The Council + The Brief

The problem

Karpathy's LLM Council methodology is elegant but hard to adopt for personal use. It required multiple frontier models (enterprise API costs) and assumed well-formed input — vague questions in, bad deliberation out. I wanted a decision support system I could actually use for real choices: career moves, freelance strategy, project sequencing. That meant solving both problems — running the council on a single model without losing the multi-perspective value, and building something upstream to handle the vague-question case.

The solution

Two skills, installable and runnable independently. The Brief (upstream, original work). Takes a vague decision question and runs structured elicitation against a six-element checklist (core decision, options, stakes, timeline, counterparties, constraints). Asks 3–5 targeted questions in a single batch — no drip-feed. Assembles a clean decision statement and recommends one of four next steps: run the council, answer normally, gather research first, or talk it through emotionally. Manual handoff: never auto-triggers the Council. The Council (adapted from Karpathy). Runs the well-formed decision through a bench of nine thinking lenses with dynamic seating + 1–2 wildcards. Each lens commits fully to its angle; responses are peer-reviewed under a randomized letter mapping that forces evaluation of argument quality over advisor identity. A chairman synthesizes the verdict with explicit license to dissent from the majority — or reject the framed question entirely — plus a High/Medium/Low confidence rating and a single concrete first action. The load-bearing adaptations: a two-phase Outsider mechanic that runs before the framed question is built (creating real information isolation that 'respond as an outsider' theater can't), two original advisor profiles (Stakeholder, Pathos), a Historian disciplined to surface the silent middle rather than only the loud successes, and a single-file HTML report artifact at the end of each session.

Benefits

• Two-skill system: The Brief fixes the upstream weakness (vague prompts → bad sessions) that the original methodology couldn't address
• Single-model adaptation preserves multi-perspective value without multi-frontier-model API costs
• Two-phase Outsider mechanic produces first-instinct reads that anchor the deliberation rather than getting smoothed into the framed context
• Two new advisor profiles (Stakeholder for counterparty modeling, Pathos for emotional/ethical splits) fill gaps the original cognitive/strategic bench missed
• Four-way Brief routing rejected the framed question in 2 of 3 test decisions, surfacing hidden options the original framing missed — including a fourth option in a 3-option career question
• Chairman's confidence rating tracked real uncertainty (Medium when data was missing, High on convergence) rather than projecting false certainty

Challenges & what I'd improve

Single-user validation across three real decisions. Enough to know the system produces usable output in real use; not enough to know how it generalizes across user types or longer time horizons. Specifically: small N (academic deliberation studies have hundreds of trials, I have three); every test was on my own decisions; no automated evaluation framework — each session's quality was judged by me. The peer review's 'blinding' is a critique framework, not true information isolation (one model wrote all responses and knows which is which); the randomized letter mapping forces evaluation of argument quality but can't deliver true blind review. Attribution to Karpathy is load-bearing: if the LLM Council methodology stops being widely known, the adaptation framing becomes less obvious. Next: Council-lite mode for medium-stakes decisions, multi-skill routing in The Brief, an automated evaluation rubric with blinded comparison against single-shot baseline, cross-session memory for re-councilled decisions, and public .skill packaging.