consult_twin

consult_twin

A sub-context twin you can ask anything without dragging persona shaping into the current working session. The skill dispatches the twin-advisor agent via the Task tool; the agent reads the brief and the recent transcript in its own context, replies in the user's voice, and the response comes back to the orchestrator as plain text. The brief never enters this session's prompt.

This is the deliberate complement to /load_persona:

	/load_persona	/consult_twin
Scope	persists for the whole session	one Task dispatch, then gone
Side effect	every reply afterward is persona-shaped	nothing in this session changes
When to use	starting a working session you want to be persona-aware end-to-end	you only need the twin's take at one moment

The two are orthogonal. The user can have neither, either, or (in principle) both — /consult_twin does not call /load_persona internally and never will.

Argument grammar

/consult_twin <free-form question>

A single free-form string. No flags, no modes. The owner explicitly chose the simplest possible interface ("越無腦越好"). Mode (advisor vs. curious) is inferred from the question shape — see Step 4.

Step 1 — Resolve $HOME and the brief path

Run a single Bash command to expand $HOME and derive the brief path:

DATA_ROOT="$(bash -c 'echo $HOME')/.claude/agent-twin/personalized"
BRIEF_PATH="$DATA_ROOT/results/profile/behavior_brief.md"

All subsequent operations use the resolved absolute paths.

Step 2 — Verify the brief exists

If the file at BRIEF_PATH does not exist, surface exactly one line and stop:

No persona profile found. Run /run_pipeline on a captured session first.

Do not attempt to dispatch the agent without a brief — the agent has nothing to read.

Step 3 — Locate the current session's transcript (best effort)

Claude Code writes the active session's JSONL under ~/.claude/projects/<encoded-project-path>/. The encoding rule is: every path separator becomes - (so D:\agent-twin → D--agent-twin, /home/user/proj → -home-user-proj). Verify on this machine before assuming:

ls "$(bash -c 'echo $HOME')/.claude/projects/" | head

Pick the directory entry that matches the current working directory under that encoding.

Then glob *.jsonl inside that directory and pick the newest by mtime — the active session writes to its file continuously, so the most recently modified JSONL is almost always the current session:

ENCODED="$(pwd | sed 's#[/\\:]#-#g')"
PROJECTS_DIR="$(bash -c 'echo $HOME')/.claude/projects/$ENCODED"
TRANSCRIPT="$(ls -t "$PROJECTS_DIR"/*.jsonl 2>/dev/null | head -n1)"

If no transcript is found (unusual configuration, brand-new session before first turn flushes), proceed with SESSION_TRANSCRIPT_PATH empty. Advisor mode still works on the brief alone.

Step 4 — Mode inference from question shape

This is a soft heuristic. If the result is wrong, the cost is small (the user just rephrases). Future agents editing this SKILL: feel free to refine the prefix list as patterns emerge.

Set MODE=curious if the question (case-insensitive, after stripping leading whitespace and punctuation) starts with one of:

• English: research, explore, let's look at, let's research, help me understand, i want to learn about, i want to understand, what should i research, what should i learn about, tell me about (when paired with an unfamiliar-domain noun phrase — soft signal only)
• Traditional Chinese: 研究, 探索, 了解, 學習, 想了解, 想學, 幫我了解, 想研究

Otherwise set MODE=advisor. Default is advisor. When uncertain, choose advisor — it's the more useful failure mode.

Step 5 — Dispatch the twin-advisor agent

Use the Task tool with subagent_type="twin-advisor". Build the prompt as a structured key-value block so the agent can parse it robustly:

BRIEF_PATH=<absolute path from Step 1>
SESSION_TRANSCRIPT_PATH=<absolute path from Step 3, or empty>
MODE=<advisor|curious from Step 4>
QUESTION=<the user's question verbatim>

The description parameter for the Task call should be a short label like Twin consultation: <first 50 chars of question>. The agent's tools are Read only — it cannot write or run commands; it just reads, thinks, and responds.

Step 6 — Return the agent's response verbatim

When the agent returns, print its response to the user without any preamble. Do not say "Here's what your twin says:" or "The twin responded:". The agent already speaks in first-person; let it land directly.

Step 7 — One-line trailer

After the response, on a new line, append exactly:

(Consulted via twin-advisor — response did not enter this session's prompt context.)

This is the integrity confirmation: the user wanted assurance that contamination did not happen. Always include it.

What this skill must not do

• Call /load_persona. The two are orthogonal; the whole point is sub-context isolation.
• Read behavior_brief.md itself. The agent reads it; the SKILL only checks existence.
• Inject the brief or the transcript into this session's prompt by quoting it back.
• Add framing or commentary around the agent's response. The agent's reply IS the answer.
• Refuse a question because mode inference is ambiguous. Default to advisor.

Out of scope

• Editing or rebuilding the brief — use /run_pipeline.
• Loading the brief into the session — use /load_persona.
• Inspecting the profile — use /show_persona.
• Auditing analyses — use /run_meta_critic.

Completion checklist

• $HOME resolved via Bash; all paths used were absolute
• Missing behavior_brief.md was reported with the standard one-line message and execution stopped
• Most recent JSONL under ~/.claude/projects/<encoded-cwd>/ was located, or empty was passed if not found
• Mode was inferred from the question shape (default advisor)
• twin-advisor was dispatched via the Task tool with the four-key prompt block
• Agent's response was printed verbatim with no preamble
• One-line trailer about sub-context isolation was appended
• No call to /load_persona was made