peterbamuhigire/ai-content-humaniser

AI Content Humaniser

<!-- dual-compat:start -->

Use when

• Transforms AI-generated marketing drafts into brand-voiced, human-quality content by applying a structured quality-control process covering vocabulary, cultural localisation, factual accuracy, and voice consistency. Invoke immediately after any AI-assisted draft and before any client delivery — for captions, blog posts, email copy, strategy documents, and proposals.
• Use this skill when it is the closest match to the requested deliverable or workflow.

Do not use when

• Do not use this skill for graphic design, video production, software development, or legal advice beyond the repository's stated scope.
• Do not use it when another skill in this repository is clearly more specific to the requested deliverable.

Workflow

1. Collect the required inputs or source material before drafting, unless this skill explicitly generates the intake itself.
2. Follow the section order and decision rules in this SKILL.md; do not skip mandatory steps or required fields.
3. Review the draft against the quality criteria, then deliver the final output in markdown unless the skill specifies another format.

Anti-Patterns

• Do not invent client facts, performance data, budgets, or approvals that were not provided or clearly inferred from evidence.
• Do not skip required inputs, mandatory sections, or quality checks just to make the output shorter.
• Do not drift into out-of-scope work such as code implementation, design production, or unsupported legal conclusions.

Outputs

• An AI-focused strategy, audit, system design, or prompt asset in markdown with human review and control points.

References

• Use the inline instructions in this skill now. If a references/ directory is added later, treat its files as the deeper source material and keep this SKILL.md execution-focused.

<!-- dual-compat:end -->

The Golden Rule

This agency is AI-powered. Every output it produces — text, captions, proposals, blog posts, emails, strategy documents — must look, feel, and sound as if it was crafted by the most skilled human creatives with meticulous attention to detail and deep knowledge of the target audience. Clients and their audiences must never sense the AI behind the work.

This skill is the enforcement layer for that standard. Use it for every piece of AI-assisted content before client delivery. The test is not "does this pass a grammar check?" The test is "would an expert in this market, reading this cold, believe a skilled human wrote it?"

Purpose

AI writing tools produce fast drafts. They also produce generic, culturally misaligned, over-smooth content that reveals its origins to any experienced reader. This skill is a quality-control process for closing the gap between an AI-generated draft and content that reads as written by a knowledgeable human expert who knows Uganda and East Africa.

The goal is not to conceal AI involvement. The goal is to ensure the output genuinely earns its place — accurate, specific, culturally grounded, and written in the client's voice.

Apply the AI-as-creative-partner principle throughout (Johnsen, 2024): AI drafts are raw material, never finished work. Editors have full authority — and full responsibility — to rewrite rather than polish. A light polish that leaves AI fingerprints is a failure of this process.

---

Required Input

Before beginning any humanisation review, confirm the following:

1. Client business name — whose brand voice does this content represent?
2. Industry — what sector is the client in?
3. Country / city — where is the audience located? (Default: Uganda / East Africa)
4. Primary goal — what is this piece of content meant to achieve?
5. Content type — caption, blog post, email, strategy document, proposal, or other?
6. Draft to review — paste the full AI-generated draft for review.
7. Brand voice notes — any existing tone-of-voice guidelines, sample copy, or adjectives the client uses to describe their brand (e.g., "professional but warm", "bold and direct").

---

The Five AI Quality Risks

Every AI-generated draft carries these risks (Ltifi and Johnsen). Check for all five before approving any piece for client delivery.

1. Originality

AI is trained on existing content and blurs into plagiarism territory. Output may reproduce phrases, structures, or ideas from source material without attribution. Run any passage that seems unusually polished through a plagiarism check before delivery.

2. Quality Inconsistency

AI does not consistently meet professional standards. A single draft may contain one strong paragraph and three weak ones. Do not approve the whole because one section is good — review every section independently.

3. Bias and Stereotyping

AI tools are trained predominantly on Western data. They perpetuate Western assumptions about audiences, markets, behaviours, and values. Content produced for a Ugandan audience without local review will often reflect these biases in ways that range from irrelevant to actively alienating.

4. Misinformation

AI generates convincing but factually incorrect claims, invented statistics, and hallucinated citations. Treat every statistic, figure, and citation in an AI draft as unverified until confirmed against a primary source. Do not pass fabricated data to a client.

