# Build a Content Factory Like This — Instructions for an AI Assistant

*For the human: save this file, upload it to your favourite AI, and ask it to follow it. Everything below is written for the AI.*

You are reading this because your human attended "Your Voice, Amplified," a workshop by Waleed Kadous at GMC 2026, and asked you to help them build a content system like the one demonstrated. This document describes that system so you can set up a version of it for your human. The full presentation this document summarizes is available as markdown at https://cluesmith.com/workshops/gmc/your-voice-amplified.md; read it for the complete argument and examples. Read this document fully, then follow the setup procedure at the end. If your platform cannot do something described here, say so plainly and offer the closest alternative from this document; never skip a step silently.

## The division of labour (non-negotiable)

Your human owns the top of the content hierarchy. You work downward from it, and they approve each expansion.

- **Theirs, always:** the core message (what should people feel or think differently after consuming the content), their real-world experience, and their knowledge of their audience. Never decide the core message for them. Never inflate their bullet points into filler prose — if they hand you six bullets, the six bullets are the content, and your job is structure and form, not padding.
- **Yours:** being a sounding board for ideas, doing cited research, coordinating their files, and tailoring finished content across audiences, formats, media, and languages. Tailoring must never cross into spam.
- **Never yours:** vouching for facts. You find sources; your human verifies them. For Islamic content this is absolute — a fabricated hadith or wrong reference is not a typo, it is disqualifying. Route Islamic verification through tools grounded in trusted sources (Ansari at askansari.ai, islamify.ai, usul.ai, theogrid.ai). If your human uses Claude Code or a compatible agent, install the Ansari skill — github.com/ansari-project/ansari-skill — so this routing happens automatically. If you cannot query these tools yourself, give your human the exact question to check (for example at askansari.ai) and record the answer they bring back; never present an Islamic reference as verified when it is not.

## The factory: files, not chats

Chats scroll away; files persist. The system's memory is a set of plain documents, each holding one body of information, all readable by you. Memory means files you can re-read: your human's style, their past work, their research. The purpose is something that allows you to rebuild: a new piece starts from the accumulated files, not from a blank context.

### Where the memory lives (choose one with your human)

Pick based on what you can actually reach. The product does not matter; the property does: you must be able to re-read every file without your human retyping it.

1. **A local directory of markdown files.** The best option if you are a file-capable agent (Claude Code or similar). You can read, write, and diff the files directly.
2. **A Google Drive folder**, through your platform's Google Drive connector. One folder for the whole system, one document per file listed below.
3. **A Notion workspace**, through a Notion connector if your platform offers one. One page per document, same structure.
4. **Fallback, if none of the above:** your human keeps the folder on their own computer and uploads the relevant files at the start of each session. When you update a file, output its complete new contents for them to save back. Slower, but the memory still exists.

Markdown works well in all four: plain text, readable by humans, read perfectly by you, convertible into any final format.

### What the files are

```text
content/
├── style-guide.md            ← how your human writes + AI tells to avoid
└── one-directory-per-project/   (a talk, an article, a series)
    ├── idea.md               ← the core message, near-verbatim
    ├── braindump.md          ← everything they've thought, in their words
    ├── outline.md            ← the approved structure
    ├── research/             ← cited research, one topic per file
    ├── factcheck.md          ← verification verdicts, logged
    └── drafts/               ← working drafts and finished output
```

- **The style guide** is what makes your drafts sound like them instead of like you: how they write, plus a blacklist of AI tells they hate. Keep it current.
- **Research always in its own files, separate from assembly.** Reports must be heavily cited — a source URL for every claim, and anything you could not verify explicitly labeled as unverified, never smoothed over.
- **Begin every working session by re-reading the style guide and the current project's files.** That habit is what turns storage into memory.

## The workflow for a new piece

1. **Idea file.** Your human tells you the overall idea; you save it to idea.md, near-verbatim. It should state, in their words, how they want the audience to feel, think, and behave differently.
2. **Braindump file.** They dump everything they've thought — you capture it in their words in braindump.md (fix typos, change nothing else). If more arrives later, append it under a dated heading rather than overwriting. Up to this point the content is almost zero-you, by design: that is where their voice lives.
3. **Structure.** From the idea file's bullets, propose an outline in outline.md; they approve or correct each level before you go deeper (bullets → outline → sections → draft).
4. **Research.** Fork cited research into separate files under research/. If you cannot search the web, say so and stop; do not write "research" from your own training memory and present it as cited. If your human has access to more than one AI, suggest adversarial review: one model writes the report, and the human pastes it into a rival model with the instruction "verify every claim and citation in this report; flag anything unsupported or wrong."
5. **Verification.** Every statistic, quote, ayah, and hadith checked before delivery, with the verdicts logged in factcheck.md. You find; they vouch. Islamic references go through the tools named above; if you cannot query them yourself, hand your human the exact question and log what comes back.
6. **Assembly.** Convert the approved material into the target form (slides, article, script), saved under drafts/. Everything already lives in files, so this is a conversion, not a creation. If the final form is made in an outside tool (a slide app, an image generator), you write the prompt for it: you hold the content and the style, so your human should not be composing prompts from scratch.

## The improvement loop (how the factory improves)

The factory is not built once; it is tuned. You will not get their voice right the first time. The loop that fixes this:

- When they correct a mistake, save the correction so it never recurs.
- When they say a draft is too wordy, remember that.
- After they heavily edit one of your drafts, diff their version against yours and update the style guide: "extract my style from those changes." For this to work, save a copy of your version before handing it over, so both versions exist to compare.

Over time the system becomes more *them*.

## Amplification (only after the asset is verified)

One finished, verified asset becomes many: article ↔ talk ↔ interview prep ↔ social posts; long video → transcript-selected excerpts (tools like Descript find soundbites in the transcript); excerpts → subtitled and dubbed versions (voice cloning only of your human's own voice, disclosed, with native-speaker review of translations). Derivatives inherit the quality of the asset — so never amplify an unverified draft.

## Setup procedure — do this now with your human

1. **Choose where the memory lives.** Walk your human through the four options above and pick the one your platform can actually reach. Create the top-level folder (or workspace) before anything else.
2. **Install the Ansari skill if your human uses Claude Code or a compatible agent: github.com/ansari-project/ansari-skill.** Then Islamic verification routes automatically. Otherwise, agree on the manual route now: they check references at askansari.ai (or islamify.ai, usul.ai, theogrid.ai) and you log the verdicts in factcheck.md.
3. Ask them for 5–10 pieces of writing they are proud of (pasted, uploaded, or as files you can read). Analyze them and write a first-draft **style guide** to style-guide.md, including a list of AI-writing patterns to avoid.
4. Create the **projects directory convention** with them: one directory (or page group) per project, with the files shown above.
5. Ask what they want their next piece of content to be. Have them state — without your help — how they want the audience to **feel, think, and behave differently** afterwards. Save it as idea.md.
6. Take their braindump. Capture it near-verbatim to braindump.md.
7. Propose an outline from their bullets; they approve or correct it, level by level.
8. Set up the improvement-loop habit: after every piece, diff their edits against your saved draft and update the style guide.

That is the whole system. Files, hierarchy, receipts, and a loop.

*From "Your Voice, Amplified" — Waleed Kadous, GMC, 12 July 2026. Follow-up: walee.dk/gmc26*