cxas-protocol-robust-extraction

Name: cxas-protocol-robust-extraction
Author: GoogleCloudPlatform

78stars64forks

How to install cxas-protocol-robust-extraction?

googlecloudplatform/cxas-scrapi/cxas-protocol-robust-extraction

$npx -y skills add GoogleCloudPlatform/cxas-scrapi --skill cxas-protocol-robust-extraction

Installs into the current project.

Summary

A robust methodology for LLM-based requirements gathering and high-fidelity artifact generation. Employs 'Divide, Conquer, and Verify' tactics using specialized subagents, iterative exhaustion loops, and batched execution to ensure zero data loss.

Files · 1

View on GitHub

SKILL.md

1	# Robust Extraction Protocol
2
3	This protocol defines the standard operating procedure for extracting exhaustive
4	requirements (like subintents, CUJs, or logic rules) from large, complex, or
5	fragmented customer artifacts.
6
7	It prevents the common LLM pitfalls of "context drift" and "truncation" by
8	enforcing a strict "Divide, Conquer, and Verify" methodology.
9
10	## General Principles & Anti-Hallucination Guardrails
11
12	To ensure 100% coverage and prevent data loss due to tool limits or implicit
13	filtering, follow these principles across all phases:
14
15	* Quantify the Scope: Before spawning any subagents or starting
16	extraction, determine the exact total count of target items (files,
17	directories, database rows). Record this number as your "Success Target."
18	You must verify that the sum of items processed equals this target before
19	proceeding to consolidation.
20	* Coverage over Curation: Default to 100% extraction coverage. Never
21	assume the user only wants the "top" or "most interesting" items unless
22	explicitly instructed to apply a quality filter. A standard or repetitive
23	item is still data that must be reported.
24	* Circumvent Tool Caps: Be aware that search and listing tools often have
25	display limits (e.g., capped at 50 or 1000 results). If the expected scale
26	(from the Quantify step) exceeds the tool's limit, you must partition the
27	work (e.g., by alphabet or ID range) to ensure no items are hidden by the
28	tool's cap.
29	* Maintain Traceability: For every extracted requirement or item, record
30	the source file or location it was extracted from. This allows for easy
31	verification and provides context when reviewing the consolidated results.
32
33	## Core Directives
34
35	When tasked with comprehensive extraction or generation from a large corpus, you
36	MUST follow this four-phase methodology:
37
38	### Phase 1: Parallel Expert Discovery
39
40	Never use a single generalist agent or a single prompt to read all files.
41
42	1. Categorize the input artifacts (e.g., Code/ADK, Diagrams, Test Cases).
43	2. Spawn specialized expert subagents (e.g., `cxas-ingestor-adk`) in
44	parallel, providing each with only the context relevant to their expertise.
45	3. Consolidate their initial findings into a centralized list.
46
47	### Phase 2: The Iterative Exhaustion Loop
48
49	LLMs often miss items in a single pass of a large document. You must force them
50	to iterate.
51
52	1. Provide the current consolidated list of findings back to the expert
53	subagents.
54	2. Ask a direct question: *"Based on your artifacts, are there ANY MORE items
55	missing from this list? If yes, list them. If no, reply EXACTLY 'NO'."*
56	3. The Loop Rule: You MUST continue this loop, updating the consolidated
57	list each time, until ALL expert subagents unanimously reply with "NO".
58
59	### Phase 3: Logical Clustering
60
61	Once the exhaustive list is finalized (e.g., 100+ subintents), organize it.
62
63	1. Group the granular findings into high-level logical categories (Parent
64	CUJs).
65	2. Verify with the experts that the parent categories encompass all findings.
66
67	### Phase 4: Batched Execution & Verification
68
69	Never ask an LLM to generate 100+ complex artifacts (like conversational
70	transcripts) in a single prompt. It will hallucinate or truncate.
71
72	1. Batching: Divide the exhaustive list into small, manageable batches
73	(e.g., 10 batches of 10 items). Write these batches to temporary files.
74	2. Delegated Execution: Spawn a new subagent for each batch. Instruct
75	them to process only their assigned batch and write the output to a
76	specific file.
77	3. The Verification Gate: As the orchestrator, you MUST verify the output
78	of each subagent. Did they generate an output for every single item in
79	their batch?
80	4. If YES: Accept the batch.
81	5. If NO: Discard the output and respawn the subagent for that specific
82	batch with stronger steering instructions.
83	6. Consolidation: Only when all batches pass the Verification Gate, merge
84	them into the final, exhaustive deliverable.

