self-reflect

name	self-reflect
description	Deep reflection and optimization with Opus. Two modes — /self-reflect analyzes completed work, extracts root causes, and stores lessons. /optimize analyzes a process or system, determines the right model and approach, then executes improvements. Use when reflecting on what happened, extracting lessons, optimizing workflows, or improving approaches. Triggers on "self-reflect", "reflect on", "optimize", "what went wrong", "what can we improve", "lessons learned".
category	debugging
runtimes	["claude"]
pii_safe	true

/self-reflect & /optimize — Deep Analysis + Action

Mode 1: /self-reflect [topic]

Opus-high reflection on what happened, why, and what to learn.

Steps

1. Switch to Opus-high:

   session_status model=opus-high
   

2. Gather full context:

   • Search haivemind: mcporter call haivemind.search_memories query="[topic]" limit=20
   • Read channel directive: contexts/[channel-id].md
   • Check recent Discord messages: message action=read channelId=[id] limit=30
   • Check git log if code-related: git log --oneline -20
   • Read any referenced plans, tickets, or docs

3. Deep reflection (Opus analyzes all of this):

   • What happened — timeline of events, decisions, outcomes
   • What worked — patterns, tools, approaches that succeeded
   • What failed — root causes (not symptoms), systemic issues
   • Why — connect failures to underlying assumptions, gaps in knowledge, process flaws
   • Counterfactuals — what would have changed the outcome
   • Patterns — recurring themes across similar past work (search haivemind for prior reflections)

4. Extract actionable lessons:

   • Each lesson = [LESSON]: [clear statement] | trigger=[what exposed it] | fix=[what to do differently]
   • Distinguish between: one-off mistakes vs systemic issues vs knowledge gaps
   • Rate severity: critical (will recur and cause damage) / moderate / minor

5. Store lessons (four destinations):

   • Channel directive (contexts/[channel-id].md) → channel-specific lessons under ## Self-Reflection [date]
   • haivemind → mcporter call haivemind.store_memory content="REFLECTION [channel-id] [date]: [lessons]" category="rules"
   • MCP Registry → if MCP tools were involved, update mcpRegistry.domains.[server].notes[]
   • MEMORY.md → only if the lesson is globally significant (rare)

6. Report: Conversational summary — lead with the most important insight. No fluff.

Mode 2: /optimize [target]

Opus-high analyzes a process/system, then determines the right model and approach to implement improvements.

Phase 1: Analysis (Opus-high — always)

1. Switch to Opus-high:

   session_status model=opus-high
   

2. Gather context (same as reflect — haivemind, channel directive, messages, code, docs)

3. Optimization analysis:

   • Current state — how does this process/system work today
   • Bottlenecks — where does time/effort/quality get lost
   • Root causes — why do the bottlenecks exist
   • Opportunities — what changes would have the highest impact
   • Dependencies — what blocks each improvement
   • Risk — what could go wrong with each change

4. Produce an optimization plan:

   • Ordered list of improvements, highest-impact first
   • For each: what to change, expected impact, effort estimate, risk level
   • Model assignment — Opus determines which model handles each optimization:

   Optimization type	Model	Why
   Code refactor / bug fix	cursor-agent/sonnet-4.6	CursorSonnet for efficient code changes
   Complex architecture change	cursor-agent/opus-4.6	Deep reasoning needed
   Config / infra change	max-high/claude-sonnet-4-6	Inline, needs care
   Skill / workflow rewrite	max-high/claude-sonnet-4-6	Inline reasoning
   Research-heavy optimization	google/gemini-3.1-pro-preview	1M context for large codebases
   Simple cleanup / formatting	max-low/claude-sonnet-4-6	Low-stakes, fast

Phase 2: Execution (Model determined by Phase 1)

5. Present plan to <USER_NAME>for approval:

   • Show the ordered improvements with model assignments
   • Wait for "do it", "go", or selective approval ("do 1 and 3")

6. Execute approved optimizations:

   • Switch model per-task if needed: session_status model=[assigned-model]
   • For code changes: spawn sub-agent with appropriate model
   • For inline changes: execute directly
   • Verify each change before moving to next

7. Post-optimization reflection (brief):

   • What was improved
   • Measured or expected impact
   • Store to haivemind: mcporter call haivemind.store_memory content="OPTIMIZATION [channel-id] [date]: [changes made] [impact]" category="operations"

Key Principles

• Opus-high always drives the thinking. Analysis and lesson extraction are never delegated to a cheaper model.
• Execution model is determined by the analysis. Opus picks the right tool for each job.
• No shallow reflections. Surface-level "we should communicate better" is useless. Find the mechanism.
• Lessons must be stored. A reflection without persistence is wasted tokens.
• Optimize requires approval. Never execute optimizations without explicit go-ahead.
• Feed the loop. Every reflection and optimization improves future /plan and /do runs.