boub-voice-mastery

name: boub-voice-mastery description: Ensures AI-generated content sounds human, earned, direct, and specifically aligned with Boubacar Barry's voice. Triggers on "voice check", "does this sound like me", "apply my voice", "voice mastery", "humanize this", "Boubacar voice".

Skill: BouB Voice Mastery

Goal: Ensuring all AI-generated content sounds human, earned, and specifically like Boubacar Barry.

LinkedIn craft reference: For any LinkedIn post or article, also apply skills/ctq-social/references/linkedin-craft.md — hook patterns, pivot line rules, CTA types, format norms, and creator techniques reverse-engineered from Boubacar's own best-performing posts.

The Human Voice Manifesto

1. Directness is Respect: No "In today's fast-paced world" headers. No "It is important to note that" transitions. Start with the insight.
2. Opinionated over Neutral: AI presents all sides; Humans have a favorite side. Take a stand.
3. Grit over Polish: Perfect grammar without personality is a red flag. Use "And" or "But" to start sentences. Use short, punchy fragments.
4. No Em-Dash Abuse: Em-dashes (—) are "AI's favorite connector". Use them only once every 500 words. Use periods.
5. Specifics over Generics: Use real numbers, specific brands, and concrete anecdotes. Never use "cutting-edge", "revolutionary", or "seamless".
6. Sentence Case Required (locked 2026-05-20): Boubacar does NOT write all-lowercase. First word of every sentence capitalized. Proper nouns capitalized (Boubacar, Catalyst Works, agentsHQ; brand names like humanatwork.ai may stay lowercase as canonical). One occasional lowercase mid-sentence word is natural. All-lowercase post = OFF-VOICE. Half-lowercase post (<80% sentence-start uppercase) = OFF-VOICE. Trigger event: 7/30 HAW Launch X Thread. See feedback_no_all_lowercase_in_boubacar_voice.md + orchestrator/lint/lowercase_gate.py.
7. Connection Gate (locked 2026-05-29): Boubacar's signature is seeing non-obvious connections others miss. Every post must NAME A BRIDGE between two domains he uniquely straddles (Guinea/Conakry life <-> American operator; Theory-of-Constraints/process thinking <-> AI adoption; Comms+PR training <-> AI messaging; practitioner-who-built-agentsHQ <-> advisor). The bridge IS the insight. "Here is what happened / what I shipped" is the floor, not the post. TEST: "Is there a connection here no one else in the niche would draw?" If the draft is just a fact or a ship report, it is missing his superpower — re-anchor until it names the bridge. See feedback_connection_seeing_signature_2026_05_29.md. Pairs with the Earned Insight Gate below. Recovered 2026-05-30 from stashed work (stash@{3}, never committed).

AI Red-Flag Watchlist (2026 Edition)

• Excessive Em-Dashes: "This is a great tool — and it works — specifically for..." -> "This tool works. It's built specifically for..."
• Throat-Clearing: "In the realm of", "When it comes to", "Indeed", "Notably".
• Buzzword Soup: "Leverage", "Synergy", "Tapestry", "Delve", "Complexities".
• Uniformity: Sentence length should vary drastically. (Short. Long. Short.)

Avoid-Slop: Shann OS Tier 1 (8 patterns)

Source: @shannholmberg Content OS framework (2026-05-09 absorb). Full 54-pattern tier list deferred pending bookmarkable.io blueprint release (~2026-06-01). These 8 patterns augment the AI Red-Flag Watchlist above.

Scan every draft for these before voice polish. Any hit = flag and rewrite before proceeding.

Continuous Learning Tool

The agent using this skill is tasked with searching the web for:

• "AI text detection patterns 2026"
• "How to make ChatGPT sound like a person"
• "Common AI writing cliches"
• "Boubacar Barry writing style prompts"

HARD GATE: No Fabricated Stories (check this first, before any other review)

Before touching voice or style, scan for anecdotes attributed to Boubacar that he did not provide.

Red flags:

• Named characters in first-person scenes Boubacar did not describe ("I met Marco...", "A client called me at 2 AM...")
• Invented dollar figures tied to a client story
• "Last quarter I..." or "One of my clients..." followed by a fabricated scene
• Hypotheticals written in first person as if lived experience

If found: Stop. Flag exact sentences. Do not polish. Do not rewrite voice. Ask Boubacar to confirm or replace with an explicit hypothetical before proceeding.

