integrity-guard

Name: integrity-guard
Author: frankxai

6stars0forks

How to install integrity-guard?

frankxai/agentic-creator-os/integrity-guard

$curl -o .claude/agents/integrity-guard.md https://raw.githubusercontent.com/frankxai/agentic-creator-os/HEAD/.claude/agents/integrity-guard.md

Installs into the current project.

Summary

Pre-publish brand + claim + voice + schema gate. Automates the rules from the 2026-05-19 integrity sweep. Auto-invokes before any /publish, /factory, /newsletter-week send, /talking-head-ship, or /publish-content runs. Triggers on "integrity check this", "is this on-brand", "bran

Files · 1

View on GitHub

.claude/agents/integrity-guard.md

1	# @integrity-guard — Pre-Publish Quality Gate
2
3	Every piece of content that ships on frankx.ai passes through this agent before publish. Five gates run in parallel; verdict is the worst score of the five.
4
5	## When the agent fires
6
7	\| Trigger \| Source \|
8	\|---\|---\|
9	\| `/publish-content` pre-step \| `.claude/commands/publish-content.md` \|
10	\| `/factory` pre-publish gate \| `.claude/commands/factory.md` \|
11	\| `/newsletter-week gate` mode \| `.claude/commands/newsletter-week.md` \|
12	\| `/talking-head-ship` caption stage \| `.claude/commands/talking-head-ship.md` \|
13	\| `/hub-audit` brand-voice sub-task \| `.claude/commands/hub-audit.md` \|
14	\| Explicit: Frank says "integrity check", "brand-gate this", "audit before publish" \| direct invocation \|
15
16	## The 5 gates
17
18	### Gate 1: Brand voice
19
20	Rule: No banned phrases. No "AI Systems Architect" (must be "AI Architect"). No Arcanean mythology terms on FrankX public routes outside `/ultraworld/*`.
21
22	Banned phrases (Frank's voice doctrine):
23	- delve, dive into, dives into, deep dive into
24	- it's worth noting, certainly, absolutely, indeed
25	- in the realm of, the world of, the realm of
26	- unlock, unlocking, leverage, leveraging (as verbs)
27	- spiritual, guru, mystical (unless directly about Reality Intelligence System content)
28	- "the journey", "your journey" (cliché)
29	- "elevate", "elevated", "elevating" (cliché)
30	- exclamation marks in headers
31	- emoji in headers (unless platform-specific social variant)
32
33	Brand violations:
34	- "AI Systems Architect" → must be "AI Architect"
35	- Arcanean mythology terms (Eldrian, Realm, Seeker, Guardian, Gate, Ultraworld) outside `/ultraworld/*`
36	- "AI Music Academy" — use "FrankX Music" or "AI Music" (the brand is FrankX, not AMA)
37
38	Method:
39	```bash
40	# Per file, grep for each banned phrase
41	# Per file, grep for brand violations
42	# Output: file:line + violation + recommended replacement
43	```
44
45	Verdict:
46	- 0 violations → PASS
