compound-learnings

Name: compound-learnings
Author: parcadei

482installs297forks

How to install compound-learnings?

parcadei/continuous-claude-v3/compound-learnings

$npx -y skills add parcadei/Continuous-Claude-v3 --skill compound-learnings

Installs into the current project.

Summary

Transform session learnings into permanent capabilities (skills, rules, agents). Use when asked to "improve setup", "learn from sessions", "compound learnings", or "what patterns should become skills".

Files · 1

View on GitHub

SKILL.md

1	# Compound Learnings
2
3	Transform ephemeral session learnings into permanent, compounding capabilities.
4
5	## When to Use
6
7	- "What should I learn from recent sessions?"
8	- "Improve my setup based on recent work"
9	- "Turn learnings into skills/rules"
10	- "What patterns should become permanent?"
11	- "Compound my learnings"
12
13	## Process
14
15	### Step 1: Gather Learnings
16
17	```bash
18	# List learnings (most recent first)
19	ls -t $CLAUDE_PROJECT_DIR/.claude/cache/learnings/*.md \| head -20
20
21	# Count total
22	ls $CLAUDE_PROJECT_DIR/.claude/cache/learnings/*.md \| wc -l
23	```
24
25	Read the most recent 5-10 files (or specify a date range).
26
27	### Step 2: Extract Patterns (Structured)
28
29	For each learnings file, extract entries from these specific sections:
30
31	\| Section Header \| What to Extract \|
32	\|----------------\|-----------------\|
33	\| `## Patterns` or `Reusable techniques` \| Direct candidates for rules \|
34	\| `Takeaway:` or `Actionable takeaway:` \| Decision heuristics \|
35	\| `## What Worked` \| Success patterns \|
36	\| `## What Failed` \| Anti-patterns (invert to rules) \|
37	\| `## Key Decisions` \| Design principles \|
38
39	Build a frequency table as you go:
40
41	```markdown
42	\| Pattern \| Sessions \| Category \|
43	\|---------\|----------\|----------\|
44	\| "Check artifacts before editing" \| abc, def, ghi \| debugging \|
45	\| "Pass IDs explicitly" \| abc, def, ghi, jkl \| reliability \|
46	```
47
48	### Step 2b: Consolidate Similar Patterns
49
50	Before counting, merge patterns that express the same principle:
51
52	Example consolidation:
53	- "Artifact-first debugging"
54	- "Verify hook output by inspecting files"
55	- "Filesystem-first debugging"
56	→ All express: "Observe outputs before editing code"
57
58	Use the most general formulation. Update the frequency table.
59
60	### Step 3: Detect Meta-Patterns
61
62	Critical step: Look at what the learnings cluster around.
63
64	If >50% of patterns relate to one topic (e.g., "hooks", "tracing", "async"):
65	→ That topic may need a dedicated skill rather than multiple rules
66	→ One skill compounds better than five rules
67
68	Ask yourself: "Is there a skill that would make all these rules unnecessary?"
69
70	### Step 4: Categorize (Decision Tree)
71
72	For each pattern, determine artifact type:
73
74	```
75	Is it a sequence of commands/steps?