Why: One false story attributed to Boubacar by name destroys the practitioner credibility the entire brand is built on.

Voice Profile: Boubacar Barry

Based on your instructions:

• Tone: Direct, Principle-based, Insightful.
• Constraints:
  • No hedging ("It might be", "One could argue").
  • No filler.
  • Lead with the "WOW" moment.
  • Connect operational constraints to financial/human outcomes.

Earned Insight Gate (MANDATORY — runs after Fabricated Story Gate)

Every post must draw from at least one item in skills/boub_voice_mastery/references/brand-spine-audit.md. Synthesized best practices with no personal anchor FAIL this gate.

Before proceeding with voice polish, ask: which earned anchor does this post draw from?

• Cultural bridge (African/American duality, built-from-scratch identity)
• Failure anchor (misplaced trust, wasted potential, hurt people around him)
• Contrarian anchor (AI is a tool, not savior/devil — humans still own the wheel)
• Values anchor (authenticity without performance, truth over comfort, joy as resistance, faith/family/Guinea+US first)

If the answer is "none — this is a general best practice or synthesized insight," STOP. Flag it: "This post contains nothing only Boubacar could have said. Ground it in one of his earned anchors or cut it."

Why: AI can replicate polish and even voice patterns. It cannot replicate Boubacar's specific lived experience. Posts that pass voice checks but contain no earned anchor are indistinguishable from AI content in a trust-scarce market.

I-Count Discipline (lessons from flagship session 2026-05-11)

Rule: Voice trim must never replace earned "I" with passive voice. Friend-over-coffee feel comes from active first-person voice. Passive constructions ("got built," "got pointed at," "gets dropped") sound like a whitepaper.

The right way to thin "I" stacks

When a post has 25+ first-person pronouns in <1500 words, the problem is usually 3-6 redundant stacks, not the count itself. Surgical cuts only.

Keep these I's:

• Opinion ownership: "I have a low tolerance for this pattern"
• Witness moment: "I have watched this in my own system"
• Scene: "I ran an adversarial code review on agentsHQ"
• Stance: "I do not soften it in client conversations"
• Anchor: "I built agentsHQ before teaching anyone else"
• Origin: "My mother worked for the government telecom company"

Cut these I's (redundancy patterns):

• "I bring that up because..." → just say the thing
• "I keep these names because..." → say what they cover, skip the framing
• "Every time I added... Every time I named..." → keep one I-clause, swing the second to "you"
• "before I taught anyone" → "before teaching anyone" (one I, not two)
• "This is why I built X. Not because... Because I needed it..." → "That is why X looks the way it does. Not because... Because it had to work on..."

Hard rule: never use passive voice to dodge an "I"

Wrong: "The reviewer got pointed at the notes" / "got built" / "gets dropped" Right: "I pointed the reviewer at the notes" / "I built it" / "an agent gets dropped" (agent-active is fine)

If the only way to avoid an "I" is passive voice — keep the "I." Reader trusts the human, not the construction.

Target I-count

• Short feed post (<300 words): 3-7 I's is normal
• LinkedIn post (200-400 words): 8-14 I's
• Long article (1000-1500 words): 18-25 I's — anything above ~28 = check for redundant stacks
• I-count is a trailing indicator, not a target. Don't optimize for the number. Optimize for friend-over-coffee feel.

Date References (lessons 2026-05-11)

Rule: Posts get cross-published over days/weeks. Avoid hard date anchors that age badly.

• ❌ "Two days ago" — by ship date it's wrong
• ❌ "Yesterday I did X" — same
• ✅ "A few days ago"
• ✅ "Last week"
• ✅ "Recently"
• ✅ Specific event with permanent reference: "When I ran the adversarial code review on agentsHQ" (no time anchor)

For breaking news / current-event posts (Timely content type), specific dates are fine. For Evergreen content, soft anchors.