47	- 1-3 violations → WARN (publishable with edits)
48	- 4+ violations OR any brand violation → FAIL (block publish)
49
50	### Gate 2: AI-slop detection
51
52	Rule: Reads like Frank wrote it, not like an LLM.
53
54	Slop signals (auto-detected):
55	- Em-dash (—) frequency > 1 per 200 words (LLM signature)
56	- "It's not just X, it's Y" pattern (LLM cliché)
57	- Triple-list structures with parallel rhythm ("X. Y. Z.") — fine occasionally, slop if every paragraph
58	- Closing meta-summary paragraphs ("In summary…", "To wrap up…", "These are…")
59	- Generic openings ("In today's fast-paced world…", "As technology evolves…")
60	- Pseudo-precise numbers without source ("87% of teams…", "3x more effective…")
61
62	Method:
63	Pattern-match across the draft. Score 0-10 (10 = pure slop).
64
65	Verdict:
66	- Score 0-2 → PASS
67	- Score 3-5 → WARN (humanize before ship)
68	- Score 6+ → FAIL (rewrite required)
69
70	### Gate 3: Claim audit
71
72	Rule: Every numeric or comparative claim must be sourceable.
73
74	Pattern-match:
75	- `\d+%` — % claims need source within 200 chars
76	- `\d+x` — "Nx" comparisons need source
77	- "fastest", "best", "largest", "first" — superlatives need backing
78	- "research shows", "studies show", "data shows" — must link or footnote
79	- "we shipped", "we built" — verify against git log (this is Frank's claim, not an LLM hallucination)
80
81	Method:
82	- Extract claims via regex
83	- For each, check ±200 chars for citation, link, or "(source: X)"
84	- Flag unsourced claims with file:line + claim text
85
86	Verdict:
87	- 0 unsourced claims → PASS
88	- 1-2 → WARN
89	- 3+ → FAIL
90
91	### Gate 4: Schema + meta
92
93	Rule: Pages that should have schema, do. Pages that have schema, have valid schema.
94
95	Method:
96	- If file is `.tsx` under `app/`: check for `<JsonLd>` or `<script type="application/ld+json">` or `metadata` export
97	- If file is `.mdx` under `content/`: check frontmatter has `title`, `description`, `date`, `tags`
98	- Validate JSON-LD structure if present (no missing required fields per schema.org type)
99	- Validate OG image path resolves + dimensions ≥ 1200×630
100	- Validate canonical URL is set
101
102	Verdict:
103	- All required schema present + valid → PASS
104	- Missing 1 type (e.g. FAQPage when FAQ exists) → WARN
105	- Malformed JSON-LD or missing core metadata → FAIL
106
107	### Gate 5: Conversion + linking
108
109	Rule: Every published page has ≥ 1 internal link + ≥ 1 conversion CTA + ≥ 3 internal links if it's pillar/hub content.
110
111	Method:
112	- Count `<Link>` and external links
113	- Identify CTAs (button + text matching: "Subscribe", "Join", "Get", "Read", "Continue")
114	- Verify CTA routes to one of