5. Flatness

AI lacks genuine texture — the finesse, inventiveness, and specificity that come from lived expertise. Content may be technically correct but feel hollow. Flatness is the hardest risk to quantify and the most damaging to brand credibility over time.

---

The Uncanny Valley Calibration Checklist

AI content that is almost-but-not-quite human triggers a stronger rejection response than content that is obviously AI. This is the uncanny valley effect applied to text — readers cannot name what is wrong, but they feel something is off, and they disengage (Ltifi, 2024).

The structural signs of uncanny AI content:

• Every paragraph is approximately the same length
• Transitions are consistently smooth — no abrupt shifts in register
• There is no natural hesitation, qualification, or colloquialism
• Phrasing is suspiciously balanced — no strongly held positions
• Vocabulary is drawn from the same formal register throughout

When uncanny valley is detected, apply these four corrective moves:

1. Break the symmetry. Vary paragraph length deliberately. Follow a long paragraph with one short sentence. Use a sentence fragment for emphasis.

2. Insert a real detail. Add one specific, locally grounded detail that only a person familiar with this market would include — a Kampala neighbourhood, a named local brand, a Uganda-specific price point, a cultural reference that lands for this audience.

3. Take a position. Replace "one approach is…" with "I recommend…" or "the evidence suggests…". AI defaults to false balance. Human experts have opinions.

4. Introduce a controlled imperfection. A conversational aside. A rhetorical question. A sentence that starts with "And". A regional idiom used correctly. These are not mistakes — they are authenticity signals.

Run this checklist on every draft before applying the Human Voice Checklist. If three or more uncanny valley signs are present, the draft requires a full rewrite, not a polish.

---

AI Vocabulary to Eliminate

These words and phrases are characteristic of AI-generated marketing content. Their presence signals an unreviewed draft. Remove or rewrite every instance.

Single words — banned: delve, tapestry, leverage (as a verb), foster, realm, seamlessly, robust, comprehensive, revolutionary, groundbreaking, game-changer, navigate (metaphorical use), landscape (metaphorical use), beacon, testament, crucial, vital, cutting-edge, innovative, empower, unlock, journey (metaphorical use), vibrant, dynamic

Phrases — banned:

• "In today's digital age"
• "In the ever-evolving"
• "It is worth noting that"
• "It is important to note that"
• "It goes without saying"
• "With that being said"
• "At the end of the day"
• "Moving forward"
• "Take your business to the next level"
• "One-stop shop"

Over-smooth transitions — rewrite or remove:

• "Furthermore,"
• "Moreover,"
• "In addition to the above,"
• "Building on this,"

Weak hedges — strengthen or cut:

• "may potentially"
• "could possibly"
• "one might consider"
• "it could be argued that"
• "in some cases"

---

The Human Voice Checklist

Apply these ten questions to every AI draft. A draft that fails more than two must be rewritten, not patched.

1. Does the opening hook grab attention without being generic or clichéd? If not, rewrite the first sentence entirely — do not smooth it; replace it.

2. Is there at least one concrete, specific detail — a number, a named place, a named person, a real event? Generality is the primary signal of AI authorship.

3. Does the content take a clear position? Replace "one approach is..." with "I recommend..." or "the evidence points to...". AI defaults to false balance; human experts have opinions.

4. Is any sentence over 35 words? If yes, split it. Long, subordinate-clause-heavy sentences are an AI signature. Short sentences carry authority.

5. Does the vocabulary feel natural to the audience, or does it read like a marketing textbook? Read the piece aloud. If you stumble, the audience will stumble.

6. Are there any facts, statistics, or citations that have not been verified against a primary source? Mark every unverified claim before proceeding.

7. Does the content feel written by someone who knows Uganda and East Africa — not as a foreign market, but as the normal context? If not, apply the Cultural Localisation Checklist below.

8. Is the call to action clear and direct? "Learn more" is not a CTA. "Send us a WhatsApp message on 0700 000 000 before Friday" is a CTA.

9. Does the draft avoid all banned vocabulary and phrases listed above? Do a word search if unsure — do not rely on reading alone.

10. Would you be comfortable if the client knew exactly how this piece was produced — AI draft, reviewed and revised by this process? If the answer is no, it is not ready.

11. Does this content have the natural micro-variation in sentence length, register, and vocabulary that characterises human writing — or does it have AI's telltale mechanical uniformity? If every sentence feels structurally similar, the draft requires deliberate variation before it can pass as human.

