Flywheel
Analyze recent session transcripts for user-correction patterns and propose agent/steering/skill configuration improvements. Use when asked to run the flywheel, improve agent configuration, or learn from past sessions.
Install
npx skills add aws-samples/sample-kiro-cli-multiagent-development --skill flywheelgit clone --depth 1 https://github.com/aws-samples/sample-kiro-cli-multiagent-development.git /tmp/sample-kiro-cli-multiagent-development
mkdir -p ~/.claude/skills
cp -r /tmp/sample-kiro-cli-multiagent-development/skills/flywheel ~/.claude/skills/flywheelPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# Flywheel: Agent Configuration Improvement Loop
Analyze recent sessions to identify patterns where the user had to correct, redirect, or steer the agent — then propose configuration changes to prevent recurrence.
Source of truth: the **v3 session transcripts** the engine writes for every session. No separate index
files or hooks are needed — parse the transcripts directly.
## Process
### Phase 1: Session Analysis
1. **Locate recent sessions.** v3 stores each session under `~/.kiro/sessions/<hash>/sess_<uuid>/` with
two files:
- `session.json` — metadata: `id`, `title`, `agentMode` (which agent ran), `modelId` (the model),
`workspacePaths` (cwd), `createdAt`/`lastModifiedAt`, `status`, `description`.
- `messages.jsonl` — the transcript; one JSON object per line, each with a `payload.type`:
- `user` → `payload.content` is the user's prompt
- `assistant` → `payload.content` is the agent's response (`operationType: "Say"` for text)
- `turn_start` → marks the start of a turn (`executionId`)
- `session_metadata` → e.g. context-usage samples
List sessions by `lastModifiedAt` descending; analyze the ~10 most recent (or a date range the user
specifies).
2. **Scan for correction events** by reading each `messages.jsonl` and examining every `user` message
(in sequence with the `assistant` messages around it). A correction event is a user message that
signals the agent did something wrong or suboptimal:Details
- Skill name
flywheel- Path
skills/flywheel/SKILL.md- License
- MIT-0
- Repo updated
- 9 September 2026
FAQ
What does the Flywheel skill do?
Analyze recent session transcripts for user-correction patterns and propose agent/steering/skill configuration improvements. Use when asked to run the flywheel, improve agent configuration, or learn from past sessions.
How do I install the Flywheel skill?
skills CLI (Claude Code, Cursor, Codex and more): npx skills add aws-samples/sample-kiro-cli-multiagent-development --skill flywheel. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Flywheel popular?
It lives in aws-samples/sample-kiro-cli-multiagent-development, which has 43 GitHub stars and 9 forks.
Related skills
Agentcore Patterns
43aws-samples/sample-kiro-cli-multiagent-development
Verified Amazon Bedrock AgentCore patterns for Runtime handlers, Gateway MCP clients with SigV4/IAM auth, Memory session management with Strands, and Bedrock cross-region inference-profile IAM policies. Use when building or debugging AgentCore-based agents.
AWS Technical Docs
43aws-samples/sample-kiro-cli-multiagent-development
AWS technical documentation patterns covering procedures, lists, tables, titles, error messages, and content patterns. Use when writing AWS service documentation, procedures, troubleshooting guides, tutorials, or getting started content.
CLI Execution
43aws-samples/sample-kiro-cli-multiagent-development
CLI execution best practices covering RTK token compression, AWS discover-then-execute patterns, output control, and bash/zsh scripting with logging. Use when running CLI commands, querying AWS resources, writing automation scripts, or working with shell tooling.
Principle Explain the Number
107Dimon94/skills
Apply before you measure, and before you trust, report, or act on a number you measured: a speedup, a regression, a throughput, a latency, or an eval result. Find what limits it, and rule out that it measured something other than the work you think.
Correct
107Dimon94/skills
从仓库历史找出 agent 反复犯的错误类别,逐类在能生效的最高层级让它不可能再犯,并证明新检查会在真实旧错误上失败。用于用户显式调用 /correct 或 $correct。
Principle Model the Domain
107Dimon94/skills
Apply when writing stateful logic, or when code branches a lot or repeats a shape assumption across files.
Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.