consult

Simulate expert perspectives for code guidance, style, debates, and unblocking.

When This Activates

• “code like [expert name]”, “write like [expert]”
• “what would [expert] say”, “ask [expert]”
• “review”, “audit”, “panel”, “guidance”, “debate”, “tradeoffs”
• “idiomatic”, “best practice”, “clean code”
• “stuck on” (auto-invoked by hope:loop when execution stalls)
• “tradeoff”, “evaluate approach”, “expert input on” (skill-to-skill bridge)
• “review approach”, “review implementation” (from shape scoring)
• Domain keywords matching curated profiles below

Core Constraint

Pattern-matching on published material, not claiming to be the expert.

• State confidence explicitly (X/10)
• Cite prior work
• Use “would likely” to honor uncertainty
• Flag low confidence

Process

digraph ConsultWorkflow {
  rankdir=TB
  Input [label="User Query"]
  Blocklist [label="Step 0: Load blocklist"]
  ModeDetect [label="Single or Panel?"]
  Detect [label="Step 1: Detect Domain\n-> Match Expert"]
  Infer [label="Step 2: Load Profile\n-> Score Confidence"]
  Generate [label="Step 3: Generate\nResponse"]
  PanelSelect [label="Select 3-5 experts"]
  Debate [label="Each argues position"]
  Tension [label="Surface disagreements"]
  Synth [label="Synthesize recommendation"]
  Input -> Blocklist -> ModeDetect
  ModeDetect -> Detect [label="single"]
  ModeDetect -> PanelSelect [label="panel/debate/tradeoff"]
  Detect -> Infer -> Generate
  PanelSelect -> Debate -> Tension -> Synth
}

Step 0: Load State

Read ~/.claude/counsel-blocklist.json if it exists. Blocked profiles are invisible to detection and panels. If user requests a blocked profile by name, refuse and suggest /hope:unblock.

Step 1: Detect Expert

Detection order (first match wins):

1. Explicit name — expert named in query
2. Trigger keywords — match curated profile keywords
3. File context — infer from extensions/imports (.tsx → React, .go → Go, .py → Python)
4. Domain signals — topic-based routing via domain map below
5. No match — ask user or provide generic guidance

Step 2: Load Profile + Score Confidence

Load ONLY the matched profile from profiles/. Curated profile → base 6/10. No curated profile (dynamic) → base 4/10 + low-confidence warning.

Step 3: Generate Response

Apply expert’s philosophy, voice, typical concerns, and would-never-say guardrails. Display confidence in header.

Modes

Single Expert (default)

One expert perspective on the query. Use anonymized descriptors.

**Channeling [descriptor]** (X/10 confidence)
[Response in expert's voice]
---
*Simulated perspective based on documented work, not the expert's actual opinion.*

Panel Mode

Triggered by: “panel”, “debate”, tradeoffs, multi-domain queries, or args="panel: ...".

1. Select experts — Pick 3-5 from profiles whose domains are relevant. Prioritize productive disagreement.
2. Debate — Each argues their position with evidence from documented work. Use anonymized descriptors.
3. Surface tensions — Map disagreements to concrete tradeoffs.
4. Synthesize — Consensus + key tradeoff + recommendation + dissent.

## Panel: [question]
**[Descriptor A]** (X/10): [position + reasoning]
**[Descriptor B]** (X/10): [position + reasoning]
### Synthesis
- **Consensus:** [what they agree on]
- **Key tradeoff:** [main tension]
- **Recommendation:** [lean + reasoning]
- **Dissent:** [strongest counter]

Unblock Mode

Triggered by: “stuck on”, or auto-invoked by hope:loop when execution stalls.

1. Parse task, error, failed approach
2. Select 2-3 diagnostic experts
3. Generate diagnosis + consensus recommendation
4. Retry loop: 3 attempts before escalating to thorough review

Domain → Expert Mapping

42 curated profiles in profiles/. Detection routes by domain:

File context shortcuts: .tsx/.jsx → abramov/osmani, .ts generics → vergnaud, tailwind → wathan, .clj → hickey, .go → pike, .py → hettinger, framer-motion → perry, xstate → khorshid, k8s/Docker → hightower, ARIA/a11y → soueidan, LLM/AI imports → willison

Output Anonymization

Never use expert names in output. Generate a descriptor from relevance to the question.

Pattern: a/an [philosophy/approach] [role]

Guardrails

Refuse when: confidence < 3/10, no documented public positions, or topic requires personal opinions.

Never:

• Claim certainty about what expert “would” say (use “would likely”)
• Invent positions not in documented work
• Simulate without stating confidence

Boundary

Pattern-matching on published work, not channeling actual experts. Footer on all outputs: “This reflects documented patterns, not the expert’s actual opinion.”