interleaved-thinking

Name: interleaved-thinking
Author: muratcankoylan

17kstars1.4kforks

How to install interleaved-thinking?

muratcankoylan/agent-skills-for-context-engineering/interleaved-thinking

$npx -y skills add muratcankoylan/Agent-Skills-for-Context-Engineering --skill interleaved-thinking

Installs into the current project.

Summary

Debug and optimize AI agents by analyzing reasoning traces, context degradation, tool confusion, instruction drift, repeated task failures, and performance regressions.

Files · 1

View on GitHub

SKILL.md

1	# Reasoning Trace Optimizer
2
3	Debug and optimize AI agents by analyzing their reasoning traces. This skill uses MiniMax M2.1's interleaved thinking to provide deep insight into agent decision-making and generate concrete improvements.
4
5	## When to Activate
6
7	- Agent reasoning traces need debugging, analysis, or prompt optimization
8	- Agent task fails and user wants to understand why
9	- User mentions "context degradation", "tool confusion", or "instruction drift"
10	- Request to improve agent performance or reduce errors
11	- User wants to generate shareable learnings from debugging sessions
12	- After repeated failures on similar tasks
13
14	## Core Concepts
15
16	### Interleaved Thinking
17
18	Unlike standard reasoning models that think once at the start, interleaved thinking allows reasoning BETWEEN each tool interaction. This is critical because:
19
20	1. Long-horizon tasks require maintaining focus across many turns
21	2. External perturbations (tool outputs, environment changes) need real-time adaptation
22	3. Debugging requires seeing HOW decisions were made, not just WHAT was output
23
24	### The Optimization Loop
25
26	```
27	Execute Agent → Capture Traces → Analyze Patterns → Optimize Prompt → Re-run
28	↑____________\|
29	```
30
31	Each iteration improves the prompt based on detected patterns until convergence.
32
33	### Pattern Detection
34
35	Common failure patterns the analyzer detects:
36
37	\| Pattern \| Description \|
38	\|---------\|-------------\|
39	\| `context_degradation` \| Model loses track of information over long contexts \|
40	\| `tool_confusion` \| Model misunderstands tool capabilities or outputs \|
41	\| `instruction_drift` \| Model gradually deviates from original instructions \|
42	\| `goal_abandonment` \| Model stops pursuing the original goal \|
43	\| `circular_reasoning` \| Model repeats similar actions without progress \|
44	\| `premature_conclusion` \| Model concludes before completing the task \|
45
46	## Usage Modes
47
48	### Mode 1: M2.1 Agent Debugging
49
50	Run a task through M2.1 and analyze its reasoning:
51
52	```python
53	from reasoning_trace_optimizer import TraceCapture, TraceAnalyzer
54
55	capture = TraceCapture()
56	trace = capture.run(
57	task="Search for Python tutorials and summarize them",
58	system_prompt="You are a research assistant.",
59	tools=[search_tool],
60	tool_executor=execute_search
61	)
62
63	analyzer = TraceAnalyzer()
64	analysis = analyzer.analyze(trace)
65
66	print(f"Score: {analysis.overall_score}/100")
67	for pattern in analysis.patterns:
68	print(f"Found: {pattern.type.value} - {pattern.suggestion}")
69	```
70
71	### Mode 2: Full Optimization Loop
72
73	Automatically iterate until the prompt is optimized:
74
75	```python
76	from reasoning_trace_optimizer import OptimizationLoop, LoopConfig
77
78	config = LoopConfig(
79	max_iterations=5,
80	min_score_threshold=80.0,
81	)
82
83	loop = OptimizationLoop(config=config)
84	result = loop.run(
85	task="Analyze this codebase and suggest improvements",
86	initial_prompt="You are a code reviewer.",
87	tools=[read_file_tool, search_tool],
88	tool_executor=execute_tool
89	)
90
91	print(f"Improved: {result.initial_score} → {result.final_score}")
92	print(f"Final prompt:\n{result.final_prompt}")
93	```
94
95	### Mode 3: Universal Session Analysis
96
97	Analyze any agent's previous thinking (works with Claude, GPT, etc.):
98
99	When this skill is activated in Claude Code, it can analyze the current session's thinking blocks to identify issues and suggest improvements.
100
101	```
102	/reasoning-trace-optimizer analyze-session
103	```
104
105	### Mode 4: Generate Shareable Skills
106
107	Convert optimization learnings into reusable Agent Skills:
108
109	```python
110	from reasoning_trace_optimizer import SkillGenerator
111
112	generator = SkillGenerator()
113	skill_path = generator.generate(
114	result=loop_result,
115	skill_name="web-search-best-practices",
116	output_dir="./skills"
117	)
118	```
119
120	## CLI Commands
121
122	```bash
123	# Capture reasoning trace
124	rto capture "Search for Python tutorials" -s "You are a helpful assistant."
125
126	# Analyze a task
127	rto analyze "Debug this code" -o analysis.txt
128
129	# Run optimization loop
130	rto optimize "Research AI papers" --max-iterations 5 --generate-skill
131
132	# Generate skill from artifacts
133	rto generate-skill my-skill-name --artifacts-dir ./optimization_artifacts
134	```
135
136	## Integration with Claude Code
137
138	### Auto-trigger on Failure
139
140	Add to your hooks to automatically analyze failures:
141
142	```json
143	{
144	"hooks": {
145	"post_tool_error": {
146	"command": "rto analyze-session --last-error"
147	}
148	}
149	}
150	```
151
152	### On-demand Analysis
153
154	Use the slash command to analyze current session:
155
156	```
157	/reasoning-trace-optimizer
158	```
159
160	This will:
161	1. Extract thinking blocks from the current session
162	2. Identify patterns and issues
163	3. Suggest prompt improvements
164	4. Optionally update the system prompt
165
166	## Guidelines
167
168	1. Preserve full context: M2.1 requires full response history including thinking blocks for o

