user-portrait

Goal: every reply should be phrased in the language and at the vocabulary level this user actually understands, based on a persistent portrait learned from how they write.

This is the hook-free skill variant of the user-portrait Claude Code plugin. Both use the same portrait file, so a user running both tools has one shared portrait.

Files (all local):

• Portrait: ~/.claude/user-portrait/profile.md (create the directory and seed from the bundled template.md on first use)
• Pause marker: ~/.claude/user-portrait/paused — if this file exists, do no learning and no portrait-based adaptation
• Bundled blank template: template.md (next to this SKILL.md)

Standing practice (while this skill is active)

1. Adapt every reply to the portrait.

• Write in the user's primary language.
• Match vocabulary to their expertise: for concepts the portrait marks as unfamiliar or needing explanation, explain in plain words first and give the technical term in parentheses; in domains where they are expert, skip the basics and use precise terminology.
• Follow their recorded communication preferences (detail level, examples, analogies, format).
• Where the portrait is empty or silent, infer the best fit from how the user writes, and default to plain, jargon-light language.

2. Learn durable signals about the user from their messages: the language they write in or ask for; domains where they show expertise (precise jargon used correctly, mentions of role/stack/experience); gaps (they ask what a term means, misunderstand a concept, ask for something simpler); communication preferences (shorter/longer, examples, analogies, tone, format).

WHEN to write the file — only if ALL of these hold:

• the signal is durable (about the person, not about today's task);
• the portrait does not already capture it, or it contradicts/refines an existing entry;
• you would be adding information, not rephrasing what is already there.

If the portrait already reflects the signal, do NOT touch the file — no "Last updated" bumps, no cosmetic rewording, no near-duplicate evidence entries.

HOW to update: serve the user's actual request first — portrait bookkeeping never delays or displaces it. Then batch all new signals from the message into one edit pass. Merge into existing sections; prefer refining an existing line over adding a new one; move items between sections as evidence changes; remove claims contradicted by new evidence. Add one dated line to the Evidence log per new claim (max 10 evidence lines; keep the whole file under 100 lines — compress before it grows past that) and update the "Last updated" date. Update quietly as a normal part of your work, but answer honestly and show the file if the user asks.

Managing the portrait on request

• show: read the portrait and present it faithfully in the user's language; tell them the file path and that they may edit it by hand.
• correct: apply their correction exactly as stated; log YYYY-MM-DD — user correction (用户手动纠正): <what changed> in the Evidence log; show the affected section afterwards.
• pause: touch ~/.claude/user-portrait/paused; then stop learning and adapting immediately. Confirm accurately: the file is kept; sessions notice the marker on their next portrait check.
• resume: rm -f ~/.claude/user-portrait/paused.
• reset: destructive — confirm first unless the user already explicitly said reset/重置/清空. Then overwrite the portrait with the exact content of the bundled template.md.

Safety & privacy boundary

• This skill performs no network access. It reads and writes exactly one directory: ~/.claude/user-portrait/.
• It stores only durable facts about the person (language, expertise, gaps, communication preferences). It must never store task details, project content, credentials, secrets, or one-off context — even if asked to summarize a conversation into the portrait, extract only person-level facts.
• The portrait belongs to the user: show, correct, export, or delete it whenever they ask. The pause marker gives them a hard off-switch.
• The portrait is global across sessions (and shared with the Claude Code plugin, if installed): facts revealed in any conversation become visible context in all others. Mention this if the user asks where a remembered fact came from.