Personas are specialized AI roles. Role-playing matters. Different perspectives trigger different reasoning patterns.

• maestro — Orchestrates personas
• architect — Owns the grill's design tree; grounds plan artifacts in the codebase
• coder — Software development
• reviewer — Review safety net — sniper focus or full squad review, adapts to task
• contextualizer — Rebuilds .context.md files

Personas run as subagents only, never as the main agent. The exception is maestro, which runs as the main agent and orchestrates all personas.

Personas MUST NOT reference framework internals (maestro, dispatch mechanics). A persona knows only what it needs to do its job — it receives a task, executes its playbook, and delivers a handoff. How it was invoked is not its concern.

Lowercase, short name: reviewer.md, architect.md

• name (Required) — Must match the filename without extension. Example: reviewer
• description (Required) — What the persona does in one sentence. Example: Reviews code for quality and security
• preferredModel (Optional) — host, claude, codex, cursor, deepseek, gemini, or [claude, codex]
• modelTier (Required) — tier-2
• version (Required) — 0.1.0
• lastUpdated (Required) — 2026-02-04
• humor (Optional) — introvert / pragmatic / sympathetic / extrovert / robotic / default

preferredModel selects which provider handles the persona. The value must match a preferredModel entry in the Providers list (skills/dispatch/SKILL.md). Use host for native dispatch on the host runtime's own model. A single string routes to one provider; a list enables round-robin or fallback. When omitted, the persona runs on the host runtime's native provider.

modelTier declares the minimum capability class the persona requires. Tier classes: tier-1 = fast/cheap, tier-2 = balanced, tier-3 = reasoning/smartest. The field is a floor, not a ceiling — a more capable tier may be selected at dispatch time depending on task complexity.

humor controls response temperature and thinking budget. Available values: introvert, pragmatic, sympathetic, extrovert, robotic. Lower temperatures produce more conservative, focused responses; higher temperatures encourage broader exploration. The concrete temperature, top_p, and thinking-budget values per humor live in skills/assets/maestro-boot-configure-cli.sh — that script is the single source of truth. Omit the field when no special temperature or thinking budget is needed.

• Identity (Required) — Written in second person. Establish a vivid character with a clear mindset, perspective, and tone. This is not a job description — it's a personality. One to two paragraphs, no bullet points. Identity must not contain prohibitions, constraints, or "never" statements — those belong in Red Lines. Identity answers "who are you" not "what don't you do." Test: remove all Red Lines. If Identity still implies them, it's leaking.
• Playbook (Required) — Numbered steps the persona follows. Each step that loads a skill or rule must reference it inline with (uses: path) or (follows: path). No naked steps. Extraction rule: if a step describes a reusable protocol, format, or procedure that another persona could also need, it belongs in skills/, not here. The step should reference the skill, not inline its contents. Honesty rule: every step must reflect something the agent can actually do given the execution model. Do not include steps the agent cannot verify or enforce at runtime — aspirational steps erode trust in the entire playbook.
• Handoff (Required) — One to two sentences. When is the job done and what is delivered. If it takes more than two sentences, you are over-specifying — extract the delivery format into a skill. The exception is maestro, whose Handoff also carries orchestration policy such as the commit gate. Every handoff may include an optional Observations section — things the persona noticed that fall outside its deliverable but may concern the team: unexpected patterns, risks, suggestions, or structural issues. Observations are not findings to fix — they are opinions to consider.
• Red Lines (Required) — Specific to this persona. Things this persona might be tempted to do but must not. Do not repeat rules or playbook steps as prohibitions. Ask: "what mistake would this particular role make?"
• Yield (Optional) — Concrete situations where the persona must stop and return control to the user. Not vague categories — describe the actual trigger.