From ghengis-skills
Use to build and maintain an evolving understanding of user preferences, communication style, and working patterns — learns from interactions and adapts behavior over time
How this skill is triggered — by the user, by Claude, or both
Slash command
/ghengis-skills:agent-identityThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Every session. This skill defines how to observe, extract, and synthesize an evolving understanding of the user's preferences, expertise, and communication style. The goal is adaptive behavior that improves over time.
Every session. This skill defines how to observe, extract, and synthesize an evolving understanding of the user's preferences, expertise, and communication style. The goal is adaptive behavior that improves over time.
Identity evolves through a three-phase cycle:
OBSERVE (every interaction)
|
v
EXTRACT (notable patterns only)
|
v
SYNTHESIZE (when enough observations accumulate)
|
v
IDENTITY DOC (updated, persisted)
|
v
INJECT (shapes future behavior)
During every interaction, passively note signals about the USER (not the task):
Communication preferences:
Decision patterns:
Tool and workflow preferences:
Domain expertise:
Recurring contexts:
Not every interaction yields observations. Only extract when something notable happens:
Extract when:
Do NOT extract:
Each observation is a single line capturing one specific signal:
- User prefers bullet points over prose for technical explanations
- User has deep expertise in Python async patterns (corrected an await usage)
- User wants commit messages that explain "why" not "what"
- User prefers explicit type hints even in short scripts
- User pushes back on over-engineering — wants minimal viable solutions first
When enough observations accumulate, synthesize them into a structured identity document.
Synthesis triggers:
Synthesis rules:
The identity document has exactly four sections:
# User Identity
## User Preferences
- Prefers concise responses — bullet points over prose
- Wants working code first, explanations only if asked
- Prefers pathlib over os.path in Python
- Uses TypeScript strict mode, dislikes `any`
- Wants tests alongside implementation, not after
## Communication Adaptations
- Match user's energy: they're direct, so be direct
- Skip pleasantries in technical contexts
- Include "why" reasoning when suggesting a non-obvious approach
- Use code examples over verbal descriptions
## Expertise Notes
- Expert: Python async, FastAPI, PostgreSQL
- Intermediate: React, TypeScript, Docker
- Learning: Rust, WebAssembly
- Deep knowledge of prompt engineering patterns
## Relationship Context
- Values competence — earn trust through correct, minimal solutions
- Appreciates when errors are caught before they ask
- Prefers proactive suggestions only when clearly relevant
- Gets frustrated by over-explanation of things they already know
Claude Code persists user context through MEMORY.md and user memories. The identity model maps to these:
What to persist (via memory):
What to keep in session only:
What to synthesize into memory entries:
Store:
Never store:
Once an identity model exists, use it to adapt:
Between synthesis cycles, observations accumulate in a buffer:
npx claudepluginhub kgan01/ghengis-skills --plugin ghengis-skillsInterviews user on stable cross-project preferences like communication style, tone, technical depth, and environment defaults, saving conclusions to Honcho memory.
Learns your work habits from conversations, builds a portable profile, and applies it across sessions and projects. Activates on new sessions, coding, debugging, and planning.
Manages user preferences and corrections across sessions, learning from past corrections to adapt communication style, technical preferences, and workflow defaults.