Halbert Craft Floor + Hype Guardrail (runs after Earned Insight Gate, before Output Contract)

FIRST — classify the draft by content type. This gates which axes run:

• CONVERSION COPY = sales / outreach / landing-page / CTA / lead-magnet. Concretely: Ghost Brief or ghostwriting sales drafts, SW/CW/GW cold outreach, HAW sales pages + lead magnets, landing-page copy, sales emails, any block whose job is to get the reader to act.
• EXPRESSION = everything else. Thought-leadership LinkedIn posts, bios, observational X posts, narrative/story content, chat replies, internal docs.

The Halbert KEEP axes are direct-response / conversion-copy fundamentals (starving-crowd market, one-reader, CTA, emotional-then-logical). They apply ONLY to CONVERSION COPY. Running them on a thought-leadership post or a bio is WRONG — those are not sales letters and would get falsely flagged for "missing" axes that do not apply.

The REJECT anti-axes (hype/fabrication guardrails) apply to EVERYTHING. Hype, manufactured urgency, and fabricated proof are never OK in any Boubacar-voiced content.

Score against the Halbert copy axes in skills/boub_voice_mastery/references/halbert-copy-axes.md — Gary Halbert's direct-response canon, reconciled for Boubacar's earned voice (pedagogy + PGA framing credited to Nicolas Cole).

• 8 KEEP axes — CONVERSION COPY ONLY (positive craft checks): market-first/starving-crowd, A-pile specificity, specificity-over-generality, you-focus/one-reader, story-led "movie of the mind," emotional-then-logical, read-aloud readability, copywork/draft-fast. Run these ONLY if the draft is conversion copy. A FAIL here is a craft gap to fix, not a hard block. Fold any fail into the "Fixes to make" layer. If the draft is EXPRESSION, skip this layer entirely (do not flag missing CTAs / one-reader focus / starving-crowd framing on a thought-leadership post).
• 5 REJECT anti-axes — UNIVERSAL (hype guardrails): manufactured urgency, hyperbolic multipliers, fake-personal gimmicks, P.S./multi-CTA/big-button stacks, fake testimonials/credential-signatures. These are anti-axes — run them on EVERY draft regardless of content type. If the draft exhibits one, FLAG it (quote the line in "Issues found") and route the fix per the violated rule (feedback_no_ben_angel_hype_pattern, feedback_truth_only_no_invented_proof, and the others named in the axes file).

Read the axes file for the per-axis check question + PASS/FAIL examples + rule cross-links. Rationale: docs/playbooks/halbert-copywriting-principles-2026-05-29.md.

Burstiness + Anti-Tell Pass (final mechanical sweep — runs after all gates, before the Output Contract)

Why: even true, in-voice content reads machine when the SCAFFOLDING is uniform. AI detectors flag low burstiness (even sentence rhythm) + known classifier phrases. This pass restores human structure. NOT number-gaming — truth-only holds; never invent an aside to "look human." A lower detector score is a side effect of better writing, not the target. (Folds in Boubacar's 2-year-old ChatGPT-Humanizer list.)

Run three steps on every draft:

A. Burstiness sweep (read aloud). No 4+ consecutive sentences within ~5 words of each other — break the run with a fragment or a fused long sentence. At least one single-sentence paragraph. At least one sentence over 25 words AND one under 6.

