recursive-decomposition

Name: recursive-decomposition
Author: massimodeluisa

164installs2forks

How to install recursive-decomposition?

massimodeluisa/recursive-decomposition-skill/recursive-decomposition

$npx -y skills add massimodeluisa/recursive-decomposition-skill --skill recursive-decomposition

Installs into the current project.

Summary

Based on the Recursive Language Models (RLM) research by Zhang, Kraska, and Khattab (2025), this skill provides strategies for handling tasks that exceed comfortable context limits through programmatic decomposition and recursive self-invocation. Triggers on phrases like "analyze

Files · 1

View on GitHub

SKILL.md

1	# Recursive Decomposition Guidelines
2
3	## References
4
5	Consult these resources as needed:
6
7	- ./references/rlm-strategies.md -- Detailed decomposition patterns from the RLM paper
8	- ./references/cost-analysis.md -- When to apply recursive vs. direct approaches
9	- ./references/codebase-analysis.md -- Full walkthrough of codebase-wide analysis
10	- ./references/document-aggregation.md -- Multi-document information extraction
11
12	## Core Principles
13
14	CRITICAL: Treat inputs as environmental variables, not immediate context.
15
16	Most tasks fail when context is overloaded. Instead of loading entire contexts into the processing window, treat inputs as environmental variables accessible through code execution. Decompose problems recursively, process segments independently, and aggregate results programmatically.
17
18	Progressive Disclosure: Load information only when necessary. Start high-level to map the territory, then dive deep into specific areas.
19
20	### When Recursive Decomposition is Required
21
22	- Tasks involving 10+ files
23	- Input exceeding ~50k tokens where single-prompt context is insufficient
24	- Multi-hop questions requiring evidence from multiple scattered sources
25	- Codebase-wide pattern analysis or migration planning
26
27	### When Direct Processing Works
28
29	- Small contexts (<30k tokens)
30	- Single file analysis
31	- Linear complexity tasks with manageable inputs
32
33	## Operational Rules
34
35	- Always identify the search space size first.
36	- Always use `grep` or `glob` before `view_file` on directories.
37	- Always partition lists > 10 items into batches.
38	- Never read more than 5 files into context without a specific plan.
39	- Verify synthesized answers by spot-checking source material.
40	- Mitigate "context rot" by verifying answers on smaller windows.
41	- Treat yourself as an autonomous agent, not just a passive responder.
42
43	## Large File Handling Protocols
44
45	CRITICAL: Do NOT read large files directly into context.
46
47	1. Check Size First: Always run `wc -l` (lines) or `ls -lh` (size) before `view_file`.
48	2. Hard Limits:
49	* Text/Code: > 2,000 lines or > 50KB -> MUST use `view_file` with `start_line`/`end_line` or `head`/`tail`.
50	* PDFs: > 30MB or > 100 pages -> MUST be split or processed by metadata only.
51	3. Strategy:
52	* For code: Read definitions first (`grep -n "function" ...`) then read specific bodies.
53	* For text: Read Table of Contents or Abstract first.
54
55	## Tool Preferences
56
57	- `grep` / `glob` not `ls -R` (unless mapping structure).
58	- `view_file` with line ranges (offset/limit) not full file reads for huge files.
59	- `wc -l` / `ls -lh` before reading unknown files.
60	- `run_command` (grep) not `read_file` for searching.
61	- `task` tool for sub-agents (recurse).
62
63	## Empowering Agentic Behavior
64
65	To maximize effectiveness:
66
67	- Self-Correction: Always verify your own work. If a result seems empty or wrong, debug the approach (e.g., check grep arguments) before giving up.
68	- Aggressive Context Management: Regularly clear irrelevant history. Don't let the context window rot with dead ends.
69	- Plan First: For any task > 3 steps, write a mini-plan.
70	- Safe YOLO Mode: When appropriate (e.g., read-only searches), proceed with confidence without asking for permission on every single step, but stop for critical actions.
71
72	## Cost-Performance Tradeoffs
73
74	- Smaller contexts: Direct processing may be more efficient.
75	- Larger contexts: Recursive decomposition becomes necessary.
76	- Threshold: Consider decomposition when inputs exceed ~30k tokens or span 10+ files.
77
78	Balance thoroughness against computational cost. For time-sensitive tasks, apply aggressive filtering. For comprehensive analysis, prefer exhaustive decomposition.
79
80	## Anti-Patterns to Avoid
81
82	- Excessive sub-calling: Avoid redundant queries over the same content.
83	- Premature decomposition: Simple tasks don't need recursive strategies.
84	- Lost context: Ensure sub-agents have sufficient context for their sub-tasks.
85	- Unverified synthesis: Always spot-check aggregated results.
86
87	## Scalability (Chunking & filtering)
88
89	### 1. Filter Before Deep Analysis
90
91	Narrow the search space before detailed processing:
92
93	```
94	# Instead of reading all files into context:
95	1. Use Grep/Glob to identify candidate files by pattern
96	2. Filter candidates using domain-specific keywords
97	3. Only deeply analyze the filtered subset
98	```
99
100	Apply model priors about domain terminology to construct effective filters. For