"Ineffable Something" Test

Source: Ching & Mothi (2025, p.64) — citing Tyler Cowen's March 2024 review of Suno AI. Named quality criterion: does this content move, surprise, or connect in a way that is unmistakably human? If a skilled reader would feel nothing in particular after reading it, the content fails the Golden Rule regardless of grammatical correctness. The test is subjective by design — that is the point. If a piece is technically correct but leaves the reviewer with no reaction, it requires a rewrite at the level of voice and meaning, not vocabulary. Apply as the final quality gate after all checklist items are satisfied.

Micro-Moment Register Matching

When editing content used at pivotal customer journey moments — a cart abandonment message, a post-browse retargeting caption, a comparison-shopping landing page — the copy must match the emotional register of that exact moment (Ltifi, 2024). Generic AI output cannot detect emotional context; human editors must impose it.

• Urgency without pressure: Time-sensitive offers need energy, not aggression.
• Reassurance without sycophancy: Trust-building moments need warmth, not flattery.
• Confidence without arrogance: Authority content must convey expertise, not dismissiveness.

Before approving any content used at a known customer journey touchpoint, ask: does the tone of this copy match what the audience is feeling at this moment?

---

Editing Process by Content Type

Captions

1. Check the first line — the hook. Rewrite if it opens with a question, a generic statement, or any banned phrase. The first line must earn the tap to expand.
2. Remove every banned word and phrase. Do not replace with a synonym from the same register — rewrite to be direct.
3. Add one specific local detail: a place name, a local reference, a Ugandan price point, a relevant local event. This alone separates local content from generic content.
4. Ensure the CTA is direct and platform-appropriate. On WhatsApp-first platforms, the CTA should reference WhatsApp where relevant.
5. Read aloud. If it sounds like a press release, simplify. If it sounds like a list, restructure around a single idea.

Blog Posts

1. Rewrite the opening if it starts with a definition ("According to the Oxford Dictionary..."), a decontextualised statistic, or any variation of "In today's...". The opening must state the piece's central argument or the reader's problem — immediately.
2. Ensure at least two clear opinions or recommendations appear per article. AI defaults to neutral description; human experts give direction.
3. Verify every statistic and citation against its primary source. Replace any that cannot be verified with either a verified alternative or a first-hand observation.
4. Add at least one East African example, case study, or reference that is specific and accurate. Generic "African" examples are not sufficient.
5. Review each section heading. Ask: does the section that follows answer a genuine question, or does it merely restate the heading in paragraph form? If the latter, restructure.

Email Copy

1. Subject line: Does it create genuine curiosity without resorting to clickbait? Test against this standard: would you open it? Would the client's specific audience open it?
2. Preview text: Does it add new information, or does it repeat the subject line? Repetition wastes the preview text. Rewrite to extend the subject line's idea.
3. Body — first paragraph: Is it written to the reader as an individual? Replace "our valued customers" and "businesses like yours" with direct, specific address.
4. CTA: Is there exactly one call to action? Multiple CTAs split attention and reduce conversion. Is it specific — does it tell the reader exactly what to do, where, and when?
5. Tone: Does the email match how the client actually communicates with their customers? Cross- reference against any existing brand voice notes.

Hallucination Gate

Source: Evelyn (2025); Mizrahi (2024). Apply to all content types. The gate checks: statistics, citations, named entities, dates, and product/service claims. For any output making factual claims, run: "Use web search to find the latest news and resources, and cite your sources" — then verify the cited sources independently before client delivery. Do not assume web-search output is correct; verify the linked source directly. The gate is not optional — it is a production standard for any content that makes claims of fact.

Strategy Documents and Proposals

1. Are all recommendations specific and actionable? "Leverage synergies across digital touchpoints" is not a recommendation. "Post three times per week on Facebook, focusing on video content between 18:00 and 20:00 EAT" is a recommendation.
2. Does the document reflect the client's specific situation, or is it generic boilerplate with the client's name inserted? Replace every generic section with a client-specific observation.
3. Verify all AI-generated market data — penetration figures, demographic statistics, platform usage data — against a named and dated primary source. Footnote the source.
4. Does the document demonstrate knowledge of the Uganda and East Africa market specifically? Check for assumptions based on Western market conditions (credit card penetration, broadband access, regulatory environment, platform algorithm behaviour in the region).

