main
md 229 lines 8.01 KB
Rendered Raw
1 ---
2 name: reflect
3 description: Learning capture system that extracts HIGH/MED/LOW confidence patterns from conversations to prevent repeating mistakes. Use after user corrections ("no", "wrong"), praise ("perfect", "exactly"), or when discovering edge cases. Complements .squad/agents/{agent}/history.md and .squad/decisions.md.
4 license: MIT
5 version: 1.0.0-squad
6 domain: team-memory, learning
7 confidence: high
8 ---
9
10 # Reflect Skill
11
12 **Critical learning capture system** for Squad. Prevents repeating mistakes and preserves successful patterns across sessions.
13
14 Analyze conversations and propose improvements to squad knowledge based on what worked, what didn't, and edge cases discovered. **Every correction is a learning opportunity.**
15
16 ---
17
18 ## Integration with Squad Architecture
19
20 **Reflect complements existing Squad knowledge systems:**
21
22 1. **`.squad/agents/{agent}/history.md`** — Permanent learnings from completed work (append-only; each agent updates their own file; Scribe propagates cross-agent updates)
23 2. **`.squad/decisions.md`** — Team-wide decisions that all agents respect
24 3. **`reflect` skill** — Captures in-flight learnings from conversations that may graduate to history.md or decisions.md
25
26 **Workflow:**
27 - Use `reflect` during work to capture learnings
28 - At session end, review captured learnings
29 - Promote HIGH confidence patterns → lead agent for decision.md review
30 - Promote agent-specific patterns → `{agent}/history.md` updates
31
32 ---
33
34 ## Triggers
35
36 ### 🔴 HIGH Priority (Invoke Immediately)
37
38 | Trigger | Example | Why Critical |
39 |---------|---------|--------------|
40 | User correction | "no", "wrong", "not like that", "never do" | Captures mistakes to prevent repetition |
41 | Architectural insight | "you removed that without understanding why" | Documents design decisions (Chesterton's Fence) |
42 | Immediate fixes | "debug", "root cause", "fix all" | Learns from errors in real-time |
43
44 ### 🟡 MEDIUM Priority (Invoke After Multiple)
45
46 | Trigger | Example | Why Important |
47 |---------|---------|---------------|
48 | User praise | "perfect", "exactly", "great" | Reinforces successful patterns |
49 | Tool preferences | "use X instead of Y", "prefer" | Builds workflow preferences |
50 | Edge cases | "what if X happens?", "don't forget", "ensure" | Captures scenarios to handle |
51
52 ### 🟢 LOW Priority (Invoke at Session End)
53
54 | Trigger | Example | Why Useful |
55 |---------|---------|------------|
56 | Repeated patterns | Frequent use of specific commands/tools | Identifies workflow preferences |
57 | Session end | After complex work | Consolidates all session learnings |
58
59 ---
60
61 ## Process
62
63 ### Phase 1: Identify Learning Target
64
65 Determine what knowledge system should be updated:
66
67 1. **Agent-specific learning**`.squad/agents/{agent}/history.md`
68 2. **Team-wide decision**`.squad/decisions/inbox/{agent}-{topic}.md`
69 3. **Skill-specific improvement** → Document in session, recommend to skill owner
70
71 ### Phase 2: Analyze Conversation
72
73 Scan for learning signals with confidence levels:
74
75 #### HIGH Confidence: Corrections
76
77 User actively steered or corrected output.
78
79 **Detection patterns:**
80 - Explicit rejection: "no", "not like that", "that's wrong"
81 - Strong directives: "never do", "always do", "don't ever"
82 - User provided alternative implementation
83
84 **Example:**
85 ```text
86 User: "No, use the azure-devops MCP tool instead of raw API calls"
87 → [HIGH] + Add constraint: "Prefer azure-devops MCP tools over REST API"
88 ```
89
90 #### MEDIUM Confidence: Success Patterns
91
92 Output was accepted or praised.
93
94 **Detection patterns:**
95 - Explicit praise: "perfect", "great", "yes", "exactly"
96 - User built on output without modification
97 - Output was committed without changes
98
99 **Example:**
100 ```text
101 User: "Perfect, that's exactly what I needed"
102 → [MED] + Add preference: "Include usage examples in documentation"
