Agent Skills
› NeverSight/learn-skills.dev
› secondme-skill
secondme-skill
GitHub构建AI分身的全流程编排技能，从个人数据提取身份、建立私有知识库、训练本地模型到治理共享内容。支持隐私优先和离线运行，包含初始化、数据摄入、身份提炼、模型训练评估及集成报告等阶段，并设有人工审批与安全发布机制。
Trigger Scenarios
需要基于个人数据创建定制化AI分身
希望本地化训练并部署具有个人风格的语音或文本模型
需要对AI分身的生成过程和共享内容进行隐私与合规治理
Install
npx skills add NeverSight/learn-skills.dev --skill secondme-skill -g -y
SKILL.md
Frontmatter
{
"name": "secondme-skill",
"license": "MIT",
"metadata": {
"author": "acnlabs",
"version": "0.1.2"
},
"description": "A complete pipeline to build your AI Second Me: distill your identity from personal data, grow a private knowledge base, train a local model, and govern what gets shared.",
"allowed-tools": "Read Write Edit Bash WebSearch",
"compatibility": "OpenPersona\/OpenClaw\/Cursor, local-first with optional network interoperability."
}
secondme-skill
secondme-skill is a complete pipeline for building your AI Second Me — distill your identity from owned data, grow a private knowledge base, train a local model that speaks like you, and govern what gets shared. Local-first, privacy-first, fully yours.
This is an orchestration skill package. It does not replace lower-level capabilities.
Source of truth
- Persona source declaration:
persona.json - Pipeline runtime state:
state/pipeline-state.json - Product and governance spec:
references/product-report.md - Generated runtime pack:
generated/persona-secondme-skill/ - Regeneration script:
scripts/regenerate-pack.sh
Dependency chain
- Foundation:
openpersona-> persona pack creation and lifecycle baseline - Orchestration:
secondme-skill-> workflow gates, state, and report contracts - Capability chain:
anyone-skill-> identity extraction and evidence gradingpersona-knowledge-> data ingestion, deduplication, wiki/KG, versioned exportpersona-model-trainer-> local training, evaluation, export, integration
Required execution policy
- Use non-interactive generation and scripts where possible.
- Keep local-first and least-privilege defaults.
- Keep stage outputs auditable with version/hash references.
- Treat
persona-secondme-skill/as generated output (read-only baseline). - When
persona.jsonchanges, regenerate the runtime pack before release. - Before publishing outside this repository, run
scripts/publish-check.sh. - Preferred release check path:
scripts/run-gates.sh(regenerate + sync + model gate + publish gate).
Stage contract
init
- Validate toolchain and directories.
- Initialize or load
state/pipeline-state.json.
ingest
- Ingest user-owned data with PII scanning.
- Require explicit source authorization from user.
distill
- Build structured persona extraction artifacts.
- Ensure minimum persona input for OpenPersona is complete.
train
- Route by hardware tier:
- Apple Silicon:
mlx - NVIDIA:
unsloth - No local GPU:
colab
- Apple Silicon:
eval
- Check thresholds:
voice_score >= 3.5probe_score >= 0.8perplexitydegradation <= 20% vs last viable version
integrate
- Integrate model artifacts only when eval gate passes.
- Require runtime pack persona model integration before marking stage pass.
report
- Emit three reports under
reports/data,reports/model,reports/deploy. - Keep
reportand deployment recommendation in blocked state if persona model gate fails.
Failure routing
- Data gate fail -> return to
ingestand request source expansion. - Train fail -> change backend or reduce model size.
- Eval fail -> augment data or retune hyperparameters, then retrain.
Always update error_code, last_error, and retry_count in pipeline state before retry.
Human approval gates
Require explicit human approval for:
- financial/legal commitments
- account-changing write actions
- external publishing/sharing of identity artifacts
Persona model gate
secondme requires trained persona model integration, not only host default model fallback.
Pass criteria:
generated/persona-secondme-skill/persona.jsoncontainsbody.runtime.models.body.runtime.modelshas at least one model entry.scripts/check-model-integration.shreturns success.
If this gate fails, report must not be marked pass and deployment recommendation remains blocked.
Sync discipline
- Edit root
persona.jsonand orchestration docs first. - Run
scripts/regenerate-pack.sh. - Run
scripts/check-sync.shto validate root and generated pack alignment. - Verify runtime pack path
generated/persona-secondme-skill/exists and updated. - Only then produce release reports under
reports/.
Version History
- e0220ca Current 2026-07-05 23:00