People who liked this also viewed

6 picks

Type

interleaved-thinking

17kstars1.4kforks

View on GitHub

How to install interleaved-thinking?

muratcankoylan/agent-skills-for-context-engineering/interleaved-thinking

$npx -y skills add muratcankoylan/Agent-Skills-for-Context-Engineering --skill interleaved-thinking

Installs into the current project.

Summary

Debug and optimize AI agents by analyzing reasoning traces, context degradation, tool confusion, instruction drift, repeated task failures, and performance regressions.

Files · 1

View on GitHub

SKILL.md

1	# Reasoning Trace Optimizer
2
3	Debug and optimize AI agents by analyzing their reasoning traces. This skill uses MiniMax M2.1's interleaved thinking to provide deep insight into agent decision-making and generate concrete improvements.
4
5	## When to Activate
6
7	- Agent reasoning traces need debugging, analysis, or prompt optimization
8	- Agent task fails and user wants to understand why
9	- User mentions "context degradation", "tool confusion", or "instruction drift"
10	- Request to improve agent performance or reduce errors
11	- User wants to generate shareable learnings from debugging sessions
12	- After repeated failures on similar tasks
13
14	## Core Concepts
15
16	### Interleaved Thinking
17
18	Unlike standard reasoning models that think once at the start, interleaved thinking allows reasoning BETWEEN each tool interaction. This is critical because:
19
20	1. Long-horizon tasks require maintaining focus across many turns
21	2. External perturbations (tool outputs, environment changes) need real-time adaptation
22	3. Debugging requires seeing HOW decisions were made, not just WHAT was output
23
24	### The Optimization Loop
25
26	```
27	Execute Agent → Capture Traces → Analyze Patterns → Optimize Prompt → Re-run
28	↑____________\|
29	```
30
31	Each iteration improves the prompt based on detected patterns until convergence.
32
33	### Pattern Detection
34
35	Common failure patterns the analyzer detects:
36
37	\| Pattern \| Description \|
38	\|---------\|-------------\|
39	\| `context_degradation` \| Model loses track of information over long contexts \|
40	\| `tool_confusion` \| Model misunderstands tool capabilities or outputs \|
41	\| `instruction_drift` \| Model gradually deviates from original instructions \|
42	\| `goal_abandonment` \| Model stops pursuing the original goal \|
43	\| `circular_reasoning` \| Model repeats similar actions without progress \|
44	\| `premature_conclusion` \| Model concludes before completing the task \|
45
46	## Usage Modes
47
48	### Mode 1: M2.1 Agent Debugging
49
50	Run a task through M2.1 and analyze its reasoning:
51
52	```python
53	from reasoning_trace_optimizer import TraceCapture, TraceAnalyzer
54
55	capture = TraceCapture()
56	trace = capture.run(
57	task="Search for Python tutorials and summarize them",
58	system_prompt="You are a research assistant.",
