Nehad Al Ghadri on stage, holding a microphone

The Board of Directors

A command-line tool I built that turns any real decision into a researched debate between several AI advisors. It writes a decision file, records the discussion as a short podcast (Kokoro text-to-speech), and checks back with me at 30, 90, and 180 days so I can see how the decision held up.

Persona-modeled advisors
12
Deep-mode decision cycle
~7m
Post-mortems at 30/90/180d
3

Claude CLI, Python, Kokoro TTS, ffmpeg, launchd, Bash

Consulting is ninety percent judgment.

Judgment only sharpens through structured reps, adversarial framing, prediction plus outcome tracking, and bias recognition. Most professionals stumble through all four accidentally, if at all. I built The Board to force all four deliberately.

What I built

A slash-command system, installed globally, that turns /board "your question" into a full decision pass.

Classify
The topic gets tagged with one of eight decision types: pricing, career, technical, relational, content, strategic, health, financial-personal.
Research
A single Claude plus web-search call pulls the last sixty days of relevant news, trends, and expert takes. The context gets injected into every advisor prompt and saved as “Recent web context” in the decision file, so the panel is reasoning about the world as it is this week, not as it was in the training data.
Framework pass
One or two decision frameworks get applied to the topic. Regret minimization, inversion, 10/10/10, Type 1 versus Type 2 reversibility, pre-mortem, Playing to Win cascade. Each framework produces a structured angle the advisors can argue about.
Advisors
A panel of twelve persona-modeled advisors, each with decision-type tags that control which ones get called in for which kind of question. Quick mode calls three advisors and deep mode calls five. Each writes a steel-man, then an actual take in their own voice, then reads the others and pushes back.
Synthesis
The synthesis shows where the advisors agree, where they fight, and what the real upstream question is, with a consensus meter from 0 to 5. A bias check scans the synthesis for recurring patterns against a personal bias log.
Podcast
Kokoro TTS synthesizes the whole dialogue, in the advisors’ own voices, to a proper MP3 I can listen to in the car instead of re-reading the transcript.

The accountability loop

This is the feature the rest depends on. Every decision logged via /decide schedules follow-up pings at 30, 90, and 180 days. A launchd job fires a macOS notification when a post-mortem is due. /postmortem asks five fixed questions comparing prediction to reality. After ten post-mortems, /biases surfaces my own recurring patterns: timeline optimism, scope creep, and status-quo bias. Every quarter, /review writes a “state of my decision-making” report covering advisor track record, framework usage, bias frequency, and confidence calibration.

Post-mortems are eighty percent of the value.

Without them, The Board is confirmation bias in a nice voice.

Modes, cost, and privacy

The decision modes matter less than the discipline of writing the decision down before the outcome is known. But the modes help. oracle picks one random advisor you don’t see coming. coach flips the panel into Socratic mode, asking questions instead of giving answers. premortem asks every matched advisor to imagine the decision has catastrophically failed in six months and write the post-mortem from that future. opposite makes every advisor argue the inverse of their usual stance to break the groove.

There is no separate API bill, because everything runs through the Claude CLI under my Max subscription. The decision files and the audio stay on my Mac. Sensitive decisions route to a separate folder with a restricted permission mode and stay out of the professional knowledge graph. It’s the first tool I’ve built that I expect to still be using in ten years, because each decision and its post-mortem add to a record that a one-shot model call never has.

All work