Blog Voice Guide & Story Worlds

🔴 READ THIS BEFORE WRITING OR EDITING ANY POST IN C:\ssClawy\aguidetocloud-revamp\content\blog\. This is Sush's blog voice gate. Part 1 (the USP, voice, story worlds, section pattern) is Sush-authored — read, don't own: change it only with his explicit sign-off. Part 2 (Anti-AI-writing gate) was added 21 Aug 2026 after a measured audit of all 71 posts, and is validated against the external standard conorbronsdon/avoid-ai-writing.

Scope: the aguidetocloud.com blog only. Plain AI content has its own separate gate at ~/.copilot/plain-ai-voice-guardrail.md — do not cross-apply them.

Promoted into the portal 21 Aug 2026 from a dead session-state folder, where it was invisible for months (Rule #14b — Law of Dead Mechanisms).

---

Part 1 — Voice (Sush-authored, sacred)

The USP (Hardcoded)

We simplify hard concepts for people like us. Not power users. Not developers. If a power user finds it "too simple" — good. They can go to Microsoft Learn. We exist for the IT admin who needs to understand AND implement by Monday morning.

The Voice

• First person. "I" and "you" — never "one should" or "the administrator must"
• Coffee chat energy. Write like explaining to a smart colleague over coffee
• Admit confusion. "This confused me too — here's how I finally got it"
• Recommend, don't just list. "If I were setting this up, I'd start with..."
• Short sentences. If a sentence has a comma, ask if it needs to be two sentences
• Active voice always. "Run this command" not "This command should be run"
• No jargon without explanation. First use = explain. Second use = just use it
• One analogy per major concept. From the 3 story worlds below

The 3 Story Worlds

1. Your Phone 📱

Use for: Features, tiers, what you get, app permissions, settings Maps to: - Free apps = Copilot Chat (free tier) - Paid apps = M365 Copilot ($30/month) - App permissions popup = data access controls ("Allow Copilot to access your files?") - Parental controls = content safety filters - Settings app = Admin Center - Software update = feature rollout - App Store = agent marketplace - Notifications settings = alert policies - Screen time reports = usage analytics

Example usage:

Think of content safety like parental controls on your phone. The phone comes with basic protections that you can't turn off — you can't install malware from the App Store. But on top of that, you can choose to enable or restrict certain content categories. That's exactly what the harmful content toggle does.

2. The Hotel 🏨

Use for: Security, governance, admin controls, who gets access to what Maps to: - Hotel = your M365 tenant - Guest rooms = user accounts/mailboxes - Room key card = licence - Front desk = Admin Center - Hotel rules (no smoking, quiet hours) = tenant policies - Safe in the room = sensitivity labels - Do Not Disturb sign = DLP policies - Fire alarm / sprinkler = security alerts (always on) - VIP floor / executive lounge = premium features (E5, Copilot) - Security cameras in hallways = audit logs - Concierge = Copilot (helps you find things) - Master key = Global Admin

Example usage:

Think of your M365 tenant like a hotel. Every guest (user) gets a room key (licence) that opens their room and the shared areas. The fire alarm and sprinklers work whether guests want them to or not — that's your core AI safety. But the hotel also lets certain guests access the executive lounge if the front desk (admin) approves it — that's your per-user policy toggles.

3. Moving House 🏠

Use for: Deployment, migration, rollout, change management, adoption Maps to: - Old house = current setup / old way of working - New house = M365 Copilot deployment - Packing = data preparation / governance audit - Moving truck = migration tools - Unpacking room by room = phased rollout (pilot → wave 1 → wave 2) - Change of address = updating DNS/domains - First night chaos = Day 1 problems - Getting to know the neighbourhood = user adoption - House inspection before moving = readiness assessment - Hiring movers vs DIY = professional services vs self-deploy - The "junk drawer" = the SharePoint site with permissions nobody understands

Example usage:

Deploying Copilot is like moving house. You wouldn't just throw everything in a truck and hope for the best — you'd sort through your stuff first, decide what to keep, label the boxes, and unpack room by room. That "sorting through your stuff" phase? That's your SharePoint permissions audit. And trust me, every organisation has a junk drawer — that one SharePoint site where permissions got weird years ago and nobody fixed it.

