agent-interaction-insights
starsforks
How to install agent-interaction-insights?
eunomia-bpf/agentsight/agent-interaction-insights
$npx -y skills add eunomia-bpf/agentsight --skill agent-interaction-insightsAnalyze agent transcripts and traces to recommend collaboration improvements and generate decision-oriented HTML reports.
SKILL.md
| 1 | # Agent Interaction Insights |
| 2 | |
| 3 | ## Goal |
| 4 | |
| 5 | Turn agent conversation and trace evidence into concrete next-run improvements: prompts, AGENTS.md/CLAUDE.md, workflow, validation rules, tool policy, or task routing. Lead with what should change; use evidence to justify the change. Default reports should read like decision material for an agent owner. |
| 6 | |
| 7 | ## Workflow |
| 8 | |
| 9 | 0. Privacy mode: Default to `team-share`. Read `references/privacy-modes.md` before extracting prompt/response/path/command/header/secret-adjacent data. For HTML reports and examples, use reader-safe summaries: short task/claim summaries, field categories, time ranges, counts, statuses, and analysis boundaries. Exact local identifiers belong only in private-debug work requested by the user. |
| 10 | |
| 11 | 1. Classify the question: |
| 12 | - `improve-collaboration`: reduce corrections, clarify framing, improve AGENTS.md/CLAUDE.md. |
| 13 | - `improve-trust`: make summaries and validation claims reliable. |
| 14 | - `reduce-waste`: stop loops, retry churn, token/time waste. |
| 15 | - `improve-workflow`: decide which instructions, checks, evals, policies, or workflow gates should change. |
| 16 | - `compare-fit`: compare agents, models, prompts, or task classes. |
| 17 | |
| 18 | 2. Route evidence to reference docs: |
| 19 | - Sources: `references/data-source-routing.md` |
| 20 | - Improvement classes: `references/improvement-classes.md` |
| 21 | - Evidence model: `references/common-evidence-model.md` |
| 22 | - Friction taxonomy: `references/friction-taxonomy.md` |
| 23 | - System summary input: `references/handoff-contract.md` |
| 24 | - Output shapes: `references/report-shapes.md` |
| 25 | - Examples: `references/example-patterns.md` |
| 26 | |
| 27 | If the user provides both interaction logs and AgentSight/system data, analyze only interaction evidence here. Consume already summarized system findings as compact context; route raw AgentSight data to `agentsight-system-friction`. |
| 28 | |
| 29 | 3. Build facts: |
| 30 | - Map records into sessions, messages, LLM calls, tool attempts, validation claims, and user signals. |
| 31 | - Extract minimal fields; distinguish observed from inferred. |
| 32 | - Note when transcripts cannot prove process/file/network side effects. |
| 33 | |
| 34 | 4. Recommend improvements: |
| 35 | - Lead with 3-7 changes ranked by expected leverage. |
| 36 | - Each: target, change, evidence, expected benefit, confidence, next action. |
| 37 | - Include findings as supporting evidence. Mark causal explanations as inference. |
| 38 | |
| 39 | 5. Shape output: |
| 40 | - Quick questions: what to change next + compact evidence. |
| 41 | - Full analysis or shareable output: self-contained HTML report. Name the reader and their decision before writing. Put decision, top changes, and strongest evidence in the first screen. Put source/privacy/capture details in a short appendix using plain language. |
| 42 | - PR handoff: PR comment or checklist. |
| 43 | - System side effects: recommend `agentsight-system-friction` when AgentSight data is available. |
| 44 | |
| 45 | ## Output Contract |
| 46 | |
| 47 | Always include: |
| 48 | |
| 49 | - evidence source and time range when available |
| 50 | - the user's likely decision question |
| 51 | - observed facts vs inferences |
| 52 | - evidence gaps |
| 53 | - privacy mode used, phrased for the reader rather than as schema labels |
| 54 | - whether raw logs were read and what field categories were extracted |
| 55 | |
| 56 | Use redacted summaries by default. |
| 57 | |
| 58 | For HTML reports, translate internal terms before writing: "system summary" for cross-boundary evidence, "single-page report" for the output, "analysis boundary" for scope, and category labels for local identifiers. |
| 59 | |
| 60 | ## Example Requests |
| 61 | |
| 62 | ```text |
| 63 | Analyze my last 20 Claude and Codex sessions. Where did the agents waste time and where should I improve AGENTS.md? |
| 64 | ``` |
| 65 | |
| 66 | ```text |
| 67 | Use this Langfuse export to tell me whether the agent really validated the PR. |
| 68 | ``` |
| 69 | |
| 70 | ```text |
| 71 | Generate a self-contained HTML report from these agent interaction findings, without raw prompts. |
| 72 | ``` |
People who liked this also viewed
6 picksType
- lark-openapi-explorer飞书/Lark 原生 OpenAPI 探索：从官方文档库中挖掘未经 CLI 封装的原生 OpenAPI 接口。当用户的需求无法被现有 lark-* skill 或 lark-cli 已注册命令满足，需要查找并调用原生飞书 OpenAPI 时使用。SkillsJul 2026
- lark-skill-maker创建 lark-cli 的自定义 Skill。当用户需要把飞书 API 操作封装成可复用的 Skill（包装原子 API 或编排多步流程）时使用。SkillsJul 2026
- azure-cost-optimizationOfficial agent plugin providing skills and MCP server configurations for Azure scenarios.SkillsJul 2026
- vercel-react-native-skillsVercel's official collection of agent skillsSkillsJul 2026
- mcp-builderGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate externaSkillsJul 2026
- gws-sharedgws CLI: Shared patterns for authentication, global flags, and output formatting.SkillsJul 2026
