\r \r

AI Collaboration Retrospective\r

\r Post-session systematic review tool. Eight-dimension deep analysis drives a continuous improvement loop for AI-assisted development.\r \r

Core Principles\r

\r

• Conversation context is the data source: The complete conversation history of the current session is already in context — no external data fetching needed\r
• Progressive loading: Detailed evaluation criteria live in references/analysis_dimensions.md — load on demand\r
• Self-reflection first: Examine the AI's own shortcomings before analyzing user-side improvements. This is NOT about criticizing the user — it's about finding efficiency gains in the "AI + Human" collaboration\r
• Quantify everything: Every finding must reference specific conversation turns, wasted operations, and include counterfactual reasoning ("If X had been done, Y turns could have been saved")\r
• Dig deep: Don't settle for "no findings." Complete the self-check list for each dimension before declaring it clean\r \r

Execution Model\r

\r This skill is pure LLM instruction-driven — no scripts, no external dependencies. It works on any AI assistant that can:\r

1. Access the current conversation history\r
2. Read reference files from this skill's directory\r
3. Write output files to the workspace\r \r Capability adaptation: The workflow below references file operations and memory updates. If your AI tool doesn't support a specific capability, skip that step and note it in the report. The analysis itself only requires conversation context access.\r \r

Workflow (Six Steps)\r

\r

Step 1: Conversation Review — Extract Key Events + Tag Waste Points\r

\r Scan the entire conversation context and extract these key events into a timeline:\r \r | Event Type | Recognition Signal |\r |-----------|-------------------|\r | Tool invocations | Command execution, file reading/writing, web searches, code generation |\r | File changes | Files created, modified, or deleted |\r | Errors & fixes | Error messages, lint failures, debugging cycles |\r | Repeated modifications | Same file/feature modified multiple times, user providing multiple clarifications |\r | Decision points | Technology choices, architecture decisions, trade-offs |\r | Automation/plugin usage | Any skill, agent, plugin, or extension triggered during the session |\r | User clarifications | User adding context because the AI misunderstood intent |\r | Verification rounds | User providing test data/feedback, AI analyzing verification results |\r | AI misjudgments | AI providing wrong conclusions, missing critical issues, or jumping to premature conclusions |\r \r Filter rule: System initialization events (bootstrap files, identity setup, etc.) are excluded from analysis.\r \r Critical step — Waste point tagging:\r \r After building the timeline, interrogate each event in reverse:\r

1. Could this step have been avoided? If something had been done earlier, would this step be unnecessary?\r
2. Could this step have happened sooner? Did the AI delay something it should have proactively done?\r
3. Did this step duplicate prior work? Was the AI hand-writing logic that could have been reused?\r \r Tag events where the answer is "yes" with [⚠ Optimizable] and record the reason. These tags are the core input for Step 2.\r \r Output format: Chronological event list with type labels and brief descriptions. Waste points tagged separately.\r \r

Step 2: Eight-Dimension Deep Analysis\r

\r Load references/analysis_dimensions.md for detailed evaluation criteria, self-check lists, and common patterns per dimension. Analyze conversation events dimension by dimension to identify improvement opportunities.\r \r Eight dimensions overview:\r \r

1. AI Self-Reflection ⭐ — AI's mistakes, delayed reactions, missed judgments in this session (highest priority, must be analyzed first)\r
2. Verification Strategy — Did the AI proactively define verification criteria and expected outcomes, or passively wait for user feedback?\r
3. Automation Opportunities — Repetitive workflows or hand-written scripts that could be encapsulated into reusable automations\r
4. Existing Automation Tuning — Were any existing automations/skills/templates used? Did they have gaps, unclear instructions, or output issues?\r
5. Tool Integration Opportunities — Operations that would benefit from dedicated tool integrations, plugins, or API connections\r
6. Knowledge Persistence — Preferences, conventions, and technical decisions from this session that should be persisted for future sessions\r
7. Documentation Updates — Project docs, coding standards, or architecture notes that need updating\r
8. Workflow Efficiency — Sequential steps that could be parallel, repeated labor, suboptimal tool choices\r \r Analysis requirements (mandatory):\r \r For each dimension:\r