103 ```
104
105 #### MEDIUM Confidence: Edge Cases
106
107 Scenarios not anticipated.
108
109 **Detection patterns:**
110 - Questions not answered
111 - Workarounds user had to apply
112 - Error handling gaps discovered
113
114 #### LOW Confidence: Preferences
115
116 Accumulated patterns over time.
117
118 ---
119
120 ### Phase 3: Propose Learnings
121
122 Present findings:
123
124 ```text
125 ┌─────────────────────────────────────────────────────────────┐
126 │ REFLECTION: {target (agent/decision/skill)} │
127 ├─────────────────────────────────────────────────────────────┤
128 │ │
129 │ [HIGH] + Add constraint: "{specific constraint}" │
130 │ Source: "{quoted user correction}" │
131 │ Target: .squad/decisions/inbox/{agent}-{topic}.md │
132 │ │
133 │ [MED] + Add preference: "{specific preference}" │
134 │ Source: "{evidence from conversation}" │
135 │ Target: .squad/agents/{agent}/history.md │
136 │ │
137 │ [LOW] ~ Note for review: "{observation}" │
138 │ Source: "{pattern observed}" │
139 │ Target: Session notes only │
140 │ │
141 ├─────────────────────────────────────────────────────────────┤
142 │ Apply changes? [Y/n/edit] │
143 └─────────────────────────────────────────────────────────────┘
144 ```
145
146 **Confidence Threshold:**
147
148 | Threshold | Action |
149 |-----------|--------|
150 | ≥1 HIGH signal | Always propose (user explicitly corrected) |
151 | ≥2 MED signals | Propose (sufficient pattern) |
152 | ≥3 LOW signals | Propose (accumulated evidence) |
153 | 1-2 LOW only | Skip (insufficient evidence) |
154
155 ### Phase 4: Persist Learnings
156
157 **ALWAYS show changes before applying.**
158
159 After user approval:
160
161 1. **For Agent History:**
162 - Append to `.squad/agents/{agent}/history.md` under `## Learnings` section
163 - Format: Date, assignment context, key learning
164
165 2. **For Team Decisions:**
166 - Create `.squad/decisions/inbox/{agent}-{topic}.md`
167 - Lead agent reviews and merges to `decisions.md` if appropriate
168
169 3. **For Skills:**
170 - Document recommendation in session notes
171 - Squad lead reviews and routes to skill owner
172
173 ---
174
175 ## Usage Examples
176
177 ### Example 1: User Correction
178
179 **Conversation:**
180 ```
181 Agent: "I'll use grep to search the repository"
182 User: "No, use the code search tools first, grep is too slow"
183 ```
184
185 **Reflection Output:**
186 ```
187 [HIGH] + Add constraint: "Use code intelligence tools before grep"
188 Source: "No, use the code search tools first, grep is too slow"
189 Target: .squad/agents/{agent}/history.md
190 ```
191
192 ### Example 2: Success Pattern
193
194 **Conversation:**
195 ```
196 Agent: [Creates PR with detailed description and test plan]
197 User: "Perfect! This is exactly the format I want for all PRs"
198 ```
199
200 **Reflection Output:**
201 ```
202 [MED] + Add preference: "Include test plan in PR descriptions"
203 Source: User praised detailed PR format
204 Target: .squad/decisions/inbox/pr-format.md (for team adoption)
205 ```
206
207 ---
208
209 ## When to Use
210
211 **Use reflect when:**
212 - User says "no", "wrong", "not like that" (HIGH priority)
213 - User says "perfect", "exactly", "great" (MED priority)
214 - You discover edge cases or gaps
215 - Complex work session with multiple learnings
216 - At end of sprint/milestone to consolidate patterns
217
218 **Don't use reflect when:**
219 - Simple one-off questions with no pattern
220 - User is just exploring ideas (no concrete decisions)
221 - Learning is already captured in history.md/decisions.md
222
223 ---
224
225 ## See Also
226
227 - `.squad/decisions.md` — Team-wide decisions
228 - `.squad/agents/*/history.md` — Agent-specific learnings
229 - `.squad/routing.md` — Work assignment patterns