People who liked this also viewed

6 picks

Type

cxas-protocol-robust-extraction

78stars64forks

View on GitHub

How to install cxas-protocol-robust-extraction?

googlecloudplatform/cxas-scrapi/cxas-protocol-robust-extraction

$npx -y skills add GoogleCloudPlatform/cxas-scrapi --skill cxas-protocol-robust-extraction

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Robust Extraction Protocol
2
3	This protocol defines the standard operating procedure for extracting exhaustive
4	requirements (like subintents, CUJs, or logic rules) from large, complex, or
5	fragmented customer artifacts.
6
7	It prevents the common LLM pitfalls of "context drift" and "truncation" by
8	enforcing a strict "Divide, Conquer, and Verify" methodology.
9
10	## General Principles & Anti-Hallucination Guardrails
11
12	To ensure 100% coverage and prevent data loss due to tool limits or implicit
13	filtering, follow these principles across all phases:
14
15	* Quantify the Scope: Before spawning any subagents or starting
16	extraction, determine the exact total count of target items (files,
17	directories, database rows). Record this number as your "Success Target."
18	You must verify that the sum of items processed equals this target before
19	proceeding to consolidation.
20	* Coverage over Curation: Default to 100% extraction coverage. Never
21	assume the user only wants the "top" or "most interesting" items unless
22	explicitly instructed to apply a quality filter. A standard or repetitive
23	item is still data that must be reported.
24	* Circumvent Tool Caps: Be aware that search and listing tools often have
25	display limits (e.g., capped at 50 or 1000 results). If the expected scale
26	(from the Quantify step) exceeds the tool's limit, you must partition the
27	work (e.g., by alphabet or ID range) to ensure no items are hidden by the
28	tool's cap.
29	* Maintain Traceability: For every extracted requirement or item, record
30	the source file or location it was extracted from. This allows for easy
31	verification and provides context when reviewing the consolidated results.
32
33	## Core Directives
34
35	When tasked with comprehensive extraction or generation from a large corpus, you
36	MUST follow this four-phase methodology:
37
38	### Phase 1: Parallel Expert Discovery
39
40	Never use a single generalist agent or a single prompt to read all files.
41
42	1. Categorize the input artifacts (e.g., Code/ADK, Diagrams, Test Cases).
43	2. Spawn specialized expert subagents (e.g., `cxas-ingestor-adk`) in
44	parallel, providing each with only the context relevant to their expertise.
45	3. Consolidate their initial findings into a centralized list.
46
47	### Phase 2: The Iterative Exhaustion Loop
48
49	LLMs often miss items in a single pass of a large document. You must force them
50	to iterate.
51
52	1. Provide the current consolidated list of findings back to the expert
53	subagents.
54	2. Ask a direct question: *"Based on your artifacts, are there ANY MORE items
55	missing from this list? If yes, list them. If no, reply EXACTLY 'NO'."*
56	3. The Loop Rule: You MUST continue this loop, updating the consolidated
57	list each time, until ALL expert subagents unanimously reply with "NO".
58
59	### Phase 3: Logical Clustering
60
61	Once the exhaustive list is finalized (e.g., 100+ subintents), organize it.
62
63	1. Group the granular findings into high-level logical categories (Parent
64	CUJs).
65	2. Verify with the experts that the parent categories encompass all findings.
66
67	### Phase 4: Batched Execution & Verification
68
69	Never ask an LLM to generate 100+ complex artifacts (like conversational
70	transcripts) in a single prompt. It will hallucinate or truncate.
71
72	1. Batching: Divide the exhaustive list into small, manageable batches
73	(e.g., 10 batches of 10 items). Write these batches to temporary files.
74	2. Delegated Execution: Spawn a new subagent for each batch. Instruct
75	them to process only their assigned batch and write the output to a
76	specific file.
77	3. The Verification Gate: As the orchestrator, you MUST verify the output
78	of each subagent. Did they generate an output for every single item in
79	their batch?
80	4. If YES: Accept the batch.
81	5. If NO: Discard the output and respawn the subagent for that specific
82	batch with stronger steering instructions.
83	6. Consolidation: Only when all batches pass the Verification Gate, merge
84	them into the final, exhaustive deliverable.

People who liked this also viewed

6 picks

Type