A safe content pipeline automates research intake, drafting, checks, review, scheduling, and audit. Publishing remains a separately authorized effect. AI must not invent sources, testimonials, results, prices, or legal claims.

// concept

Define Artifact States

// prompt — copy me4 lines

BRIEF → SOURCED → DRAFT → CHECKED → REVIEWED → APPROVED
                                              ↓
                                          SCHEDULED → PUBLISHED
REJECTED | NEEDS_EVIDENCE | FAILED | ROLLED_BACK

Each artifact stores brief version, source URLs/snapshots where permitted, prompt/model version, claim checks, reviewer, approval hash, channel, schedule, publish ID, and rollback owner.

// concept

Build Separate Stages

authorized topic/brief intake;
retrieve allowlisted sources;
draft with source references;
deterministic brand/length/link/schema checks;
claim and copyright review;
human edit/approval bound to final artifact;
publish using least-privilege channel credential;
verify live URL/content and monitor;
rollback/correct if required.

Do not scrape or reproduce copyrighted material beyond permitted use. A citation must support the nearby claim and be reachable.

Use a publication ledger keyed by channel and artifact ID. It records approval hash, scheduled time, external post ID, canonical URL, verification result, and correction or withdrawal state. This lets retries locate the existing draft or post and gives editors one place to audit what actually reached an audience.

// worked_example

Worked Example

A Pakistani AI school drafts a blog update from official n8n documentation. The workflow retrieves approved URLs, produces a DRAFT with links, and flags pricing/licensing claims for dated review. A human removes an unsupported “cheapest” claim and approves the exact hash.

The publisher creates one CMS draft, not a live post, until a separate publish approval. After publishing, the workflow fetches the canonical URL and checks title, H1, links, and status. A retry uses the CMS artifact ID, preventing duplicates.

// failure_cases

Failure Cases to Diagnose

6 cases to diagnose

AI-generated URL accepted
verify reachability and support.
Approval precedes final edit
invalidate hash after change.
Publisher credential is administrator
scope to required channel/action.
Retry creates duplicate post
use external ID/idempotency.
No live verification
fetch canonical output.
Old price/stat reused
date and revalidate unstable claims.

// pakistan_angle

Pakistan Angle

English and Roman Urdu copy must preserve the same offer, terms, and consent. Do not invent Pakistan market statistics, client results, income promises, or testimonials.

For religious, legal, health, finance, and political content, require appropriate expert/editorial review or exclude it according to product policy. Protect customer examples and obtain permission.

// hands_on

Hands-On Exercise

5 steps

Define content states and metadata.
build source → draft → checks → review.
bind approval to final hash.
publish to a synthetic/local CMS draft.
test duplicate, broken source, post-approval edit, and rollback.

// completion_rubric

Completion Rubric

6 checks — tick as you verify

0/6

// sources

Sources

3 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-generated URL accepted.” What does the lesson tell you to do about it?