claude_skill

Name: claude_skill
Author: HKUST-KnowComp

85stars6forks

How to install claude_skill?

hkust-knowcomp/deeprefine-skill/claude_skill

$npx -y skills add HKUST-KnowComp/DeepRefine-Skill --skill claude_skill

Installs into the current project.

Summary

Claude Code adapter for the DeepRefine agent-native refinement loop. Use when the user invokes /deeprefine, or asks to refine, diagnose, review, or apply changes to a Graphify / LLM-Wiki knowledge graph. Must follow the canonical DeepRefine skill rules and stop for review before

Files · 1

View on GitHub

SKILL.md

1	# DeepRefine - Claude Code Adapter
2
3	This file is the Claude Code-specific entrypoint. It keeps the platform rules
4	small and loads longer DeepRefine procedure details only when needed:
5
6	- Full workflow, queue selection, refinement branch logic, and review rules:
7	[references/deeprefine-workflow.md](references/deeprefine-workflow.md)
8	- Verbatim judgement, abduction, and refinement prompts:
9	[references/llm-prompts.md](references/llm-prompts.md)
10	- Checklist, command sequence, trace schema, paths, and CLI mode:
11	[references/trace-and-commands.md](references/trace-and-commands.md)
12
13	Do not reimplement or shorten the algorithm from memory. Load the relevant
14	reference file before executing that part of the workflow.
15
16	## Claude Code Invocation
17
18	Trigger this skill when the user:
19
20	- explicitly invokes `/deeprefine`;
21	- asks to refine, improve, diagnose, repair, inspect, or review a Graphify
22	knowledge graph;
23	- asks to apply a previously reviewed DeepRefine refinement.
24
25	Run from the knowledge-base project root, where `graphify-out/graph.json`
26	exists. If the user is planning or asking how DeepRefine works, explain the
27	workflow and do not mutate files.
28
29	If `deeprefine` is unavailable, tell the user to install it:
30
31	```bash
32	pip install deeprefine-cli
33	```
34
35	For source development:
36
37	```bash
38	pip install -e /path/to/DeepRefine-Skill
39	```
40
41	## Hard Safety Policy
42
43	A normal `/deeprefine` invocation is dry-run only and MUST NEVER call
44	`deeprefine apply`.
45
46	The default workflow must stop after:
47
48	1. `deeprefine loop validate`
49	2. `deeprefine review`
50	3. showing the proposed actions and HIGH/MEDIUM/LOW review report to the user
51
52	Then ask for explicit approval.
53
54	Only if the user's next message explicitly says to approve/apply/write the graph
55	may you run:
56
57	```bash
58	deeprefine apply --trace-file ... --refinement-file ...
59	deeprefine loop finish --trace-file ... --refinement-file ...
60	```
61
62	Do not treat any of these as approval:
63
64	- generation of a `<refinement>` block;
65	- a valid `loop_trace_<query_id>.json`;
66	- a prior user message;
67	- a successful `deeprefine review`.
68
69	If the review contains LOW-confidence actions, use
70	`--allow-low-confidence` only when the user's current approval message
71	explicitly accepts that risk.
72
73	## Mode Selection
74
75	### Full workflow
76
77	Use for `/deeprefine`, or requests to refine/improve/fix the graph.
78
79	Follow the canonical reference in this order:
80
81	1. `references/deeprefine-workflow.md`
82	2. `references/llm-prompts.md` when producing tagged LLM outputs
83	3. `references/trace-and-commands.md` when writing traces or running commands
84
85	Do not copy only the latest query if pending history exists. Process all
86	unrefined history queries first, preserving canonical dedupe/order rules.
87
88	### Review only
89
90	Use when the user asks to review, audit, inspect, dry-run, check evidence, or
91	show what would change.
92
93	Run validation and review only:
94
95	```bash
96	deeprefine loop validate --trace-file ... --refinement-file ...
97	deeprefine review --trace-file ... --refinement-file ...
98	```
99
100	Show the HIGH/MEDIUM/LOW evidence report. Do not modify `graph.json`.
101
102	### Apply only
103
104	Use only when the user's current message explicitly approves a previously
105	reviewed refinement.
106
107	Before applying, verify that the trace and refinement file match
108	`references/trace-and-commands.md` and the review rules in
109	`references/deeprefine-workflow.md`. Then run:
110
111	```bash
112	deeprefine loop validate --trace-file ... --refinement-file ...
113	deeprefine apply --trace-file ... --refinement-file ...
114	deeprefine loop finish --trace-file ... --refinement-file ...
115	```
116
117	Use the LOW-confidence override only with explicit risk acknowledgement in the
118	same user message:
119
120	```bash
121	deeprefine apply --allow-low-confidence --trace-file ... --refinement-file ...
122	```
123
124	## Non-Negotiable Rules
125
126	These are restated here so Claude always sees the hard stops before loading any
127	reference.
128
129	Do not:
130
131	1. Run `deeprefine refine` unless the user explicitly asks for CLI/FAISS mode.
132	2. Call `deeprefine apply` without a valid `loop_trace_<query_id>.json`.
133	3. Call `deeprefine apply` before `deeprefine review` and explicit approval.
134	4. Ignore LOW-confidence review warnings without explicit risk acceptance.
135	5. Skip any hop's `<judge>Yes</judge>` / `<judge>No</judge>` judgement.
136	6. Skip error abduction when `len(interaction_history) > 1`.
137	7. Write `<refinement>` before abduction when refinement is required.
138	8. Hand-edit `graphify-out/graph.json` with Python or ad-hoc JSON patches.
139	9. Ignore pending history and refine only one latest query.
140	10. Invent a shorter pipeline such as "read file -> write refinemen

