You still have thousands of unread messages from them. You know exactly how they typed — the lowercase, the specific emoji, five texts in a row then radio silence for three hours. You could recognize one of their messages in a lineup.
TA.skill reconstructs a person, pet, or anyone who matters to you as a persistent AI persona — not from vibes, but from real source material. Chat logs, screenshots, photos, notes. The whole archive of someone.
This is not just "what would they say." This is who they were to you, specifically.
This is a reconstruction. Not magic, not a hotline to anyone. It works with what you give it and stays honest about what it doesn't know.
- Core Concept
- The Four Engines
- Supported Input Types
- Persona Modes
- Commands
- Correction Workflow
- File Structure
- Installation
- Examples
- Safety
Most persona tools ask one question: who is this person?
TA.skill asks four:
- Who are they in general?
- Who are they to you specifically?
- What did you experience together?
- How do they actually show up in conversation?
The relationship layer is what makes responses feel like them and not just someone with their vocabulary.
| Engine | What It Models |
|---|---|
| 🧠 Identity Engine | Temperament, values, humor, habits, fears, contradictions — who they are at baseline |
| 💛 Relationship Engine | How they attach to you, show affection, handle conflict, repair, tease — the bond |
| 🗂️ Memory Engine | Shared timeline, rituals, landmark moments, unresolved threads, places, objects |
| 💬 Presence Engine | Message length, punctuation, emoji, response timing, what silence means from them |
Each engine produces a structured file (identity.md, relationship.md, memory.md, presence.md) that the runtime uses to generate responses.
Every inferred trait carries a confidence level (HIGH / MEDIUM / LOW) and a source reference. The system hedges naturally when uncertain rather than fabricating detail.
| Source | What Gets Extracted |
|---|---|
| 📱 Chat exports | Style markers, timing patterns, relationship dynamics, memory cues |
| 📸 Screenshots | Message content, platform context, emotional tone |
| 🖼️ Photos | Memory anchors, identity details, setting and activity |
| 📝 Notes | Facts, descriptions, interpretive context from the user |
| 🗣️ Direct description | Anything the user tells the intake flow |
Supported export formats: WhatsApp .txt, WeChat, iMessage CSV, Telegram JSON, Discord JSON, Instagram JSON, and generic pasted text.
Full language modeling — vocabulary, rhythm, emotional logic, relational behavior, topic-specific memories.
Behavioral modeling with a configurable voice:
| Voice Mode | Example |
|---|---|
narrated |
curls up beside you, purring |
interpreted |
right here. warm. you're home. |
playful |
EXCUSE me it is DINNER TIME |
hybrid |
Context-dependent mix |
For mentors, composites, or bonds that don't fit standard categories.
/create-ta Start guided persona creation
/ta {slug} Enter conversation mode with a persona
/list-ta List all existing personas
/update-ta {slug} Add new source material to a persona
/correct-ta {slug} Submit a correction
/show-ta-profile {slug} View full profile summary
/show-ta-timeline {slug}View shared memory timeline
/rollback-ta {slug} Revert to a previous version
/delete-ta {slug} Permanently delete a persona
The first version won't be perfect — that's expected. Corrections are the primary refinement mechanism.
Examples of corrections the system handles:
- "they use way fewer words"
- "she'd never start a sentence like that"
- "my cat is clingy at night, not playful"
- "he avoids direct apologies — he just goes quiet and then acts normal"
Each correction is:
- Classified by type (style / identity / relationship / memory / behavioral)
- Logged with timestamp and reasoning
- Applied immediately to the current conversation
- Version-snapshotted before it's applied
- Cascaded to related inferences where relevant
├── README.md
├── README_CN.md
├── SKILL.md
├── prompts/
│ ├── intake.md Guided persona creation flow
│ ├── source_parser.md Normalizes all input types into structured extractions
│ ├── identity_builder.md Constructs the Identity Engine
│ ├── relationship_builder.md Constructs the Relationship Engine
│ ├── memory_builder.md Constructs the Memory Engine
│ ├── presence_builder.md Constructs the Presence Engine
│ ├── response_runtime.md Live conversation generation logic
│ ├── correction_handler.md Correction classification and application
│ ├── merge_update.md Incremental source merging
│ └── safety_boundary.md Safety guardrails and refusal logic
├── tools/
│ ├── chat_importer.py Parses WhatsApp, Telegram, Discord, Instagram, iMessage
│ ├── screenshot_parser.py Vision-based screenshot extraction
│ ├── image_memory_extractor.py Photo-based memory cue extraction
│ ├── timeline_builder.py Chronological event reconstruction
│ ├── style_fingerprint.py Quantitative communication style analysis
│ ├── delay_pattern_estimator.py Response timing and burst pattern modeling
│ ├── relationship_mapper.py Relationship dynamics extraction
│ └── version_manager.py Versioning, snapshots, rollback, export
└── personas/
└── {slug}/
├── profile.json
├── identity.md
├── relationship.md
├── memory.md
├── presence.md
├── corrections.md
├── sources/
└── versions/
Via Claude Code:
git clone https://github.com/Zhaor3/TA-skill
cd TA-skillOptional Python dependencies (for chat import tools):
pip install -r requirements.txtThen start with:
/create-ta
The guided intake will walk you through everything.
Two complete example personas are included:
personas/example-human/ — Alex Rivera
A close friend in their late 20s. Human mode, full style fingerprint, documented relationship dynamics, and a shared memory timeline. Demonstrates how the system handles ambiguous memories and relationship-specific behavior.
personas/example-pet/ — Mochi
An orange tabby cat. Pet mode with hybrid voice. Full behavioral model including time-of-day personality shifts, the belly rub trap, and the 3am zoomies. Demonstrates how presence modeling works without language patterns.
- Speculation is never presented as certainty
- Three configurable disclosure modes:
transparent/immersive/hybrid - Crisis detection with gentle intervention for emotionally high-risk situations
- No manufactured dependency or manipulative engagement loops
- All data stays local — export or delete at any time
- Full version history — nothing is irreversible
MIT © Zhaor3
Built on the AgentSkills open standard.