A Full AI TikTok Slideshow Workflow, Step by Step

The repeatable slideshow workflow is:

This is the practical version inside Ghostfeed.

Set up the project around the input

Ghostfeed does not force every idea through one generator:

Whichever route you choose, the output is the same kind of saved project: editable slides, backgrounds with source provenance, text boxes, caption, and a variant family.

Step 1: run a short analytical session

Do not begin with "give me 30 viral ideas." Begin with:

• yesterday's strongest hook and weakest drop;
• customer language from reviews or support;
• one audience problem;
• one product mechanism you can show;
• one format hypothesis for today.

The output is a brief for a deck, not a pile of captions.

Step 2: write the complete slide script

Give each slide one role:

Keep the deck useful without the caption. If the app suddenly appears on the final slide, the story did not integrate the product.

Step 3: create the Ghostfeed deck

For the agent-driven scratch route, use create_blank_slideshow with the target workspace and slide count. The returned slideshow ID is the saved deck that every later operation edits. In the dashboard, From Prompt and From Social create the first deck through the Director and open the same editor once it is ready.

The deck stays structured: slides, backgrounds, text boxes, caption, and variants.

Step 4: art-direct the backgrounds

Search for real scenes, inspect the candidates, and place one background per slide. You can use Pinterest, collections, product images, approved avatar images, or uploads.

The best slideshow backgrounds pass a camera-roll test: they look like moments a person could have captured, not a generated poster with baked-in copy.

Step 5: add text in TikTok-safe positions

Pass one to five editable text blocks with heading, subheading, body, or CTA roles and top, center, or bottom placement. Ghostfeed computes the safe-zone coordinates.

This makes copy changes cheap and keeps typography consistent.

Step 6: set the caption and inspect the saved result

get_slideshow returns every slide's text and background provenance. Read it back as a sequence and check:

• does slide one earn slide two?
• does every body slide add information?
• is the product the organizing answer rather than a surprise ad?
• are claims grounded?
• does the close match the promised job?

The full Ghostfeed slideshow walkthrough: from a source idea to a saved, editable photo-mode deck.

Step 7: create a controlled variant

Choose the ready base in Generate variants. Enter the topic or angle, choose up to the allowed count, and let Ghostfeed prepare distinct hooks. The base carries style, cadence, slide count, and text geometry; each hook controls the fresh copy.

Images do not all reroll. Pinterest-like backgrounds are recast, collections and avatars rotate within their saved pools, and deliberate uploads or AI images without an avatar remain fixed.

One base deck and one variant are more informative than ten decks where every variable changed.

Step 8: export, post, and read the signal

Ghostfeed's creation tools stop at a reviewable deck. Downloading, audio choice, and authorized platform posting remain explicit steps.

Log the result against the hook, beat order, images, and CTA. Tomorrow's analytical session should start there.

What automation must preserve

At larger volume, the loop above needs saved state and safe retries. That is the role of the technical workflow below.

An AI TikTok slideshow workflow is not:

Claude writes 30 scripts, an image model makes 240 slides, and a scheduler sprays them across ten accounts.

That is three batch jobs and a platform-risk problem.

A reliable implementation also has:

• a state machine;
• persistent artifacts;
• bounded retries;
• human approvals;
• authorized publishing;
• performance lineage.

The difference only becomes visible when a provider times out, a client changes a claim, or one bad slide appears in 20 queued posts.

The workflow in one line

Every arrow creates a saved artifact or a recorded decision.

1. Ingest evidence

A research item needs an ID and provenance:

The content model should cite these IDs when proposing a theme.

If a claim has no source, mark it speculative or reject it. Do not let a confident paragraph promote itself into customer truth.

2. Approve a creative hypothesis

One hypothesis contains:

• audience situation;
• single argument;
• hook mechanic;
• proof;
• format;
• success metric;
• prohibited claims.

Example:

This is the strategic approval. Do not hide it inside a 30-script batch.

3. Create the storyboard

Each slide gets:

• narrative job;
• exact overlay copy;
• visible scene;
• evidence;
• continuity requirements;
• rejection criteria.

