skillward-audit-skill

Name: skillward-audit-skill
Author: Fangcun-AI

131stars4forks

How to install skillward-audit-skill?

fangcun-ai/skillward/skillward-audit-skill

$npx -y skills add Fangcun-AI/SkillWard --skill skillward-audit-skill

Installs into the current project.

Summary

Security-audit a third-party skill bundle (folder with SKILL.md, or .zip / .tar.gz archive) before installing it, using the SkillWard cloud scanner. Trigger on any natural safety question about an external skill — "is this skill safe?", "audit / scan / review this skill", "should

Files · 1

View on GitHub

SKILL.md

1	# SkillWard Audit
2
3	Thin client for the SkillWard public scanning service at
4	`https://skillward.fangcunleap.com`. Given a local skill bundle (folder with
5	`SKILL.md`, or a `.zip` / `.tar.gz` archive), this skill packages it, uploads it
6	to the remote service, streams scan progress, and reports the verdict.
7
8	## When to use
9
10	- Explicit invocation — user asks to "audit", "scan", "review", or "check
11	the safety of" a skill bundle.
12	- Natural safety question about an external skill — even without the words
13	"audit" or "SkillWard", treat any of these as an invocation request when a
14	skill path / archive is in scope:
15	- "Is this skill safe?" / "Is it safe to install?"
16	- "Should I install this?" / "Should I trust this?"
17	- "Any malware / backdoors / suspicious behavior in here?"
18	- "Is this trustworthy?" / "Does this look sketchy?"
19	- "Anything I should know before I install this?"
20	- User references a skill bundle path — a folder containing `SKILL.md`, or
21	a `.zip` / `.tar.gz` / `.tgz` archive — in a safety-adjacent context.
22	- Pre-install safety check — user is about to install a third-party /
23	downloaded skill and wants a security pass first.
24	- Source mentioned — user mentions ClawHub, OpenClaw, GitHub, or another
25	third-party source for a skill bundle in a safety context.
26
27	## When NOT to use
28
29	- The user is authoring their own skill and asking for code review,
30	feedback, or "does this look right?" on something they just wrote in this
31	session. Do a normal code review instead — SkillWard is for vetting
32	external code the user did not write.
33	- The user just edited the skill in this session — what they want is
34	review of their own changes, not a third-party scanner.
35	- Built-in / first-party skills — Claude's own bundled skills don't need
36	third-party scanning.
37	- Non-skill inputs — a regular codebase, Python package, generic repo,
38	or arbitrary file that isn't shaped like a skill bundle. Decline and
39	explain.
40	- General malware scanning of arbitrary files — decline; this tool only
41	understands skill bundles.
42
43	## Gathering input
44
45	1. User gave a path — verify it exists. If it's a folder, confirm it contains
46	`SKILL.md` (or a nested folder that does). If it's an archive, accept `.zip`
47	/ `.tar.gz` / `.tgz`.
48	2. User said "this skill" / "the skill" without a path — ask one clarifying
49	question: "Which folder or archive should I scan? Please give me a path (e.g.
50	`C:\path\to\skill-folder` or `./foo.zip`)."
51	3. User gave a URL or GitHub link — v1 does not support remote fetching.
52	Reply: "I can't fetch URLs directly for auditing. Please `git clone` the repo
53	or download the release zip first, then point me at the local path."
54
55	## Choosing language
56
57	Inspect the user's most recent 2–3 messages. If any contain CJK (Chinese /
58	Japanese / Korean) characters — Unicode range `U+4E00–U+9FFF` — pass
59	`--lang zh`. Otherwise pass `--lang en`. Default to `en` when ambiguous.
60
61	The report JSON always contains both `text` (Chinese) and `text_en` (English)
62	in `warnings[]`; use whichever matches `--lang` when summarizing back to the
63	user.
64
65	## Choosing scan depth
66
67	Default: `--depth sandbox` (static + LLM + server-side Docker sandbox, no
68	after-tool inspection). Override only when the user signals otherwise:
69
70	- "quick" / "fast" / "just static" / "快速" / "只要静态" → `--depth static`
71	(seconds; skips sandbox)
72	- "deep" / "thorough" / "after-tool" / "深度" / "彻底" → `--depth deep`
73	(3-15 min; full pipeline incl. after-tool capability check)
74
75	When running `deep`, warn the user upfront: "Deep mode runs the Docker
76	sandbox plus after-tool capability analysis — this typically takes 3–10 minutes.
77	Starting now."
78
79	## How to invoke
80
81	### ⚠️ Required — set the tool timeout explicitly in every call
82
83	Do not rely on your exec / bash tool's default timeout. Defaults vary
84	across harnesses and installations (Claude Code default is 120 s; OpenClaw
85	default is 1800 s but some deployments override it down to 300 s or less).
86	A too-short default will SIGKILL `scan.py` mid-scan, the SSE stream is cut
87	before the server emits the final `report` event, and nothing gets saved.
88
89	Pass the timeout explicitly on every invocation. Match the depth you're
90	running:
91
92	\| Depth \| Claude Code (`timeout`, ms) \| OpenClaw (`timeoutSec`, s) \| Typical duration \|
93	\| --------- \| --------------------------- \| -----

