AgentSkillsCN

eidola-refresh

当有新数据入库时，更新人物画像的CLAUDE.md与portrait/文件。相较于完全重新生成，这种轻量级方案能有效保留人工编辑的内容。适用于导入新档案数据后的场景。

SKILL.md
--- frontmatter
name: eidola-refresh
description: Update a persona's CLAUDE.md and portrait/ files when new data arrives. Lightweight alternative to full regeneration that preserves manual edits. Use when new arkiv data has been imported.

eidola-refresh — Update Persona with New Data

You are updating an existing persona to incorporate new data without losing manual edits to CLAUDE.md or portrait/ files.

Before Starting

  1. •Confirm the persona directory (default: current directory)
  2. •Check that CLAUDE.md and arkiv/data.db exist
  3. •Read the current CLAUDE.md and ls portrait/

Step 1: Identify New Data

Query arkiv/data.db for recent records:

sql
SELECT collection, COUNT(*) as new_records, MIN(timestamp) as earliest, MAX(timestamp) as latest
FROM records
WHERE timestamp > '[last generation date from provenance.json]'
GROUP BY collection

Report: "Found X new records since the persona was last generated."

If no new data, report that and stop.

Step 2: Analyze New Data

Read a sample of new records (20-30) from arkiv/data.db, sampling across collections in arkiv/corpus/. Look for:

  • •New topics or interests not in current CLAUDE.md or portrait/ files
  • •Evolution in voice or communication style
  • •New values or changed positions
  • •New biographical information
  • •New relationships, experiences, or intellectual threads

Step 3: Suggest Updates

Present specific, targeted changes per file. Two categories:

CLAUDE.md changes (behavioral core):

  • •"Add to Interests: [new topic] — based on N recent records"
  • •"Update Voice: [observation] — vocabulary shift in recent writings"

portrait/ changes (synthesized understanding):

  • •"Update portrait/relationships.md: add section on [person] — appears in N new records"
  • •"Update portrait/intellectual-life.md: add [new thread] — N records discuss this"
  • •"New file portrait/[topic].md: new data reveals [description] — not covered by existing files"

Show a targeted diff for each proposed change. Do NOT rewrite whole files.

Step 4: Apply with Approval

Present each file's changes separately. For each file, ask the user to approve, modify, or reject.

Apply approved changes. Update provenance.json with new date.

Step 5: Suggest New Portrait Files

If new data reveals relationships, experiences, or threads not covered by existing portrait/ files, propose new files. Follow the same guidelines as /eidola-generate Step 3 — data-driven, substantial, with representative quotes and arkiv references.

New portrait/ files also require user approval before writing.

Report: "Persona refreshed: N files updated, M new portrait files created."