How to Use Story Worlds

• Rotate between them — don't use the same one twice in a row
• One analogy per h2 section — open the section with it, then get technical
• Don't force it — if a concept doesn't map naturally, skip the analogy
• Keep it to 2-3 sentences — the analogy opens the door, the explanation walks through it
• Never ONLY use the analogy — always follow with the real technical detail

Section Pattern for Blog Posts

Every h2 section should follow this rhythm: 1. Hook — analogy or "here's what most people miss" (1-2 sentences) 2. Simple explanation — what it is in plain English (2-3 sentences) 3. The detail — tables, steps, technical specifics 4. The action — "Here's what I'd do" or "Your next step is..."

What NOT to Do

• Don't start sections with definitions ("X is a feature that...")
• Don't use passive voice ("It should be noted that...")
• Don't hide the recommendation ("There are several options...")
• Don't write paragraphs longer than 4 sentences
• Don't use "leverage", "utilize", "implement", "facilitate", "stakeholders"

---

Part 2 — Anti-AI-writing gate (added 21 Aug 2026)

Why this section exists. Sush asked whether the blog reads as AI-written. An audit of all 71 posts / 247,101 words against conorbronsdon/avoid-ai-writing found the answer is mostly no — but the formatting gives it away. Part 1 above was already doing its job on vocabulary; nothing was protecting punctuation and emphasis.

The measured baseline (21 Aug 2026, before the de-bolding pass)

Signal	Measured	Standard
Tier 1A AI marker words	0.36 / 1k	low	🟢
Tier 1B clarity words (leverage, utilize, robust)	0	0	🟢
Tier 2 clusters (2+ per paragraph)	0	0	🟢
Tier 3 saturation	0.02%	flag at 3%	🟢
Chatbot artifacts / sycophancy / cutoff disclaimers	0	0	🟢
Em dashes in prose	20.3 / 1k	hard max 1	🔴
Bold runs	30.1 / 1k	~1 per section	🔴
Headings with em dash	622 / 2,294	—	⚠️
Title Case headings	532 / 2,294	sentence case	⚠️

Conclusion: the vocabulary gate works; the formatting gate did not exist. That is what Part 2 adds.

🔴 Rule 1 — Bold is a budget, not a highlighter

Bold exists so a scanner can find the one thing that matters in a block. Every extra bold spends the budget and makes the page look machine-generated.

• At most one emphasis bold per line, bullet, or paragraph. A structural lead-in label (**Why it matters:**, **Both can run at once.**) does not count against the budget — the label and one emphasis can coexist.
• Never bold a clause or a sentence. If a whole sentence matters that much, restructure it, pull it into a callout, or make it its own line. Bold >5 words is always wrong.
• Bold the first mention of a term only. Repeat mentions read as nervous emphasis.
• Tables, headings and callout shortcodes are exempt — bold inside them is structural.

Enforced by scripts/debold-propose.mjs (dry-run by default; --cap=1 --apply to rewrite).

🔴 Rule 2 — Earn the em dash (corrected 2026-08-21 — read this before acting on it)

The original version of this rule was wrong, and it was wrong in the direction that would have damaged the blog. It claimed a splice is "the single most recognisable AI punctuation habit" and put hard counts against it. Gate A killed both claims.

What the evidence actually says. The external standard published a human-control corpus (corpus/README.md, v3.22.0, 875 human + 779 machine paragraphs) measuring which of its own categories separate human from machine text:

Category	Human	Machine	Lift
uniformity (sentence rhythm)	2.1%	25.1%	11.7×
filler	2.4%	8.3%	3.4×
tier1 (the 112-word list)	8.0%	7.4%	0.9×
em-dash	9.9%	1.9%	0.2× — inverted

Its maintainer's own conclusion: "On this corpus an em dash is evidence the text is human." Pooled ROC-AUC for the composite score is 0.501 — a coin flip.

So: em dashes are a style question, not an authorship question. Stripping them would have made this blog read less human, not more. Never again cite dash density as evidence that writing "looks AI-generated".