Review the storyboard before rendering. Moving slide four to slide two costs nothing in a table and wastes six images after generation.

4. Generate slide candidates

Create an immutable attempt record:

Persist the output immediately.

Do not update asset-7312 in place after a retry. A new attempt produces a new asset. Otherwise the review history points to something that no longer exists.

5. Run mechanical QC

Automated checks can reject:

• wrong dimensions;
• corrupt files;
• missing safe-area space;
• unreadable product detail;
• visible generated gibberish;
• identity below a chosen similarity threshold;
• obvious mismatch with the slide spec.

Return reason codes:

Reason codes make retry prompts and reporting possible. "Looks off" does not.

6. Select and composite

A human selects the image candidate.

Then deterministic code adds:

• exact text;
• font;
• line breaks;
• contrast treatment;
• logo if appropriate;
• disclosure;
• crop;
• export dimensions.

Keep copy separate from pixels so a wording correction does not require image regeneration.

The selected image and final composite are two distinct assets.

7. Review the sequence

Slide-level approval does not imply sequence approval.

Check:

• promise and payoff;
• duplicate beats;
• visual continuity;
• factual continuity;
• pacing;
• product arrival;
• disclosure;
• CTA fit.

The reviewer can:

• approve;
• reject one slide;
• return the storyboard;
• request claim evidence;
• archive the concept.

Avoid a single approved: true boolean. You need to know what was approved and at which version.

8. Create the publishing package

The package contains:

• ordered final assets;
• caption;
• disclosure;
• alt text or source description;
• account ID;
• scheduled window;
• approval ID;
• campaign and hypothesis lineage.

Do not automate the last mile through unofficial device farms, proxy tricks, or account fingerprints.

TikTok's current integrity rules prohibit bulk account operation and high-volume spam intended to mislead or manipulate the platform. They allow multiple accounts for authentic creative expression, not deceptive networks. Read the current Integrity and Authenticity rules.

Use an authorized path. TikTok's Content Posting API supports direct posting and uploading a draft for the creator to finish. Its Direct Post flow requires creator information, appropriate UI, explicit user consent, and an audited client for public visibility.

For many teams, uploading a draft for human review is the right initial route.

9. Make the publish operation idempotent

Use a key such as:

Store:

• initiation response;
• upload state;
• platform post ID;
• final status;
• last checked time;
• terminal error.

If the request times out after upload, query status before retrying. "We did not receive the response" is not proof that the platform did not receive the post.

10. Join performance back to the hypothesis

Keep raw platform metrics and derived metrics separate.

Raw:

• views;
• likes;
• comments;
• shares;
• saves where available;
• profile visits where available;
• clicks and conversions from your own analytics.

Derived:

• completion rate;
• qualified profile visits per 1,000 views;
• conversion rate;
• cost per approved asset;
• cost per qualified outcome.

Snapshot raw data over time. Platform totals can update late, and definitions can change.

Then analyze by lineage:

That lets the next brief ask a useful question: "Did recognition hooks improve qualified visits for subscription-visibility concepts?" rather than "Which post went viral?"

Failure handling

Provider outage

• leave job queued;
• apply exponential backoff with jitter;
• stop at retry budget;
• alert a person;
• do not regenerate completed slides.

Claim changes

• invalidate affected slide specs and composites;
• preserve old versions;
• re-run sequence approval;
• block the stale publishing package.

Client delay

• pause upstream generation at the WIP limit;
• expire trend-dependent packages;
• do not build inventory forever.

Platform rejection

• store the exact response;
• do not disguise or repeatedly resubmit prohibited content;
• route to policy review.

Minimal data model

That may look like more engineering than "generate slides."

It is less engineering than reconstructing which prompt, claim, and image produced a bad post after it is public.

Ghostfeed already treats frames, videos, slideshows, and approvals as real state. The promise is not that no one has to think. It is that people spend their time approving the argument and the output, not copying files between disconnected tools.

For the creative spec behind this state machine, use the AI slideshow storyboard workflow. For capacity math, read AI slideshow production capacity.