People who liked this also viewed

6 picks

Type

claude_skill

85stars6forks

View on GitHub

How to install claude_skill?

hkust-knowcomp/deeprefine-skill/claude_skill

$npx -y skills add HKUST-KnowComp/DeepRefine-Skill --skill claude_skill

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# DeepRefine - Claude Code Adapter
2
3	This file is the Claude Code-specific entrypoint. It keeps the platform rules
4	small and loads longer DeepRefine procedure details only when needed:
5
6	- Full workflow, queue selection, refinement branch logic, and review rules:
7	[references/deeprefine-workflow.md](references/deeprefine-workflow.md)
8	- Verbatim judgement, abduction, and refinement prompts:
9	[references/llm-prompts.md](references/llm-prompts.md)
10	- Checklist, command sequence, trace schema, paths, and CLI mode:
11	[references/trace-and-commands.md](references/trace-and-commands.md)
12
13	Do not reimplement or shorten the algorithm from memory. Load the relevant
14	reference file before executing that part of the workflow.
15
16	## Claude Code Invocation
17
18	Trigger this skill when the user:
19
20	- explicitly invokes `/deeprefine`;
21	- asks to refine, improve, diagnose, repair, inspect, or review a Graphify
22	knowledge graph;
23	- asks to apply a previously reviewed DeepRefine refinement.
24
25	Run from the knowledge-base project root, where `graphify-out/graph.json`
26	exists. If the user is planning or asking how DeepRefine works, explain the
27	workflow and do not mutate files.
28
29	If `deeprefine` is unavailable, tell the user to install it:
30
31	```bash
32	pip install deeprefine-cli
33	```
34
35	For source development:
36
37	```bash
38	pip install -e /path/to/DeepRefine-Skill
39	```
40
41	## Hard Safety Policy
42
43	A normal `/deeprefine` invocation is dry-run only and MUST NEVER call
44	`deeprefine apply`.
45
46	The default workflow must stop after:
47
48	1. `deeprefine loop validate`
49	2. `deeprefine review`
50	3. showing the proposed actions and HIGH/MEDIUM/LOW review report to the user
51
52	Then ask for explicit approval.
53
54	Only if the user's next message explicitly says to approve/apply/write the graph
55	may you run:
56
57	```bash
58	deeprefine apply --trace-file ... --refinement-file ...
59	deeprefine loop finish --trace-file ... --refinement-file ...
60	```
61
62	Do not treat any of these as approval:
63
64	- generation of a `<refinement>` block;
65	- a valid `loop_trace_<query_id>.json`;
66	- a prior user message;
67	- a successful `deeprefine review`.
68
69	If the review contains LOW-confidence actions, use
70	`--allow-low-confidence` only when the user's current approval message
71	explicitly accepts that risk.
72
73	## Mode Selection
74
75	### Full workflow
76
77	Use for `/deeprefine`, or requests to refine/improve/fix the graph.
78
79	Follow the canonical reference in this order:
80
81	1. `references/deeprefine-workflow.md`
82	2. `references/llm-prompts.md` when producing tagged LLM outputs
83	3. `references/trace-and-commands.md` when writing traces or running commands
84
85	Do not copy only the latest query if pending history exists. Process all
86	unrefined history queries first, preserving canonical dedupe/order rules.
87
88	### Review only
89
90	Use when the user asks to review, audit, inspect, dry-run, check evidence, or
91	show what would change.
92
93	Run validation and review only:
94
95	```bash
96	deeprefine loop validate --trace-file ... --refinement-file ...
97	deeprefine review --trace-file ... --refinement-file ...
98	```
99
100	Show the HIGH/MEDIUM/LOW evidence report. Do not modify `graph.json`.
101
102	### Apply only
103
104	Use only when the user's current message explicitly approves a previously
105	reviewed refinement.
106
107	Before applying, verify that the trace and refinement file match
108	`references/trace-and-commands.md` and the review rules in
109	`references/deeprefine-workflow.md`. Then run:
110
111	```bash
112	deeprefine loop validate --trace-file ... --refinement-file ...
113	deeprefine apply --trace-file ... --refinement-file ...
114	deeprefine loop finish --trace-file ... --refinement-file ...
115	```
116
117	Use the LOW-confidence override only with explicit risk acknowledgement in the
118	same user message:
119
120	```bash
121	deeprefine apply --allow-low-confidence --trace-file ... --refinement-file ...
122	```
123
124	## Non-Negotiable Rules
125
126	These are restated here so Claude always sees the hard stops before loading any
127	reference.
128
129	Do not:
130
131	1. Run `deeprefine refine` unless the user explicitly asks for CLI/FAISS mode.
132	2. Call `deeprefine apply` without a valid `loop_trace_<query_id>.json`.
133	3. Call `deeprefine apply` before `deeprefine review` and explicit approval.
134	4. Ignore LOW-confidence review warnings without explicit risk acceptance.
135	5. Skip any hop's `<judge>Yes</judge>` / `<judge>No</judge>` judgement.
136	6. Skip error abduction when `len(interaction_history) > 1`.
137	7. Write `<refinement>` before abduction when refinement is required.
138	8. Hand-edit `graphify-out/graph.json` with Python or ad-hoc JSON patches.
139	9. Ignore pending history and refine only one latest query.
140	10. Invent a shorter pipeline such as "read file -> write refinemen

People who liked this also viewed

6 picks

Type