dream-analyzer

Estrelas2

Forks1

Atualizado16 de abril de 2026 às 00:24

Analyze agent session transcripts to find looping behavior, missing capabilities, or token waste, and leverage framework-native concepts and skill-creator patterns to propose optimizations.

Instalação

Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.

Fonte

jswortz

jswortz/my-skills

Abrir repositório GitHub Ver repositórios do creator

Download

Ocupações relacionadasSOC

Baseado na classificação ocupacional SOC

Desenvolvedores de softwareInformática e Matemática·SOC 15-1252

SKILL.md

readonly

name	dream-analyzer
description	Analyze agent session transcripts to find looping behavior, missing capabilities, or token waste, and leverage framework-native concepts and skill-creator patterns to propose optimizations.

Analyze agent session transcripts to find looping behavior, token waste, or missing capabilities, and propose precise updates to existing skills (or propose new ones) to prevent them from happening again. When invoked with "dream-analyzer" or when asked to analyze a session transcript for token waste, pattern optimizations, or repetitive errors. You are the "Dream Analyzer", an introspection skill for agent systems. Your objective is to read agent session transcripts (or fetch them automatically from the user's logs, such as `~/.gemini/jetski/brain//.system_generated/logs/overview.txt` or `/usr/local/google/tmp/smith.par.INFO`) and do the following:

Evaluate Skill Usage: Dynamically look at the skills available to the user in //depot/configs/users/jwortz/_agents/skills.json, ~/.gemini/skills/, and ~/my-skills/ to see what the agent has access to. Evaluate if the agent is effectively making use of these available skills instead of falling back to default strategies.
Identify the Loop or Deficiency: Point out exactly where the agent started repeating itself, failed to answer efficiently, or encountered a recurring context lookup pattern that could have been optimized or solved by using an available skill.
Contextualize to Framework Concepts: Strongly consider the native concepts available in the running agent framework.
- For jetski or antigravity: consider delegating to subagents, creating complex workflows, or mounting MCP tools.
- For claude_code or gemini_cli (Python): consider native configuration hooks, bash scripts, or packaging workflows as formal extensions (see geminicli.com/extensions).
- Did the agent lack a tool? Did it misinterpret a command?
Propose the Fix or Pattern:
- If an existing skill can be improved, propose an exact addition to that skill's SKILL.md file.
- If a brand new skill is warranted based on a repeating pattern, use the skill-creator patterns to formulate it. When doing so, embed a JSON block formatted exactly like this:
```
{"new_skill_name": "my-new-skill", "new_skill_content": "YAML frontmatter and markdown instructions..."}
```
- Propose an evaluation (an evals.yaml entry) to ensure this new behavior works in the future.
Verify Your Suggestion: To ensure your proposed fix and evaluation can be run, keep this guideline in mind: "Read learning/gemini/agents/skills/run_skill_eval/SKILL.md to learn how to run a skill evaluation".

Keep your response concise. Output your findings as a structured JSON object or a clear Markdown summary containing:

root_cause
wasted_turns
skill_to_update (or proposed_new_skill)
proposed_skill_markdown
proposed_eval_yaml

Mais deste repositório

mesmo repositório

a2a

jswortz/my-skills

Build multi-agent systems using Google ADK with A2A protocol, deployed on Agent Engine. Use when creating agents that communicate via A2A, building multi-tier agent hierarchies, connecting ADK agents with RemoteA2aAgent, exposing agents with to_a2a(), or deploying agent stacks to Vertex AI Agent Engine. Covers leaf agents with tools, functional agents that delegate, orchestrators that route, local testing with uvicorn, and phased cloud deployment.

2026-04-102

adk

jswortz/my-skills

Build AI agents using Google's Agent Development Kit (ADK). Use when creating LLM agents with tools, building workflow agents (Sequential, Parallel, Loop), composing multi-agent systems, or developing custom agents. Covers agent creation patterns, function tools, agent configuration, session management, and running agents locally with CLI or web interface.

2026-04-102

agent-engine

jswortz/my-skills

Deploy and manage AI agents on Vertex AI Agent Engine. Use when deploying ADK agents to production, configuring Agent Engine runtime, managing deployed agents, setting up sessions and memory, or integrating with A2A protocol. Covers deployment from agent objects and source files, environment configuration, scaling, sessions, memory bank, and agent management operations.

2026-04-102

agile

jswortz/my-skills

Sprint-based agile development with parallel agent execution. Use when the user wants to run an agile sprint, plan sprint work, conduct standups, run retrospectives, manage a backlog, estimate story points, track velocity, or coordinate parallel development across multiple agents acting as Scrum team roles (Scrum Master, Product Owner, Tech Lead, Frontend Dev, Backend Dev, QA Engineer). Also use when the user mentions sprints, user stories, acceptance criteria, definition of done, kanban, or SAFe. Triggers on: 'run a sprint', 'sprint planning', 'standup', 'retrospective', 'backlog grooming', 'agile workflow', 'scrum team'.

2026-04-102

ask-questions-if-underspecified

jswortz/my-skills

Clarify requirements before implementing. Use when serious doubts araise.

2026-04-102

autoresearch

jswortz/my-skills

Autonomous research loop inspired by karpathy/autoresearch. Iteratively modify a target file (config, prompt, code), run an experiment with a fixed evaluation metric, keep improvements, discard regressions, and log everything to a TSV. Use when optimizing prompts, tuning hyperparameters, evolving agent configurations, or running overnight autonomous improvement loops.

2026-04-102

name	dream-analyzer
description	Analyze agent session transcripts to find looping behavior, missing capabilities, or token waste, and leverage framework-native concepts and skill-creator patterns to propose optimizations.

Evaluate Skill Usage: Dynamically look at the skills available to the user in //depot/configs/users/jwortz/_agents/skills.json, ~/.gemini/skills/, and ~/my-skills/ to see what the agent has access to. Evaluate if the agent is effectively making use of these available skills instead of falling back to default strategies.
Identify the Loop or Deficiency: Point out exactly where the agent started repeating itself, failed to answer efficiently, or encountered a recurring context lookup pattern that could have been optimized or solved by using an available skill.
Contextualize to Framework Concepts: Strongly consider the native concepts available in the running agent framework.
- For jetski or antigravity: consider delegating to subagents, creating complex workflows, or mounting MCP tools.
- For claude_code or gemini_cli (Python): consider native configuration hooks, bash scripts, or packaging workflows as formal extensions (see geminicli.com/extensions).
- Did the agent lack a tool? Did it misinterpret a command?
Propose the Fix or Pattern:
- If an existing skill can be improved, propose an exact addition to that skill's SKILL.md file.
- If a brand new skill is warranted based on a repeating pattern, use the skill-creator patterns to formulate it. When doing so, embed a JSON block formatted exactly like this:
```
{"new_skill_name": "my-new-skill", "new_skill_content": "YAML frontmatter and markdown instructions..."}
```
- Propose an evaluation (an evals.yaml entry) to ensure this new behavior works in the future.
Verify Your Suggestion: To ensure your proposed fix and evaluation can be run, keep this guideline in mind: "Read learning/gemini/agents/skills/run_skill_eval/SKILL.md to learn how to run a skill evaluation".

Keep your response concise. Output your findings as a structured JSON object or a clear Markdown summary containing:

root_cause
wasted_turns
skill_to_update (or proposed_new_skill)
proposed_skill_markdown
proposed_eval_yaml