People who liked this also viewed

6 picks

Type

recursive-decomposition

164installs2forks

View on GitHub

How to install recursive-decomposition?

massimodeluisa/recursive-decomposition-skill/recursive-decomposition

$npx -y skills add massimodeluisa/recursive-decomposition-skill --skill recursive-decomposition

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Recursive Decomposition Guidelines
2
3	## References
4
5	Consult these resources as needed:
6
7	- ./references/rlm-strategies.md -- Detailed decomposition patterns from the RLM paper
8	- ./references/cost-analysis.md -- When to apply recursive vs. direct approaches
9	- ./references/codebase-analysis.md -- Full walkthrough of codebase-wide analysis
10	- ./references/document-aggregation.md -- Multi-document information extraction
11
12	## Core Principles
13
14	CRITICAL: Treat inputs as environmental variables, not immediate context.
15
16	Most tasks fail when context is overloaded. Instead of loading entire contexts into the processing window, treat inputs as environmental variables accessible through code execution. Decompose problems recursively, process segments independently, and aggregate results programmatically.
17
18	Progressive Disclosure: Load information only when necessary. Start high-level to map the territory, then dive deep into specific areas.
19
20	### When Recursive Decomposition is Required
21
22	- Tasks involving 10+ files
23	- Input exceeding ~50k tokens where single-prompt context is insufficient
24	- Multi-hop questions requiring evidence from multiple scattered sources
25	- Codebase-wide pattern analysis or migration planning
26
27	### When Direct Processing Works
28
29	- Small contexts (<30k tokens)
30	- Single file analysis
31	- Linear complexity tasks with manageable inputs
32
33	## Operational Rules
34
35	- Always identify the search space size first.
36	- Always use `grep` or `glob` before `view_file` on directories.
37	- Always partition lists > 10 items into batches.
38	- Never read more than 5 files into context without a specific plan.
39	- Verify synthesized answers by spot-checking source material.
40	- Mitigate "context rot" by verifying answers on smaller windows.
41	- Treat yourself as an autonomous agent, not just a passive responder.
42
43	## Large File Handling Protocols
44
45	CRITICAL: Do NOT read large files directly into context.
46
47	1. Check Size First: Always run `wc -l` (lines) or `ls -lh` (size) before `view_file`.
48	2. Hard Limits:
49	* Text/Code: > 2,000 lines or > 50KB -> MUST use `view_file` with `start_line`/`end_line` or `head`/`tail`.
50	* PDFs: > 30MB or > 100 pages -> MUST be split or processed by metadata only.
51	3. Strategy:
52	* For code: Read definitions first (`grep -n "function" ...`) then read specific bodies.
53	* For text: Read Table of Contents or Abstract first.
54
55	## Tool Preferences
56
57	- `grep` / `glob` not `ls -R` (unless mapping structure).
58	- `view_file` with line ranges (offset/limit) not full file reads for huge files.
59	- `wc -l` / `ls -lh` before reading unknown files.
60	- `run_command` (grep) not `read_file` for searching.
61	- `task` tool for sub-agents (recurse).
62
63	## Empowering Agentic Behavior
64
65	To maximize effectiveness:
66
67	- Self-Correction: Always verify your own work. If a result seems empty or wrong, debug the approach (e.g., check grep arguments) before giving up.
68	- Aggressive Context Management: Regularly clear irrelevant history. Don't let the context window rot with dead ends.
69	- Plan First: For any task > 3 steps, write a mini-plan.
70	- Safe YOLO Mode: When appropriate (e.g., read-only searches), proceed with confidence without asking for permission on every single step, but stop for critical actions.
71
72	## Cost-Performance Tradeoffs
73
74	- Smaller contexts: Direct processing may be more efficient.
75	- Larger contexts: Recursive decomposition becomes necessary.
76	- Threshold: Consider decomposition when inputs exceed ~30k tokens or span 10+ files.
77
78	Balance thoroughness against computational cost. For time-sensitive tasks, apply aggressive filtering. For comprehensive analysis, prefer exhaustive decomposition.
79
80	## Anti-Patterns to Avoid
81
82	- Excessive sub-calling: Avoid redundant queries over the same content.
83	- Premature decomposition: Simple tasks don't need recursive strategies.
84	- Lost context: Ensure sub-agents have sufficient context for their sub-tasks.
85	- Unverified synthesis: Always spot-check aggregated results.
86
87	## Scalability (Chunking & filtering)
88
89	### 1. Filter Before Deep Analysis
90
91	Narrow the search space before detailed processing:
92
93	```
94	# Instead of reading all files into context:
95	1. Use Grep/Glob to identify candidate files by pattern
96	2. Filter candidates using domain-specific keywords
97	3. Only deeply analyze the filtered subset
98	```
99
100	Apply model priors about domain terminology to construct effective filters. For

People who liked this also viewed

6 picks

Type