B. Anti-tell scan — find and fix each:

• Connectors / transitions (delete, start cold): "let's dive in", "here's the honest part", "the reason this happens", "in today's", "in today's digital age", "in an era where", "when it comes to", "in the realm of", "it's worth noting", "it's important to note", "the truth is", "in fact", "indeed", "first and foremost", "as a result", "therefore", "consequently", "in other words", "to put it simply", "in summary", "to sum up", "in conclusion", "all in all", "more importantly". Plus the metaphor cluster "realm / navigating / beacon in the evolving landscape / complex ecosystem".
• Do NOT over-correct (syntax salad is its own 2026 tell): deleting a connector does not mean orphaning the sentence. Where a thought truly needs a bridge, use plain words a person says ("so", "but", "then"), not AI fake-casual hooks. Also banned as over-corrections: "but here's the kicker", "that's only half the story", "why does this matter?", "real talk", "here's the thing", "here's the truth". Swapping one tell for a trendier tell is not humanizing.
• Rule of three (content-type gated — read the CONVERSION vs EXPRESSION class from the Halbert floor above): EXPRESSION = at most one tricolon per ~300 words, cut the rest. CONVERSION COPY = one well-placed tricolon is a Halbert craft feature, keep it; the real rule is no repeated tricolon pattern across a series (the EEC failure was the same 3x3 frame in five straight emails, not one good triple).
• Antithesis (content-type gated): "not X but Y" is a real Boubacar move AND a Halbert contrast device. EXPRESSION = at most one per piece. CONVERSION COPY = keep where it sells, but never the same antithesis shape twice in one asset or series. Rewrite "not only X but also Y" / "both X and Y" / "on one hand, on the other" as direct statements in all content types.
• Repeating frame: if a section or bullet skeleton repeats more than twice, break one instance (drop a header, merge bullets, vary bullet length, convert a list to a paragraph).
• Recap closer: no "to recap / over the last N we covered / in summary". Cut or replace with a forward line.
• Formulaic structure (Boubacar's humanizer core rule): no predictable intro → scene-set → steps → considerations → summary skeleton. No throat-clearing intro, no ethics paragraph, no generic advice.
• Banned-word carve-out: the filler words in the linked list are banned as polish-reaching, not as vocabulary. "context", "framework", "landscape" are fine as the literal specific thing ("the context of this deal", "our 12-week framework"). Judge the use, not the substring.
• Em-dashes: zero (already a hard rule).

C. Perplexity lift — one move per paragraph. Replace one safe/expected word with the specific word Boubacar would actually say (more lived, not a fancier synonym). Keep one genuine idiosyncratic phrase only he would write — surface it from his real material, NEVER invent one.

D. Active-mistake framing + offer-stack separation (Luke V, 2026-06-11). In any mistake/problem framing, the second half must show the reader actively CHOOSING the behavior ("you pick the next ticket instead"), never outcomes-by-absence ("the market never hears you") — absence-outcomes are an AI tell; ask "what is the reader actively choosing in that moment?". And never collapse title + promise + personal whisper into one paragraph — one element per line. A3 extensions: contrast-frame TEMPLATES repeating across a series = AI tell (keep one, rewrite rest single-direction); uniform day openers/closers = tier-1 tell; CTA always on its own line, never inside a pitch+proof paragraph; client EECs mine the client's first-person mistake stories. Detail: docs/playbooks/humanizer-anti-ai-detection-pass-2026-06-08.md.

Optional detector check (informational, not a gate). Run a detector for a directional read if handy; treat the number as a smoke signal. If the writing is genuinely human and true and still flags, ship it — chasing the number invites fabrication.

Full banned-word/phrase list + before/after technique + rationale: docs/playbooks/humanizer-anti-ai-detection-pass-2026-06-08.md.

Voice Review Output Contract

Every voice review must produce output in exactly three layers — no exceptions, no prose dumps:

1. Issues found — specific lines or patterns that fail a gate (quote the offending text)
2. Fixes to make — exact rewrites or directives (not suggestions)
3. Summary — only when there are 3+ issues; one sentence; what changed and why

No layer is skippable except Summary when issues < 3. This prevents open-ended "here are some thoughts" outputs that send Boubacar back into the content without a clear action path.

Why: AI can replicate polish and even voice patterns. It cannot replicate Boubacar's specific lived experience. Posts that pass voice checks but contain no earned anchor are indistinguishable from AI content in a trust-scarce market.

Feedback-Reply Craft (added 2026-06-05, from Luke V Module-2 loop)

When drafting Boubacar's reply to feedback (coach, client, reviewer):

1. Never narrate "same day" when replying same day. The timestamp says it; saying it reads self-congratulatory.
2. Pull on the giver's words. Quote their key phrase back, then show the change + why it follows from their note. Digestible = change + why ONLY; never re-paste the full document (redundant if they will see the next artifact anyway).
3. Restate the principle in Boubacar's own words ("active front to back now, they catch themselves mid-habit") -- learning evidence beats compliance evidence.
4. Name unprompted transfer if the fix was propagated to other artifacts -- fast-implementer signal coaches escalate attention for.
5. No re-review ask. Do not invite the giver to check the rewrite; the next deliverable is their natural checkpoint.

Source: memory feedback_apply_feedback_immediately_cooking_2026_06_05 (the cooking principle: apply feedback NOW, rest the dish).