An evidence-safe toolkit connects every factual claim to proof and every person’s image, message, or story to permission. It prevents corrections, copyright complaints, and privacy breaches that damage trust.

You will build folders, a toolkit map, evidence and consent logs, and an AI no-fabrication instruction. Local folders and a spreadsheet are enough; paid software is optional.

// concept

Give Each Tool One Job

Assign narrow roles instead of making one AI app researcher, writer, designer, fact-checker, and publisher.

Role	Tool options	What goes in	What must come out
Research	Browser, Perplexity, or search	A question without private data	Direct URLs, dates, and notes
Draft	ChatGPT or another assistant	Verified claim IDs and a brief	Copy with claim IDs and uncertainty marked
Design	Canva or a local editor	Approved copy and owned/licensed assets	Export plus an asset-rights record
Schedule	Native scheduler or calendar	Approved asset and caption	Platform, time, owner, and status

Use those columns for your toolkit map. Check each cloud tool’s current official privacy controls. ChatGPT documents Settings → Personalization → Custom Instructions on web/desktop and Settings → Customize ChatGPT on mobile; interfaces change, so confirm the page. A local standing-instruction.txt is the free alternative.

// concept

Build Folders That Preserve Evidence and Permission

Create this structure locally or in storage you control:

// prompt — copy me9 lines

content-system/
  01-idea-inbox/
  02-evidence/source-log.csv
  03-consent/consent-register.csv
  04-drafts/
  05-design-assets/rights-log.csv
  06-approved/
  07-schedule/publishing-log.csv
  standing-instruction.txt

Restrict 03-consent to the smallest necessary team. Google Drive folder permissions are inherited, so use a limited-access folder for consent evidence. Never make consent screenshots “anyone with the link.” Use local encrypted storage when cloud access is unsuitable.

Use one inbox row per post: idea_id, audience_question, format, owner, status. Draft only after factual claims have evidence IDs.

// concept

Create the Evidence and Consent Logs

The evidence log records why a claim is publishable. Give each source its own row.

Field	Example — sample data	Rule
`claim_id`	`C-001`	Stable ID used in the draft
`claim_text`	`Dimensions: 30 × 20 × 15 cm`	Exact, testable wording
`source_type`	`supplier specification`	Prefer primary or official evidence
`source_url_or_file`	`supplier-spec-v2.pdf`	Direct URL or file path
`publisher_or_owner`	`Sample supplier`	Evidence issuer
`published_or_captured`	`2026-07-18`	Source date
`checked_on`	`2026-07-18`	Verification date
`status`	`approved`	`approved`, `reject`, or `needs-check`
`note`	`Page 2; model ABC only`	Scope or contradiction

For screenshots, save source, date, and context—not a cropped number. Quote only what verifies the claim; a source is not copying permission.

The consent register needs consent_id, person_alias, content, allowed_channels, allowed_edits, granted_on, expires_on, proof_file, and status. WhatsApp repost permission does not cover a face, number, voice, or advertisement. Redact unnecessary data before using AI or design tools.

Save this standing instruction in the AI tool or local file:

// prompt — copy me8 lines

Use only facts tied to the evidence IDs I provide. Never invent a statistic,
quote, customer result, price, trend, source, or permission. Keep each evidence
ID beside the sentence it supports while drafting. If evidence is missing,
write [EVIDENCE NEEDED] instead of guessing. Do not infer consent or usage rights.
Flag names, faces, voices, phone numbers, addresses, payment details, CNIC data,
private messages, and children’s data for removal or explicit approval. Do not
copy source wording beyond a short necessary quotation. End with a list of every
claim ID and consent ID used so a human can approve the final version.

// worked_example

Worked Example

This hypothetical uses sample data for an unnamed Daraz seller. The supplier sheet supports dimensions under C-001. A private customer message says “fully waterproof” and shows a phone number, but there is no product evidence or consent.

The first drafting input is deliberately unsafe:

// prompt — copy me2 lines

Write a Daraz carousel saying this organizer is fully waterproof.
Use this customer screenshot as proof: [sample private screenshot].

A weak output repeats “fully waterproof” and exposes the number. That creates both a privacy defect and an unsupported product claim.

The fixed input uses the toolkit:

// prompt — copy me6 lines

Create a four-slide carousel for a sample home-organizer listing.
Approved evidence: C-001 = dimensions 30 × 20 × 15 cm from supplier-spec-v2.pdf.
Rejected claim: “fully waterproof” — no primary evidence. Do not include it.
No customer material has consent. Do not quote, describe, or upload it.
Keep [C-001] beside the supported sentence and mark any other fact
[EVIDENCE NEEDED]. Return a claim-ID list after the draft.

The corrected draft says: “Measure your shelf first: this model is 30 × 20 × 15 cm [C-001].” It omits “waterproof” and the customer. rights-log.csv records the seller-owned product photo; the scheduler receives only the approved export and caption.

// failure_cases

Failure Cases to Diagnose

6 cases to diagnose

AI citation without opened source
replace the answer summary with the direct document and supporting section.
One source attached to five claims
split the copy into claim IDs; reject any sentence the source does not explicitly support.
Consent folder shared widely
limit access and share only redacted, approved assets downstream.
Screenshot treated as permanent permission
check channels, edits, commercial use, and expiry.
Rights unknown for an image
record ownership or licence, or replace it.
Scheduled draft differs from approval
compare its file hash or version name with the publishing log.

// pakistan_angle

Pakistan Angle

A CNIC image, Pakistani mobile number, home address, JazzCash or easypaisa receipt, bank detail, and WhatsApp photo are private data—not content material. Never paste them into an AI prompt. Keep an original only for a legitimate need; redact the CNIC number and barcode, phone, transaction ID, balance, address, and unrelated messages from the working copy.

Ask permission in a language the person understands: state the content, platform, advertising use, and duration. A casual “theek hai” for WhatsApp Status is not permission for YouTube or paid promotion. During load-shedding, update local CSVs and sync later; verify restricted sharing before scheduling. If USD pricing or card access blocks a tool, folders, LibreOffice, a calendar, and native platform tools preserve the process. Check current vendor pages before assuming a free or paid feature.

// hands_on

Hands-On Exercise

7 steps

Create the seven folders and standing-instruction.txt exactly as shown.
Build toolkit-map.csv with the four roles, chosen tool, input boundary, output, reviewer, and free alternative.
Add one sample idea to the inbox and divide its copy into at least two claim IDs.
Add one approved and one rejected row to source-log.csv; include the reason for rejection.
Create the consent register and add one sample not-granted row without real personal data.
Run the standing instruction with your approved claim IDs. Save the draft with claim markers, then create a redacted approval copy.
Finish when another person can trace every factual sentence and asset without opening private data.

// completion_rubric

Completion Rubric

6 checks — tick as you verify

0/6

// sources

Sources

5 official sources — check every claim yourself

// check_yourself

Check yourself

4 questions · answers and options are taken word-for-word from this course

0/4

1 / 4 · diagnose
Your work shows this failure mode: “AI citation without opened source.” What does the lesson tell you to do about it?