---

Cultural Localisation Checklist for East Africa

Run this checklist on every piece of content before client delivery.

• Does any content assume Western payment methods — credit cards, PayPal, Stripe — when Mobile Money (MTN, Airtel) is the dominant payment method for the target audience?

• Are examples, case studies, or statistics drawn from the United States, United Kingdom, or Europe when East African equivalents exist or should be used?

• Does the tone match Ugandan professional communication norms — warm, respectful, relationship- first, and community-oriented — rather than the transactional directness common in Western marketing copy?

• Where prices appear, are they in UGX (Ugandan Shillings) for local audiences, or in the relevant local currency for the market in question?

• Does the content acknowledge the WhatsApp-first mobile environment where relevant? In Uganda, WhatsApp is the primary channel for customer communication. Content that directs audiences to email-first or website-first contact points will underperform.

• Are any cultural references, idioms, or examples likely to be unfamiliar or meaningless to the target audience? Replace them with locally resonant equivalents.

• If the content mentions internet access, streaming, or data usage — does it account for the reality of mobile data costs and connectivity patterns in the region?

• For any content translated or adapted into Luganda, Swahili, Amharic, or another EA language: verify cultural idiom accuracy, not just linguistic accuracy. A linguistically correct translation may contain examples, proverbs, or cultural references that are foreign or inappropriate for the target community. Commission a native-speaker review for all translated content before client delivery.

• Cultural Bias Audit (Source: Ching & Mothi, 2025): For any AI-generated content depicting people, cultures, identities, or communities — run an explicit bias check: does this reflect the actual demographic and cultural reality of the target audience? AI tools were trained predominantly on Western datasets and systematically produce Western-centric, gender-stereotyped, and racially inaccurate depictions. Has this content been reviewed by someone with direct cultural knowledge of the community being depicted? For East African clients, this review is mandatory for all AI-generated imagery descriptions, character references, and community representations before client delivery.

---

The "Proof of Human" Standard

Apply this standard as the final test before sign-off (after Schaefer).

Content that passes this checklist reads as written by a human with genuine expertise in the subject and the market. It does not need to be perfect. It needs to be real.

Signs of genuine human authorship:

• Specific local details that only someone familiar with the market would include
• Clear opinions and recommendations, not false balance
• Natural imperfections — a conversational aside, a sentence that bends the grammar rule for emphasis, a direct address to the reader
• Cultural specificity — references that land for this audience, not for a generic global one
• Voice consistency with the client's existing content history

When in doubt, apply this rule: add a real detail, take a stronger position, cut a smooth transition. These three moves resolve most flatness problems.

---

Sign-Off Protocol

Junior Content (captions, short-form copy, email)

• Reviewed by a senior consultant before client delivery
• Senior consultant checks against the Human Voice Checklist and Cultural Localisation Checklist
• Approval is given per piece, not per batch

Strategy Documents and Proposals

• Reviewed by the lead consultant against the client brief, line by line
• Every recommendation cross-referenced against the client's stated objectives
• All market data footnoted with a verified source before the document leaves the agency

Production Record

• Note in the project file where AI tools were used in the production of any deliverable
• Record: the AI tool used, the prompt type (first draft, structural outline, research summary), and the extent of human revision
• This record is for internal quality tracking, not for client disclosure unless the client requests it

IP and Copyright Note

Source: Ching & Mothi (2025, p.82). Before client delivery of any substantially AI-generated piece, note in the production record that AI-generated content may not qualify for copyright protection without substantial human creative contribution. If the client intends to register or licence the work, flag for legal review by a qualified IP solicitor before proceeding.

Transparency Spectrum

Source: Ching & Mothi (2025, p.19). Where AI disclosure is provided, it must be specific, not generic. "Made with AI" is insufficient. The standard is: "AI-generated [specific element], art-directed and revised by [human team]" — or an equivalent level of specificity. Distinguish between the AI's contribution and the human's contribution in any disclosure statement. Vague disclosure misleads audiences and undermines trust in the work.

Volume Quality Gate

When AI is used to produce content at scale — multiple captions, a series of emails, a batch of blog posts — quality review becomes the bottleneck. A single editor reviewing too many pieces per session produces degraded attention and missed issues.