59	tools=[search_tool],
60	tool_executor=execute_search
61	)
62
63	analyzer = TraceAnalyzer()
64	analysis = analyzer.analyze(trace)
65
66	print(f"Score: {analysis.overall_score}/100")
67	for pattern in analysis.patterns:
68	print(f"Found: {pattern.type.value} - {pattern.suggestion}")
69	```
70
71	### Mode 2: Full Optimization Loop
72
73	Automatically iterate until the prompt is optimized:
74
75	```python
76	from reasoning_trace_optimizer import OptimizationLoop, LoopConfig
77
78	config = LoopConfig(
79	max_iterations=5,
80	min_score_threshold=80.0,
81	)
82
83	loop = OptimizationLoop(config=config)
84	result = loop.run(
85	task="Analyze this codebase and suggest improvements",
86	initial_prompt="You are a code reviewer.",
87	tools=[read_file_tool, search_tool],
88	tool_executor=execute_tool
89	)
90
91	print(f"Improved: {result.initial_score} → {result.final_score}")
92	print(f"Final prompt:\n{result.final_prompt}")
93	```
94
95	### Mode 3: Universal Session Analysis
96
97	Analyze any agent's previous thinking (works with Claude, GPT, etc.):
98
99	When this skill is activated in Claude Code, it can analyze the current session's thinking blocks to identify issues and suggest improvements.
100
101	```
102	/reasoning-trace-optimizer analyze-session
103	```
104
105	### Mode 4: Generate Shareable Skills
106
107	Convert optimization learnings into reusable Agent Skills:
108
109	```python
110	from reasoning_trace_optimizer import SkillGenerator
111
112	generator = SkillGenerator()
113	skill_path = generator.generate(
114	result=loop_result,
115	skill_name="web-search-best-practices",
116	output_dir="./skills"
117	)
118	```
119
120	## CLI Commands
121
122	```bash
123	# Capture reasoning trace
124	rto capture "Search for Python tutorials" -s "You are a helpful assistant."
125
126	# Analyze a task
127	rto analyze "Debug this code" -o analysis.txt
128
129	# Run optimization loop
130	rto optimize "Research AI papers" --max-iterations 5 --generate-skill
131
132	# Generate skill from artifacts
133	rto generate-skill my-skill-name --artifacts-dir ./optimization_artifacts
134	```
135
136	## Integration with Claude Code
137
138	### Auto-trigger on Failure
139
140	Add to your hooks to automatically analyze failures:
141
142	```json
143	{
144	"hooks": {
145	"post_tool_error": {
146	"command": "rto analyze-session --last-error"
147	}
148	}
149	}
150	```
151
152	### On-demand Analysis
153
154	Use the slash command to analyze current session:
155
156	```
157	/reasoning-trace-optimizer
158	```
159
160	This will:
161	1. Extract thinking blocks from the current session
162	2. Identify patterns and issues
163	3. Suggest prompt improvements
164	4. Optionally update the system prompt
165
166	## Guidelines
167
168	1. Preserve full context: M2.1 requires full response history including thinking blocks for o

People who liked this also viewed

6 picks

Type