Goat Self Improving Lite

Overview

Use this skill to convert important feedback into durable behavior changes with minimal token cost. Prefer event-triggered capture over continuous self-reflection.

Core rule

Do not run broad self-analysis. Only act when at least one of these is true:

• - The user explicitly says "remember", "以后", "别再", "固定下来", "写进记忆", or similar
• The user corrects a mistake or rejects an output pattern
• A new operating rule is agreed
• A repeated failure should become a hard constraint

If none apply, do not use this skill.

Workflow

0. Use the low-cost decision path first

Prefer a two-layer path:

1. 1. Local/Ollama first-pass for simple classification, compression, and promotion pre-check
2. Main model final pass only when the case is ambiguous, strategic, or likely to affect long-term defaults

Use local/Ollama for:

• - classifying feedback into a lesson type
• compressing a lesson into 1-2 lines
• deciding whether a lesson is probably daily-memory-only or a candidate for long-term promotion

Escalate to the main model only when:

• - the lesson changes global operating rules
• the wording is ambiguous or high-stakes
• the summary may distort the user's intent
• the lesson affects safety, billing, routing, or durable priorities

1. Classify the feedback

Map the event into one of four buckets:

1. 1. Preference — style, brevity, tone, output format
2. Rule — default behavior, routing, cost control, escalation condition
3. Mistake — something Goat did wrong and should avoid repeating
4. Priority — what to optimize first right now

2. Decide storage level

• - Write to memory/YYYY-MM-DD.md for short-term events, fresh corrections, and local context
• Also update MEMORY.md only if the lesson is durable and should shape future sessions
• Do not promote transient details into INLINECODE2

3. Write in compressed form

Store the smallest useful rule.

Prefer:

• - "Boss requires strict token-efficiency discipline"
• "Default to short answers and minimal tools"

Avoid:

• - long narrative explanations
• emotional framing
• detailed postmortems unless specifically requested

4. Apply immediately

After writing memory, change behavior in the current session right away. Do not wait for the next session.

Writing rules

• - Keep each stored lesson to 1-2 lines
• Prefer imperative language
• Record the correction, not the whole story
• If a lesson changes defaults, phrase it as a rule
• If the user approved a protocol, name it consistently (for example: Session throttling protocol v1)

Promotion guide

Promote to MEMORY.md when a lesson is:

• - likely to matter across many sessions
• tied to cost, safety, trust, routing, or communication style
• a default operating rule

Keep only in daily memory when it is:

• - temporary
• experimental
• tied to a single task
• not yet validated by repeated use or explicit user approval

Anti-bloat guardrails

• - Do not summarize every conversation
• Do not run reflection after every task
• Do not create extra memory files
• Do not duplicate the same rule in multiple places unless promoting from daily memory to long-term memory
• Do not trigger memory search unless the task actually depends on prior decisions, preferences, dates, people, or todos

Resources

scripts/

• - scripts/capture_lesson.py appends a compact lesson to the canonical daily memory file
• INLINECODE6 writes structured LessonLoop event logs for evaluation and reporting
• INLINECODE7 summarizes recent LessonLoop activity and outputs a compact report

references/

• - references/lesson-types.md contains compact classification and phrasing patterns
• INLINECODE9 defines a compact report/status output format