76	→ YES → SKILL (executable > declarative)
77	→ NO ↓
78
79	Should it run automatically on an event (SessionEnd, PostToolUse, etc.)?
80	→ YES → HOOK (automatic > manual)
81	→ NO ↓
82
83	Is it "when X, do Y" or "never do X"?
84	→ YES → RULE
85	→ NO ↓
86
87	Does it enhance an existing agent workflow?
88	→ YES → AGENT UPDATE
89	→ NO → Skip (not worth capturing)
90	```
91
92	Artifact Type Examples:
93
94	\| Pattern \| Type \| Why \|
95	\|---------\|------\|-----\|
96	\| "Run linting before commit" \| Hook (PreToolUse) \| Automatic gate \|
97	\| "Extract learnings on session end" \| Hook (SessionEnd) \| Automatic trigger \|
98	\| "Debug hooks step by step" \| Skill \| Manual sequence \|
99	\| "Always pass IDs explicitly" \| Rule \| Heuristic \|
100
101	### Step 5: Apply Signal Thresholds
102
103	\| Occurrences \| Action \|
104	\|-------------\|--------\|
105	\| 1 \| Note but skip (unless critical failure) \|
106	\| 2 \| Consider - present to user \|
107	\| 3+ \| Strong signal - recommend creation \|
108	\| 4+ \| Definitely create \|
109
110	### Step 6: Propose Artifacts
111
112	Present each proposal in this format:
113
114	```markdown
115	---
116
117	## Pattern: [Generalized Name]
118
119	Signal: [N] sessions ([list session IDs])
120
121	Category: [debugging / reliability / workflow / etc.]
122
123	Artifact Type: Rule / Skill / Agent Update
124
125	Rationale: [Why this artifact type, why worth creating]
126
127	Draft Content:
128	\`\`\`markdown
129	[Actual content that would be written to file]
130	\`\`\`
131
132	File: `.claude/rules/[name].md` or `.claude/skills/[name]/SKILL.md`
133
134	---
135	```
136
137	Use `AskUserQuestion` to get approval for each artifact (or batch approval).
138
139	### Step 7: Create Approved Artifacts
140
141	#### For Rules:
142	```bash
143	# Write to rules directory
144	cat > $CLAUDE_PROJECT_DIR/.claude/rules/<name>.md << 'EOF'
145	# Rule Name
146
147	[Context: why this rule exists, based on N sessions]
148
149	## Pattern
150	[The reusable principle]
151
152	## DO
153	- [Concrete action]
154
155	## DON'T
156	- [Anti-pattern]
157
158	## Source Sessions
159	- [session-id-1]: [what happened]
160	- [session-id-2]: [what happened]
161	EOF
162	```
163
164	#### For Skills:
165	Create `.claude/skills/<name>/SKILL.md` with:
166	- Frontmatter (name, description, allowed-tools)
167	- When to Use
168	- Step-by-step instructions (executable)
169	- Examples from the learnings
170
171	Add triggers to `skill-rules.json` if appropriate.
172
173	#### For Hooks:
174	Create shell wrapper + TypeScript handler:
175
176	```bash
177	# Shell wrapper
178	cat > $CLAUDE_PROJECT_DIR/.claude/hooks/<name>.sh << 'EOF'
179	#!/bin/bash
180	set -e
181	cd "$CLAUDE_PROJECT_DIR/.claude/hooks"
182	cat \| node dist/<name>.mjs
183	EOF
184	chmod +x $CLAUDE_PROJECT_DIR/.claude/hooks/<name>.sh
185	```
186
187	Then create `src/<name>.ts`, build with esbuild, a

