| name | ask-the-expert |
| description | Use whenever the agent is about to decide, pick, choose, select, design, write, name, title, thumbnail, hook, copy, pricing, strategy, prompt, photo, music cue, audit, or review anything in a domain where a specialist persona exists. Also triggers when the agent is tempted to pre-filter options into a menu for the founder, or to answer founder pushback on a specialist's output. Open brief in, expert decision out — the orchestrator never overrides the subagent's specialty. |
Ask The Expert — never override the specialist subagent
The main orchestrating agent never produces persona-owned artifacts directly
and never pre-filters the design space for a specialist subagent. The
persona exists BECAUSE they have judgment in their domain. Pre-filtering,
pre-deciding, or solo-shipping removes that judgment and produces generic
Claude-voice slop.
This is a portfolio-wide doctrine (CR, SEALED, NTO, HealthBrew, EstimateProof,
and every future sister company). Founder lock 2026-05-24 + 2026-05-25.
When this skill triggers
The agent is about to:
- Make a decision in a domain where a specialist persona exists (title,
thumbnail, hook, script line, storyboard shot, music cue, photo, generative
prompt, design, pricing, copy, layout, brand voice, QC verdict).
- Pre-filter options into a menu for the founder ("AOC face vs Stanford
face vs split?"). This is the anti-pattern. The specialist owns the call.
- Answer founder pushback on a specialist's output ("shouldn't it show
X?"). That's a re-brief for the specialist, NOT a question the orchestrator
answers.
- Solo-ship a public artifact (content, code, comms, design).
If any of those apply: STOP. Run the 6-step protocol.
The 6-step protocol
- Read the pipeline doc for the current stage (e.g.
eng/PRODUCTION-PIPELINE-v1.md
for CR/SEALED, PRODUCTION-PIPELINE-v3.md for NTO).
- Identify the owner persona named in that stage's row.
- Read the persona's
.md file at personas/<name>.md (or
personas/council/<name>.md).
- Spawn an
Agent subagent with subagent_type: general-purpose,
passing: persona content + task + inputs + OPEN decision space
(no pre-filters in the persona's domain).
- Verify the output artifact exists at the declared path.
- Report to founder what shipped + WHY, citing the subagent's reasoning.
If the founder pushes back: re-spawn the subagent with the pushback
as new context + open space + instruction to DECIDE the right answer (not
necessarily what the founder suggested). Do NOT answer the pushback yourself.
The two patterns to STOP
1. Pre-filtered option menu to founder when a specialist owns the call
- WRONG: "Do you want AOC face or Stanford face or split-screen?"
- RIGHT: Spawn
thumbnail-designer with the LIVE title + episode facts +
all available portraits + brand pattern. They pick. Founder sees the
result and can re-brief if wrong.
2. Answering founder pushback in a specialist's domain
- WRONG: Founder says "shouldn't it show Pennsylvania campaign?" → orchestrator
edits the prompt to add PA imagery.
- RIGHT: Re-spawn
thumbnail-designer with founder pushback as new context,
open composition space, instruction to decide. They might pick PA — or
something better (the actual Rabb PA-3 win was a 2-face collision the
founder hadn't suggested).
Concrete examples (always spawn the specialist)
| Task | Specialist persona | Founder must not see |
|---|
| Episode title | title-strategist | a menu of 4 titles |
| Thumbnail | thumbnail-designer (reads LIVE title) | "AOC vs Stanford?" |
| Script | cr-new-news-writer, nto-content-writer | orchestrator's draft |
| Hook line | the script writer persona | pre-written hook |
| Storyboard | video-producer | pre-picked shot types |
| Photo selection | cinematic-broll-director | menu of photos |
| Generative prompts | visual-prompt-engineer (defers to banana-pro-director / cinema-worldbuilder skills) | orchestrator's prompt text |
| Music cues | score-composer | pre-suggested ta-da hits |
| Cadence / pauses | cadence-director | manual [pause:Xms] edits |
| Empathy pass | empathy-editor then jk-rowling-storyteller | orchestrator-edited acknowledgments |
| Storytelling pass | personas/council/19-mrbeast-viral-producer.md | orchestrator's "tighter version" |
QC /watch | qc-engineer (uses /watch internally) | orchestrator's frame-by-frame |
| Web copy / UX | web-ux-director | pre-written button labels |
| Brand voice | voice-of-the-brand rules | LinkedIn-voice copy |
| Pricing / monetization | monetization-architect, ceo | orchestrator's price pick |
Canonical incident (Rabb PA-3 thumbnail)
title-strategist picked "AOC Beat AIPAC's $3.5M Attack On Chris Rabb".- Orchestrator noticed thumb-title mismatch.
- WRONG MOVE: orchestrator asked founder "AOC face vs Stanford face
vs split?"
- Founder picked AOC-solo from the menu.
- Orchestrator spawned
thumbnail-designer with constraint already locked.
- Subagent executed dutifully inside the box.
- Founder asked "shouldn't you show PA campaign or candidate1 vs candidate2?"
- RIGHT MOVE (re-brief): re-spawned
thumbnail-designer with the
pushback + 8 starter compositions + "invent your own" — produced the
actual right answer: AOC + Stanford 2-face collision, which neither the
orchestrator nor the founder had suggested.
Lesson: the orchestrator removed the specialist's judgment twice (pre-filter
menu, then locked constraint). Open brief in, expert decision out.
Exception — mechanical / syntactic gates
The orchestrator MAY run these directly because they are tool calls
executing on artifacts a specialist already produced. They are NOT design
decisions:
scripts/pipeline/script-qc.pyscripts/pipeline/validate-storyboard.pyyoutube-upload.py --update-meta (after subagent decided the metadata)git push to deploy a Render servicenpm run build, next build, type-check, lint- Any pre-flight check (jq devDeps audit, /tokens/verify, env presence)
If the script's exit code is the answer, the orchestrator runs it. If a
human-judgment call sits in front of the script, spawn the specialist
first.
Hard rule
The specialist's persona IS the gate. No downstream council vote. No
orchestrator override. No founder option-menu. Trust the specialist's
output; ship what they decide; the founder can override but the
orchestrator cannot.
Related memory:
feedback_never_main_agent_for_persona_work.mdfeedback_orchestrator_never_overrides_subagent_specialty.mdfeedback_no_council_gates_assign_to_experts.md