Do not quote the old counts. 5,025 and 6,589 were both wrong. ai-writing-audit.mjs had a preprocessing bug (bare-URL regex ran before link targets, and \S+ ate the closing paren) that silently swallowed 962 dashes and 70,789 words. Fixed 2026-08-21. The defensible statement is "roughly 6.5–6.6k render-relevant em dashes across 74 posts, depending on parsing rules."

What survives as house style — a readability rule, owned by Sush, not compliance with anything:

Use em dashes deliberately. Review any paragraph where repeated dashes make every explanation arrive with the same pause.

• The concrete queue is paragraphs with 3+ dashes (83 of them, 34 posts) — repetition that has drained the punctuation of its effect. Not singletons, and not "two in one sentence", which preferentially finds legitimate paired parentheticals.
• — as typography is correct and stays — - **Term** — definition list items, headings, separators. That is the notebook grammar, deliberately.
• The expressive pause is voice and must be protected — "It's sitting right there — in Teams, Outlook, Word". Replacing these with commas makes the prose flatter and more corporate.
• Never use -- as an em dash substitute.

⚠️ Not remediated, and mostly should not be. Any edit here is semantic, so it is propose-and-approve per paragraph — never a bulk --apply. Removing ** preserves every word; replacing a dash is an editorial decision.

The signal that actually matters is rhythm. Run scripts/rhythm-audit.mjs. Current state: 0 of 74 posts flagged for uniformity (median sentence-length CV 0.789 against a 0.25 threshold) and 0 of 74 for vocabulary. The blog is clean on the dimension that discriminates.

🔴 Rule 3 — Words that are never worth it

Part 1 already bans leverage, utilize, implement, facilitate, stakeholders. Add these, from the external standard's Tier 1A/1B lists. Keep them as writing advice, not as authorship evidence — that same standard measured its 112-word vocabulary table at a lift of 0.9×, so it fires slightly more on human writing than machine writing. delve is still worth replacing because it is a limp word, not because it proves anything about who typed it.

delve · robust · seamless · comprehensive · landscape · realm · tapestry · navigate the complexities · deep dive · best practices · it's worth noting · in today's fast-paced world · unlock the power of · game-changer · testament to

Hollow intensifiers — cut when the sentence means the same without them: genuinely · truly · really · actually · simply · to be honest

🧩 The genuinely nuance (Sush's call, 21 Aug 2026). Keep it where it marks a threshold or a contrast — "only when developers genuinely need control", "genuinely different from regular Chat", "there is a genuinely free tier". That is honest-broker work and it is his voice. Cut it before a plain evaluative adjective — "genuinely impressive" says nothing more than "impressive". 32 of 86 uses were cut on that rule; 54 stayed.

🔴 Rule 4 — Headings

• Sentence case, not Title Case. "What mobile is good for", not "What Mobile Is Good For".
• No em dash in a heading where a colon or a shorter heading works.
• Emoji in headings: the site's handwritten over emoji rule already prefers a typographic signal. Don't add new ones.

🔴 Rule 5 — List shape

• Don't write runs of 5+ bare noun-phrase bullets. Bullets should be sentences or **Label:** explanation pairs. A wall of bare fragments is a strong AI tell.
• **Label:** not **Label.** — a colon reads as a label, a full stop reads as generated.
• Avoid the rule of three as a default rhythm ("faster, cheaper, and more reliable"). Use two, or four, or just say the one that matters.

Running the gate

cd C:\ssClawy\aguidetocloud-revamp

# Detect only — never rewrites. Reports every signal above, per post and in aggregate.
node scripts/ai-writing-audit.mjs

# Bold budget: dry run, then apply
node scripts/debold-propose.mjs --cap=1
node scripts/debold-propose.mjs --post=<slug> --cap=1 --samples
node scripts/debold-propose.mjs --cap=1 --apply

The standard's own caveat, which must be repeated whenever these numbers are quoted: these are signals, not proof. They fire more often on second-language writers (Sush is not a native English speaker) and on technical genres. A high score is a prompt to look, never a verdict.

Cross-references

• blog-notebook-system.md — the visual/typographic system these posts render into
• voice-and-tone.md — brag-allergy, phrasing fingerprints, creative-juice pattern
• ~/.copilot/plain-ai-voice-guardrail.md — Plain AI only, do not cross-apply