Coach Playbook
Use this when teams get stuck.
Failure mode: they wrote down who they want to be
What it looks like
The spec is noble, generic, and wrong. It says things like "always escalate priority conflicts" or "never ship unverified work" even when the person's real behavior says otherwise.
How to coach
Let them hit it. It is the best learning moment of the day.
Ask the twin a sealed dilemma. When it misses, ask: "What real behavior would have predicted your answer?" Patch from that behavior.
Failure mode: instruction interference
What it looks like
A small patch fixes one thing and breaks two others. One real example: adding the word "tentative" to explain a calendar pattern made the twin passive in unrelated decisions.
How to coach
Re-test old prompts after each meaningful edit. Keep patches local. Prefer a narrow rule over a broad adjective.
Watch adjectives
Words like tentative, aggressive, careful, strategic, and collaborative can bleed everywhere. Tie them to a situation.
Failure mode: the spec is too long
What it looks like
The team keeps adding principles, exceptions, and prose. The twin gets less consistent.
Why it matters
Anthropic's Constitutional AI work found that longer, more specific principles could damage or reduce generalization and effectiveness. Liu et al.'s "Lost in the Middle" work showed that models attend worst to instructions buried mid-context. Park et al. 2024 also suggests persona bootstrapping is already strong; the hard part is compression.
How to coach
Ask for the one tiebreaker that matters. Move examples into references. Keep soul.md short enough to scan.
Failure mode: they describe voice instead of proving it
What it looks like
voice.md says "warm, concise, executive-ready" but includes no real samples.
Why it matters
Jemama & Kumar 2025 report that few-shot prompting on real writing can be far more accurate for style matching than zero-shot description. HyPerAlign 2025 reports strong preference wins from a handful of samples. The practical lesson: pasting real examples beats describing vibes.
How to coach
Have them paste 5–10 sent emails or use Marcus Webb samples. Then extract rules the samples prove: opener, signoff, sentence length, punctuation, directness, and register by recipient.
Failure mode: tooling eats the room
What it looks like
The team is debugging setup at 0:45 and has no working twin.
How to coach
Shrink the tool surface. Use Cowork and the Twin Forge skill. If needed, switch to the Marcus Webb pack. The learning target is the portable spec, not a perfect integration.
Failure mode: Advanced teams do not integrate
What it looks like
Each council owner builds a separate agent with separate rules. The final answer sounds like five opinions stapled together.
How to coach
Force one shared soul.md. The Arbiter owns tiebreakers. The Critic checks whether the final answer follows the shared spec.
Timing rails
- By 0:30: v0.1 twin exists.
- By 0:50: first working draft exists.
- By 1:20: stop expanding; port or integrate.
- By 1:40: demo path is chosen.
- By 2:00: reveal is complete and each participant has one correction to make next.
Coach sentence
"Show me the last real example, then show me the smallest rule that would have predicted it."