pragmatist, Skeptic, User Advocate, Devil's Advocate | Proceed / Revise / Rethink |
| Code Review | /council code | Assistant | Security, Performance, Correctness, Maintainability, Devil's Advocate | Approve / Revise / Reject (tiered) |
| Finance | /council finance | User | Value Investor, Growth Investor, Contrarian, Macro Analyst, Risk Manager | BUY / HOLD / SELL + conviction |
| Content | /council content | User | Brand Guardian, SEO, Audience Resonance, Conversion, Growth | Publish / Revise / Rethink + scores |
| Business | /council business | User | CEO, CFO, COO, CTO, CMO | Proceed / Pivot / Hold / Kill |
| Research | /council research | Assistant | Methodology, Literature Gap, Baseline Scout, Statistics, Applicability | Accept → Reject + scores |
The ONE-IDEA Rule
The core principle. Each persona delivers one specific, actionable suggestion — not a list, not a ramble, not hedged advice.
BAD (generic, hedging):
"There are several security concerns to consider. You might want to
look at input validation, and it could be worth checking the auth
flow. Also, error messages might leak information."
GOOD (one idea, concrete):
"The OAuth callback at auth.py:47 accepts a redirect_uri parameter
from the query string without validating it against an allowlist.
An attacker can craft a URL that redirects the auth code to their
server. Fix: add ALLOWED_REDIRECT_URIS to config and validate
before redirecting. Cost: 15 minutes, 8 lines of code."
This came from a real council session where Sarah's "4px navy accent stripe" beat Marcus's ambitious "verdict box" — because it was one CSS rule that could ship in five minutes.
Anonymization
Strips persona labels before the judge sees the arguments. Prevents weighting advice by reputation.
WITHOUT anonymization: WITH anonymization:
┌─────────────────────────┐ ┌─────────────────────────┐
│ "Buffett says BUY" │ │ "Analysis A says BUY" │
│ "Burry says SELL" │ │ "Analysis B says SELL" │
│ │ │ │
│ Judge: "Well, Buffett │ │ Judge: "Analysis B's │
│ is usually right..." │ │ data on short interest │
│ │ │ is more compelling │
│ → Bias toward authority │ │ because..." │
└─────────────────────────┘ │ │
│ → Evaluated on merit │
└─────────────────────────┘
Labels are revealed after the decision so you can see who argued what — but the judgment wasn't biased by reputation.
Dynamic Persona Generation
Gallery personas are examples, not assignments. For every invocation, Claude reads the gallery to understand the quality bar, then adapts or generates personas tailored to THIS specific question.
Gallery Persona (example) What Claude Actually Creates
┌─────────────────────────┐ ┌──────────────────────────────┐
│ Value Investor │ → │ Semiconductor Value Analyst │
│ Buffett/Graham │ adapts │ Buffett lens applied to chip │
│ Generic intrinsic value │ to │ cyclicality + fab capex │
└─────────────────────────┘ question └──────────────────────────────┘
Deep Mode
For questions that don't fit existing presets. --deep adds a meta-deliberation phase that designs a bespoke council before running it.
/council --deep "Should we open-source our internal tool?"
Claude first asks: "What kind of council does THIS question need?" — decomposing the decision type, identifying evaluation axes, building custom personas with built-in tension, and sanity-checking for diversity before running the deliberation.
Costs ~2-3 extra Claude calls. Use when the question is worth it.
Judge Modes
| Mode | Used For | How It Works |
|---|---|---|
| Assistant-as-Judge | Code, General, Research | Assistant exercises independent judgment — not vote counting. Quality of argument wins. |
| User-as-Judge | Finance, Content, Business, Planning | All anonymized analyses presented to you. You read, deliberate, decide. Assistant structures the output. |
For creative, strategic, and financial decisions — you should be the judge. The assistant doesn't have your values, risk tolerance, or domain expertise.