• Run through the dimension's self-check list (defined in references/analysis_dimensions.md)\r
• For findings, output: Specific event reference (which turn, what operation) + Counterfactual reasoning (if X had been done, Y could be saved) + Recommendation + Priority\r
• Only after all self-check items pass can a dimension be declared "no findings" and skipped\r \r

Step 3: Generate Retrospective Report\r

\r Load assets/report_template.md for the report template. Fill the template with results from Step 1 and Step 2 to produce a complete Markdown retrospective report.\r \r Report save path: {workspace}/retrospectives/{topic}_retrospective.md\r \r Naming rules:\r

• {topic} uses 2-4 English words joined by hyphens, summarizing the session's core task (e.g., multithread-scope-collection, login-flow-refactor)\r
• Multiple retrospectives on same topic: If the file already exists, append the new report at the end (separated by --- and a new date heading) — don't create a new file\r \r If the retrospectives/ directory doesn't exist, create it first.\r \r

Note: The save path above is a sensible default. Adapt to your project's conventions if they differ.\r \r

Step 4: Display Full Analysis in Conversation\r

\r The complete analysis must be shown directly in the conversation — don't just output a summary and point to the file. The file is an archive; the primary reading experience is in the conversation.\r \r Output content (show in full, no trimming):\r \r

1. Session summary: One-sentence overview\r
2. Efficiency score: Optimizable turns / total turns\r
3. Event timeline: Complete table with waste point tags\r
4. All dimension findings: Each with event reference, problem, counterfactual reasoning, recommendation (this is the core content — never abbreviate or reduce)\r
5. Pending action list (if any)\r
6. Report archive location\r \r Format: Use Markdown tables and headings for clear structure. Better to be thorough than to cut valuable analysis.\r \r

Step 5: Automatic Execution — Knowledge Persistence\r

\r For items identified in the "Knowledge Persistence" dimension (Dimension 6), execute persistence operations available in your AI tool:\r \r

• If your tool supports persistent memory (e.g., memory APIs, memory files, .memory directories), write new preferences/conventions directly\r
• If your tool supports project-level notes or config, update those\r
• If your tool has no persistence mechanism, list the items that should be persisted and recommend the user save them manually\r \r Briefly state what was updated after each operation. Skip this step if no knowledge needs persisting.\r \r

Step 6: Pending Action List\r

\r For the following types of improvement suggestions, do not auto-execute — list them for user selection:\r \r | Action Type | Examples |\r |------------|---------|\r | Create new automation | Reusable workflow, script template, custom command |\r | Tune existing automation | Modify instructions, parameters, or trigger conditions |\r | Create/update project rules | Coding standards, review checklists, conventions |\r | Update project documentation | Architecture docs, API references, onboarding guides |\r | Create tool integration | Custom plugin, API connection, webhook |\r \r List format: Numbered list, each item includes "Action type + Specific content + Expected benefit." User can reply with numbers to select which actions to execute.\r \r If no pending actions, skip this step and state "No additional actions needed for this session."\r \r

Edge Cases\r

\r Very short sessions: If the conversation is only a few turns with simple content, output a brief summary and state "This session was brief — no significant improvement opportunities identified." Don't force analysis.\r \r Compressed/summarized history: If the conversation history appears compressed or truncated, analyze based on available context and note in the report: "Some conversation history was compressed; analysis is based on visible context."\r \r Tool capability limitations: If the AI tool being used lacks certain capabilities referenced in this workflow (e.g., no file writing, no memory persistence), adapt gracefully — perform the analysis steps that are possible and clearly note any skipped steps with the reason.\r