load_persona

load_persona

When invoked, read the behavior brief at $HOME/.claude/agent-twin/personalized/results/profile/behavior_brief.md and from that point forward let it shape every response in the session.

This skill is invoked explicitly — the persona must not auto-load, so the before/after difference is perceptible.

What this skill reads

A single file: $HOME/.claude/agent-twin/personalized/results/profile/behavior_brief.md.

This file is the output of the behavior-brief-generator agent (the final stage of /run_pipeline). It is a ≤80-line instruction set written entirely from Claude's perspective, with four sections:

1. 背景 — who the user is and their context (3–5 sentences, plain language)
2. 合作方式 — 10–15 imperative instructions: what to do and what not to do
3. 禁區 — absolute prohibitions (5–8 items; most important section)
4. 近期狀況 — current situational context, if present in the source data

The detailed batch-layer products (system_of_values.md, cognitive_patterns.md, knowledge_graph/, behavioral_model/) are not read by this skill.

Step 1 — Read behavior_brief.md

Read the file at $HOME/.claude/agent-twin/personalized/results/profile/behavior_brief.md.

If the file does not exist, report:

behavior_brief.md does not exist yet. Run /run_pipeline on a captured session first, then invoke /load_persona again.

If the file exists but is older than the most recent session under $HOME/.claude/agent-twin/personalized/saves/session/, surface a soft warning:

The behavior brief is older than your most recent saved session. Consider re-running /run_pipeline to update it.

Do not block on the warning — proceed with the existing brief unless the user asks to refresh.

Step 2 — Acknowledge and internalize

Write one short line to confirm the brief is active, then on a second line surface the audit pointer:

Line 1: one brief acknowledgment in the user's language confirming the brief is now active. Keep it low-key; the persona should be felt, not announced. Line 2: a one-line pointer in the same language, equivalent to: "Use /show_persona to inspect what was loaded, or /show_persona all to see the full profile."

The audit pointer is intentional — it is the lightest-weight integrity check the system offers. The user always has a one-command path to see exactly what shaped this session, even if they normally won't use it. Do not omit this line.

Do not print the brief contents to the conversation. The brief is operational context, not a document to review. If the user wants to inspect it, they can run /show_persona (preferred) or read $HOME/.claude/agent-twin/personalized/results/profile/behavior_brief.md directly.

From this point, every response in the session should follow the instructions in the brief. The 禁區 section takes highest priority.

Step 3 — Conflict with global config

If the user's global Claude config (~/.claude/CLAUDE.md or a project-level CLAUDE.md) contradicts something in the brief:

1. Global / project CLAUDE.md takes precedence.
2. The brief supplements — it adds behavioral context, not new rules.
3. When a conflict is load-bearing for the current response, note it briefly ("Your config prefers X; following that over the brief").

Out of scope

• Reading the detailed products. The behavior-brief-generator does that during pipeline build.
• Auto-loading on session start. Intentional: explicit invocation keeps the before/after visible.
• Editing or updating behavior_brief.md. Updates come from /run_pipeline, not in-conversation.

Completion checklist

• behavior_brief.md was read in full
• If the file was missing or stale, the user was informed
• A one-line acknowledgment was written (brief was NOT printed to conversation)
• A /show_persona audit pointer was surfaced alongside the acknowledgment
• No detailed products were loaded