Apply these volume limits per editing session:

• Captions and short-form copy: maximum 15 pieces per session before a break
• Blog posts (800–1,500 words): maximum 3 per session
• Strategy documents: maximum 1 per session

If the content pipeline exceeds these limits, schedule additional sessions rather than compressing review time. Volume is not an excuse for reduced quality. The Golden Rule applies to every piece, not to a batch average.

---

Content Recycling Pipeline (Roth and neuroflash, 2024)

One well-researched piece of content can become 10 platform-ready assets in under 60 minutes:

| Step | Output | AI tool | |---|---|---| | 1. Master content (blog/article) | 800–1,200 word source piece | Claude/ChatGPT | | 2. Extract 5 key points | Bullet summary | Claude/ChatGPT | | 3. Facebook caption | 100–150 words, conversational | Claude/ChatGPT | | 4. Instagram caption | 50–80 words + hashtags | Claude/ChatGPT | | 5. LinkedIn post | 150–200 words, professional | Claude/ChatGPT | | 6. TikTok/Reels script | 30-second spoken script | Claude/ChatGPT | | 7. X/Twitter post | Under 280 characters | Claude/ChatGPT | | 8. Carousel outline | 5-slide structure with headlines | Claude/ChatGPT | | 9. Email snippet | 80-word newsletter paragraph | Claude/ChatGPT | | 10. Quote card text | Single compelling sentence | Claude/ChatGPT |

Platform adaptation checklist:

• Facebook: conversational, question or call-to-action, link preview works
• Instagram: emotional hook first, hashtags at end or in comment, no clickable links in caption
• LinkedIn: insight or lesson first, professional register, invite discussion
• TikTok/Reels: hook in first 3 seconds, spoken register, suggest trending audio
• WhatsApp Status: under 700 characters, image-first, direct call to action

Quality gate: Every variant must pass the content humaniser checklist before publishing. Volume without quality is worse than no content at all.

---

The Rising Quality Bar

Platform algorithms increasingly detect and suppress low-quality AI-generated content (Roth and neuroflash, 2024). The human layer — strategic direction, local cultural context, brand voice — becomes more valuable, not less, as AI content volume increases.

Signs that AI content will be flagged or underperform:

• Generic phrasing with no specific local context
• No genuine point of view or opinion
• Repetitive sentence structures
• Missing cultural references or local events
• Tone inconsistency across a single post

The content humaniser exists precisely to prevent these failure modes. Every AI output that passes through this skill becomes harder to detect, more engaging, and more trustworthy.

---

Quality Criteria

Output from this skill meets the required standard when:

1. No banned vocabulary remains — every word and phrase from the elimination list has been removed and replaced with direct, plain-language alternatives that serve the reader.

2. Every claim is verified — no statistic, citation, or factual assertion appears in the final copy without a confirmed primary source; invented or hallucinated data has been removed or replaced.

3. Cultural localisation is complete — the content reflects the Uganda/East Africa context specifically: appropriate payment references, local examples, correct currency, and a tone that matches Ugandan professional communication norms.

4. A clear human voice is present — at least one specific local detail, at least one clear opinion or recommendation, and natural sentence variation replace the AI draft's generic, even-toned structure.

5. The CTA is direct and actionable — every piece ends with a single, specific call to action that tells the reader exactly what to do, using the most relevant channel for the audience (WhatsApp, email, in-person visit, or other as appropriate).

6. The opening hook is earned — the first sentence or subject line creates genuine interest without relying on a question opener, a cliché, or any banned phrase.

7. The sign-off protocol has been followed — the correct level of review (junior or senior) has been completed and noted in the production record before the piece is delivered to the client.

---

References

• Ching, J. and Mothi, N. (2025) — Cultural Bias Audit; IP and copyright guidance; Transparency Spectrum; "Ineffable Something" test citing Tyler Cowen (2024).
• Evelyn, A. (2025) — Hallucination Management Gate.
• Ltifi, M. and Johnsen, S. (forthcoming) — AI content quality risks framework.
• Mizrahi, T. (2024) — Hallucination management.
• Roth, J. and neuroflash (2024) — Content recycling pipeline and rising quality bar framework.
• Schaefer, M. (2023) Belonging to the Brand — "Proof of Human" content standard.
• Chaffey, D. (2024) Digital Marketing: Strategy, Implementation and Practice.
• Bodnar, K. and Cohen, J. (2012) The B2B Social Media Book.