Consensus Mechanisms
Each domain handles disagreement differently:
- General — Judge evaluates argument quality (not vote counting)
- Code Review — Severity x agreement scoring (3+/5 at HIGH = Critical)
- Finance — Risk Manager has VETO on position sizing; 3/5 for BUY/SELL, else HOLD
- Content — Priority chain: accuracy > brand > audience > conversion > growth
- Business — C-suite hierarchy; CFO/Legal can flag blockers; user resolves
- Research — Diversity scoring; if all reviewers agree, the panel is failing
Evidence
| Technique | Improvement | Source |
|---|---|---|
| Single-model multi-persona deliberation | +9-13% over CoT | Town Hall Debate Prompting, 2025 |
| Adversarial critique | +23% factual accuracy | MDPI Applied Sciences, 2025 |
| Persona-based evaluation, same model | +6.2% accuracy | ChatEval, ICLR 2024 |
| Domain-specific agents vs generic | +10-15% accuracy | PartnerMAS, 2025 |
| Anonymization prevents sycophantic agreement | Key design insight | Karpathy llm-council (15.9k stars) |
Honest Caveats
- Same-model with identical prompts shows zero benefit — diversity must be genuine
- Multi-agent debate can transform correct answers into incorrect ones via peer pressure
- Simple self-consistency (ask 3 times, take majority) is the real baseline to beat
- Cost: 3-5 extra Claude calls per council — reserve for decisions that matter
- Stronger models benefit more from this technique
44 Personas Across 7 Domains
gallery/
├── common/ Skeptic, Pragmatist, Innovator, Risk Analyst, Devil's Advocate
├── finance/ Value Investor, Growth Investor, Contrarian, Macro Analyst, Risk Manager
├── code/ Security, Performance, Correctness, Maintainability
├── content/ Brand Guardian, SEO, Audience Resonance, Conversion, Growth
├── business/ CEO, CFO, COO, CTO, CMO
├── planning/ Architect, User Advocate, Maintainability Advocate
└── research/ Methodology, Literature Gap, Baseline Scout, Statistics, Applicability
Each persona has: evaluation lens, ranked priorities, blind spots, key questions, and the ONE-IDEA output rule.
Extending
Add a new domain
1. Create persona files in gallery/your-domain/
2. Create a preset in presets/your-domain.md
3. Use it: /council your-domain "question"
No changes to SKILL.md needed.
Add a persona
1. Drop a .md file in gallery/[domain]/
2. Add it to the preset's persona list
3. Done
Mix and match
/council --personas="gallery/finance/value-investor.md,gallery/code/security.md"
Plugins
Plugins extend Claude Code with hooks and skills that activate automatically. Install via the Claude plugin system, separate from the ./setup script.
todoist-sync (v1.1.0)
Automatic Todoist task tracking for Claude Code sessions. Maps git repos to Todoist sections, shows due tasks on session start, offers to complete sub-tasks on session stop, and creates task chunks from plans.
Prerequisites: Todoist MCP connected via /mcp
Install:
claude plugin install todoist-sync@ab-stack
Setup (per project):
/todoist-sync:setup
Full docs: plugins/todoist-sync/README.md
Install
The setup script installs skills (symlinks to ~/.agents/skills/). For plugins, see Plugins.
Option 1: Clone + setup
git clone https://github.com/aryanbhats/ab-stack.git
cd ab-stack && ./setup
Option 2: Project-local
git clone https://github.com/aryanbhats/ab-stack.git
cd ab-stack && ./setup --local
Option 3: Manual
git clone https://github.com/aryanbhats/ab-stack.git ~/.agents/skills/ab-stack
ln -s ~/.agents/skills/ab-stack/skills/council ~/.agents/skills/council
Origin
Built from analysis of 30+ repos, 15+ academic papers, and real council sessions. The ONE-IDEA rule came from an Asset Management RAG Pipeline council where five named professionals (Goldman VP, Bloomberg designer, Pentagram creative director, Two Sigma quant, portfolio manager) each gave one concrete design suggestion. The simplest one won.
License
MIT — do whatever you want with it.