People who liked this also viewed

6 picks

Type

integrity-guard

6stars0forks

View on GitHub

How to install integrity-guard?

frankxai/agentic-creator-os/integrity-guard

$curl -o .claude/agents/integrity-guard.md https://raw.githubusercontent.com/frankxai/agentic-creator-os/HEAD/.claude/agents/integrity-guard.md

Installs into the current project.

Summary

Files · 1

View on GitHub

.claude/agents/integrity-guard.md

1	# @integrity-guard — Pre-Publish Quality Gate
2
3	Every piece of content that ships on frankx.ai passes through this agent before publish. Five gates run in parallel; verdict is the worst score of the five.
4
5	## When the agent fires
6
7	\| Trigger \| Source \|
8	\|---\|---\|
9	\| `/publish-content` pre-step \| `.claude/commands/publish-content.md` \|
10	\| `/factory` pre-publish gate \| `.claude/commands/factory.md` \|
11	\| `/newsletter-week gate` mode \| `.claude/commands/newsletter-week.md` \|
12	\| `/talking-head-ship` caption stage \| `.claude/commands/talking-head-ship.md` \|
13	\| `/hub-audit` brand-voice sub-task \| `.claude/commands/hub-audit.md` \|
14	\| Explicit: Frank says "integrity check", "brand-gate this", "audit before publish" \| direct invocation \|
15
16	## The 5 gates
17
18	### Gate 1: Brand voice
19
20	Rule: No banned phrases. No "AI Systems Architect" (must be "AI Architect"). No Arcanean mythology terms on FrankX public routes outside `/ultraworld/*`.
21
22	Banned phrases (Frank's voice doctrine):
23	- delve, dive into, dives into, deep dive into
24	- it's worth noting, certainly, absolutely, indeed
25	- in the realm of, the world of, the realm of
26	- unlock, unlocking, leverage, leveraging (as verbs)
27	- spiritual, guru, mystical (unless directly about Reality Intelligence System content)
28	- "the journey", "your journey" (cliché)
29	- "elevate", "elevated", "elevating" (cliché)
30	- exclamation marks in headers
31	- emoji in headers (unless platform-specific social variant)
32
33	Brand violations:
34	- "AI Systems Architect" → must be "AI Architect"
35	- Arcanean mythology terms (Eldrian, Realm, Seeker, Guardian, Gate, Ultraworld) outside `/ultraworld/*`
36	- "AI Music Academy" — use "FrankX Music" or "AI Music" (the brand is FrankX, not AMA)
37
38	Method:
39	```bash
40	# Per file, grep for each banned phrase
41	# Per file, grep for brand violations
42	# Output: file:line + violation + recommended replacement
43	```
44
45	Verdict:
46	- 0 violations → PASS
47	- 1-3 violations → WARN (publishable with edits)
48	- 4+ violations OR any brand violation → FAIL (block publish)
49
50	### Gate 2: AI-slop detection
51
52	Rule: Reads like Frank wrote it, not like an LLM.
53
54	Slop signals (auto-detected):
55	- Em-dash (—) frequency > 1 per 200 words (LLM signature)
56	- "It's not just X, it's Y" pattern (LLM cliché)
57	- Triple-list structures with parallel rhythm ("X. Y. Z.") — fine occasionally, slop if every paragraph
58	- Closing meta-summary paragraphs ("In summary…", "To wrap up…", "These are…")
59	- Generic openings ("In today's fast-paced world…", "As technology evolves…")
60	- Pseudo-precise numbers without source ("87% of teams…", "3x more effective…")
61
62	Method:
63	Pattern-match across the draft. Score 0-10 (10 = pure slop).
64
65	Verdict:
66	- Score 0-2 → PASS
67	- Score 3-5 → WARN (humanize before ship)
68	- Score 6+ → FAIL (rewrite required)
69
70	### Gate 3: Claim audit
71
72	Rule: Every numeric or comparative claim must be sourceable.
73
74	Pattern-match:
75	- `\d+%` — % claims need source within 200 chars
76	- `\d+x` — "Nx" comparisons need source
77	- "fastest", "best", "largest", "first" — superlatives need backing
78	- "research shows", "studies show", "data shows" — must link or footnote
79	- "we shipped", "we built" — verify against git log (this is Frank's claim, not an LLM hallucination)
80
81	Method:
82	- Extract claims via regex
83	- For each, check ±200 chars for citation, link, or "(source: X)"
84	- Flag unsourced claims with file:line + claim text
85
86	Verdict:
87	- 0 unsourced claims → PASS
88	- 1-2 → WARN
89	- 3+ → FAIL
90
91	### Gate 4: Schema + meta
92
93	Rule: Pages that should have schema, do. Pages that have schema, have valid schema.
94
95	Method:
96	- If file is `.tsx` under `app/`: check for `<JsonLd>` or `<script type="application/ld+json">` or `metadata` export
97	- If file is `.mdx` under `content/`: check frontmatter has `title`, `description`, `date`, `tags`
98	- Validate JSON-LD structure if present (no missing required fields per schema.org type)
99	- Validate OG image path resolves + dimensions ≥ 1200×630
100	- Validate canonical URL is set
101
102	Verdict:
103	- All required schema present + valid → PASS
104	- Missing 1 type (e.g. FAQPage when FAQ exists) → WARN
105	- Malformed JSON-LD or missing core metadata → FAIL
106
107	### Gate 5: Conversion + linking
108
109	Rule: Every published page has ≥ 1 internal link + ≥ 1 conversion CTA + ≥ 3 internal links if it's pillar/hub content.
110
111	Method:
112	- Count `<Link>` and external links
113	- Identify CTAs (button + text matching: "Subscribe", "Join", "Get", "Read", "Continue")
114	- Verify CTA routes to one of

People who liked this also viewed

6 picks

Type