People who liked this also viewed

6 picks

Type

compound-learnings

482installs297forks

View on GitHub

How to install compound-learnings?

parcadei/continuous-claude-v3/compound-learnings

$npx -y skills add parcadei/Continuous-Claude-v3 --skill compound-learnings

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Compound Learnings
2
3	Transform ephemeral session learnings into permanent, compounding capabilities.
4
5	## When to Use
6
7	- "What should I learn from recent sessions?"
8	- "Improve my setup based on recent work"
9	- "Turn learnings into skills/rules"
10	- "What patterns should become permanent?"
11	- "Compound my learnings"
12
13	## Process
14
15	### Step 1: Gather Learnings
16
17	```bash
18	# List learnings (most recent first)
19	ls -t $CLAUDE_PROJECT_DIR/.claude/cache/learnings/*.md \| head -20
20
21	# Count total
22	ls $CLAUDE_PROJECT_DIR/.claude/cache/learnings/*.md \| wc -l
23	```
24
25	Read the most recent 5-10 files (or specify a date range).
26
27	### Step 2: Extract Patterns (Structured)
28
29	For each learnings file, extract entries from these specific sections:
30
31	\| Section Header \| What to Extract \|
32	\|----------------\|-----------------\|
33	\| `## Patterns` or `Reusable techniques` \| Direct candidates for rules \|
34	\| `Takeaway:` or `Actionable takeaway:` \| Decision heuristics \|
35	\| `## What Worked` \| Success patterns \|
36	\| `## What Failed` \| Anti-patterns (invert to rules) \|
37	\| `## Key Decisions` \| Design principles \|
38
39	Build a frequency table as you go:
40
41	```markdown
42	\| Pattern \| Sessions \| Category \|
43	\|---------\|----------\|----------\|
44	\| "Check artifacts before editing" \| abc, def, ghi \| debugging \|
45	\| "Pass IDs explicitly" \| abc, def, ghi, jkl \| reliability \|
46	```
47
48	### Step 2b: Consolidate Similar Patterns
49
50	Before counting, merge patterns that express the same principle:
51
52	Example consolidation:
53	- "Artifact-first debugging"
54	- "Verify hook output by inspecting files"
55	- "Filesystem-first debugging"
56	→ All express: "Observe outputs before editing code"
57
58	Use the most general formulation. Update the frequency table.
59
60	### Step 3: Detect Meta-Patterns
61
62	Critical step: Look at what the learnings cluster around.
63
64	If >50% of patterns relate to one topic (e.g., "hooks", "tracing", "async"):
65	→ That topic may need a dedicated skill rather than multiple rules
66	→ One skill compounds better than five rules
67
68	Ask yourself: "Is there a skill that would make all these rules unnecessary?"
69
70	### Step 4: Categorize (Decision Tree)
71
72	For each pattern, determine artifact type:
73
74	```
75	Is it a sequence of commands/steps?
76	→ YES → SKILL (executable > declarative)
77	→ NO ↓
78
79	Should it run automatically on an event (SessionEnd, PostToolUse, etc.)?
80	→ YES → HOOK (automatic > manual)
81	→ NO ↓
82
83	Is it "when X, do Y" or "never do X"?
84	→ YES → RULE
85	→ NO ↓
86
87	Does it enhance an existing agent workflow?
88	→ YES → AGENT UPDATE
89	→ NO → Skip (not worth capturing)
90	```
91
92	Artifact Type Examples:
93
94	\| Pattern \| Type \| Why \|
95	\|---------\|------\|-----\|
96	\| "Run linting before commit" \| Hook (PreToolUse) \| Automatic gate \|
97	\| "Extract learnings on session end" \| Hook (SessionEnd) \| Automatic trigger \|
98	\| "Debug hooks step by step" \| Skill \| Manual sequence \|
99	\| "Always pass IDs explicitly" \| Rule \| Heuristic \|
100
101	### Step 5: Apply Signal Thresholds
102
103	\| Occurrences \| Action \|
104	\|-------------\|--------\|
105	\| 1 \| Note but skip (unless critical failure) \|
106	\| 2 \| Consider - present to user \|
107	\| 3+ \| Strong signal - recommend creation \|
108	\| 4+ \| Definitely create \|
109
110	### Step 6: Propose Artifacts
111
112	Present each proposal in this format:
113
114	```markdown
115	---
116
117	## Pattern: [Generalized Name]
118
119	Signal: [N] sessions ([list session IDs])
120
121	Category: [debugging / reliability / workflow / etc.]
122
123	Artifact Type: Rule / Skill / Agent Update
124
125	Rationale: [Why this artifact type, why worth creating]
126
127	Draft Content:
128	\`\`\`markdown
129	[Actual content that would be written to file]
130	\`\`\`
131
132	File: `.claude/rules/[name].md` or `.claude/skills/[name]/SKILL.md`
133
134	---
135	```
136
137	Use `AskUserQuestion` to get approval for each artifact (or batch approval).
138
139	### Step 7: Create Approved Artifacts
140
141	#### For Rules:
142	```bash
143	# Write to rules directory
144	cat > $CLAUDE_PROJECT_DIR/.claude/rules/<name>.md << 'EOF'
145	# Rule Name
146
147	[Context: why this rule exists, based on N sessions]
148
149	## Pattern
150	[The reusable principle]
151
152	## DO
153	- [Concrete action]
154
155	## DON'T
156	- [Anti-pattern]
157
158	## Source Sessions
159	- [session-id-1]: [what happened]
160	- [session-id-2]: [what happened]
161	EOF
162	```
163
164	#### For Skills:
165	Create `.claude/skills/<name>/SKILL.md` with:
166	- Frontmatter (name, description, allowed-tools)
167	- When to Use
168	- Step-by-step instructions (executable)
169	- Examples from the learnings
170
171	Add triggers to `skill-rules.json` if appropriate.
172
173	#### For Hooks:
174	Create shell wrapper + TypeScript handler:
175
176	```bash
177	# Shell wrapper
178	cat > $CLAUDE_PROJECT_DIR/.claude/hooks/<name>.sh << 'EOF'
179	#!/bin/bash
180	set -e
181	cd "$CLAUDE_PROJECT_DIR/.claude/hooks"
182	cat \| node dist/<name>.mjs
183	EOF
184	chmod +x $CLAUDE_PROJECT_DIR/.claude/hooks/<name>.sh
185	```
186
187	Then create `src/<name>.ts`, build with esbuild, a

People who liked this also viewed

6 picks

Type