People who liked this also viewed

6 picks

Type

skillward-audit-skill

131stars4forks

View on GitHub

How to install skillward-audit-skill?

fangcun-ai/skillward/skillward-audit-skill

$npx -y skills add Fangcun-AI/SkillWard --skill skillward-audit-skill

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# SkillWard Audit
2
3	Thin client for the SkillWard public scanning service at
4	`https://skillward.fangcunleap.com`. Given a local skill bundle (folder with
5	`SKILL.md`, or a `.zip` / `.tar.gz` archive), this skill packages it, uploads it
6	to the remote service, streams scan progress, and reports the verdict.
7
8	## When to use
9
10	- Explicit invocation — user asks to "audit", "scan", "review", or "check
11	the safety of" a skill bundle.
12	- Natural safety question about an external skill — even without the words
13	"audit" or "SkillWard", treat any of these as an invocation request when a
14	skill path / archive is in scope:
15	- "Is this skill safe?" / "Is it safe to install?"
16	- "Should I install this?" / "Should I trust this?"
17	- "Any malware / backdoors / suspicious behavior in here?"
18	- "Is this trustworthy?" / "Does this look sketchy?"
19	- "Anything I should know before I install this?"
20	- User references a skill bundle path — a folder containing `SKILL.md`, or
21	a `.zip` / `.tar.gz` / `.tgz` archive — in a safety-adjacent context.
22	- Pre-install safety check — user is about to install a third-party /
23	downloaded skill and wants a security pass first.
24	- Source mentioned — user mentions ClawHub, OpenClaw, GitHub, or another
25	third-party source for a skill bundle in a safety context.
26
27	## When NOT to use
28
29	- The user is authoring their own skill and asking for code review,
30	feedback, or "does this look right?" on something they just wrote in this
31	session. Do a normal code review instead — SkillWard is for vetting
32	external code the user did not write.
33	- The user just edited the skill in this session — what they want is
34	review of their own changes, not a third-party scanner.
35	- Built-in / first-party skills — Claude's own bundled skills don't need
36	third-party scanning.
37	- Non-skill inputs — a regular codebase, Python package, generic repo,
38	or arbitrary file that isn't shaped like a skill bundle. Decline and
39	explain.
40	- General malware scanning of arbitrary files — decline; this tool only
41	understands skill bundles.
42
43	## Gathering input
44
45	1. User gave a path — verify it exists. If it's a folder, confirm it contains
46	`SKILL.md` (or a nested folder that does). If it's an archive, accept `.zip`
47	/ `.tar.gz` / `.tgz`.
48	2. User said "this skill" / "the skill" without a path — ask one clarifying
49	question: "Which folder or archive should I scan? Please give me a path (e.g.
50	`C:\path\to\skill-folder` or `./foo.zip`)."
51	3. User gave a URL or GitHub link — v1 does not support remote fetching.
52	Reply: "I can't fetch URLs directly for auditing. Please `git clone` the repo
53	or download the release zip first, then point me at the local path."
54
55	## Choosing language
56
57	Inspect the user's most recent 2–3 messages. If any contain CJK (Chinese /
58	Japanese / Korean) characters — Unicode range `U+4E00–U+9FFF` — pass
59	`--lang zh`. Otherwise pass `--lang en`. Default to `en` when ambiguous.
60
61	The report JSON always contains both `text` (Chinese) and `text_en` (English)
62	in `warnings[]`; use whichever matches `--lang` when summarizing back to the
63	user.
64
65	## Choosing scan depth
66
67	Default: `--depth sandbox` (static + LLM + server-side Docker sandbox, no
68	after-tool inspection). Override only when the user signals otherwise:
69
70	- "quick" / "fast" / "just static" / "快速" / "只要静态" → `--depth static`
71	(seconds; skips sandbox)
72	- "deep" / "thorough" / "after-tool" / "深度" / "彻底" → `--depth deep`
73	(3-15 min; full pipeline incl. after-tool capability check)
74
75	When running `deep`, warn the user upfront: "Deep mode runs the Docker
76	sandbox plus after-tool capability analysis — this typically takes 3–10 minutes.
77	Starting now."
78
79	## How to invoke
80
81	### ⚠️ Required — set the tool timeout explicitly in every call
82
83	Do not rely on your exec / bash tool's default timeout. Defaults vary
84	across harnesses and installations (Claude Code default is 120 s; OpenClaw
85	default is 1800 s but some deployments override it down to 300 s or less).
86	A too-short default will SIGKILL `scan.py` mid-scan, the SSE stream is cut
87	before the server emits the final `report` event, and nothing gets saved.
88
89	Pass the timeout explicitly on every invocation. Match the depth you're
90	running:
91
92	\| Depth \| Claude Code (`timeout`, ms) \| OpenClaw (`timeoutSec`, s) \| Typical duration \|
93	\| --------- \| --------------------------- \| -----

People who liked this also viewed

6 picks

Type