📄 MTC article pipelines
article-pipeline-8-stages · version 16 · experimental

The 8-stage pipeline, step by step

The same 20 jobs as the default pipeline, packed into 8 bigger steps, in a slightly different order. Same helpers, same rules, same checks.

Jump to a chapter

What it is

🧪

Experimental

The 20-stage pipeline is still the default. Pick this one to test the new order.

🧰

Same helpers and rules

The same scripts, examples, limits and checks. A test fails if any of the old stages stops being covered.

🧩

New grouping

Each kind of job happens together: all editing at once, all checks at once, all pictures at once.

⚖️

No promises

Its own instructions say it does not promise faster, cheaper or equal quality. Compare with the same model and similar keywords.

The map: 20 jobs in 8 boxes

🔎

1. Research and plan

1
✍️

2. Write the draft

2
🧽

3. Edit and normalize

4 6 15 16
🔗

4. Connect articles

3 18
🧪

5. Review and repair

5 7 18
🖼️

6. Finish images

9 10 11 14 19
🏷️

7. Assemble the page

8 13 15 17
✅

8. Verify the result

20

The run, in old stage numbers

½ means half of stage 18; ² means stage 15 runs a second time.

1 › 2 › 4 › 6 › 15 › 16 › 3 › 18½ › 5 › 7 › 18½ › 9 › 10 › 11 › 14 › 19 › 8 › 13 › 15² › 17 › 20

The 8 stages

1Learn before writing2Write from the brief3Four separate editing passes4Connect the topic5Check and repair6Complete image work7Finish the page8Verify before saving
1
Stage 1 of 8 · Learn before writing

Research and plan 🔎

🔎 Research🤖⚙️ The AI writes, a script checks

Before writing anything, the AI studies Google's top results, collects facts from trusted sources, and writes a plan of what the article must cover and how long it should be.

Does these 20-stage jobs: 1

The same job as stage 1 of the 20-stage pipeline, plus an extra check that MTC can record the pipeline and version.

📥 Goes in

  • Your keyword and site
  • The full content map
  • Writer, style and writing model

⚙️ What happens, step by step

  1. The AI checks MTC's save tools, and makes sure the save tool can record the pipeline name and version. If MTC is too old, it stops and asks for an update instead of losing that record.
  2. It reads the site settings, the article list and the whole content map (every page, not just the first page of results), and prepares the list of possible link targets.
  3. It locks in the writing model, the pipeline (article-pipeline-8-stages, version 16), the writer and the style.
  4. Rewrites: if the page already ranks in Google's top 3, it asks you before touching it.
  5. 3 fresh Google searches, averaged. The first 3 real articles with at least 300 words become the competitors (big sites like Reddit, Amazon, news, .gov and .edu never count).
  6. From them: a word budget of 98 to 110 percent of the longest, their average link count, and a table budget.
  7. TextRazor finds the important "things" to mention. Product lists check products on Amazon.
  8. The AI writes the research brief: 13 sections, 4 to 6 facts with exact quotes from trusted sources, every "People also ask" question.

📤 Comes out

  • The research brief (13 sections)
  • Word budget, table budget, link average
  • The list of possible links

🚦 What it checks, and what stops it

  • Same stops as the 20-stage research: wrong model, changed version, top-3 rewrite, quotes not word for word.

🧰 Tools it uses

  • Google (through SerpAPI)
  • TextRazor
  • Amazon (product lists only)
  • MTC
2
Stage 2 of 8 · Write from the brief

Write the draft ✍️

✍️ Write🤖⚙️ The AI writes, a script checks

The AI writes the whole article from scratch, using only the facts in its plan, and copying the layout (not the words) of a model article.

Does these 20-stage jobs: 2

The same job as stage 2. It also fills every required setting from the start, and keeps one FAQ in the settings box from the beginning.

📥 Goes in

  • The research brief
  • The word budget
  • A model article of the same type

⚙️ What happens, step by step

  1. Keeps a safety copy of the old article (rewrites), and only its identity: dates, author, category, type, tags and main picture.
  2. Reads the model article for this type and copies its layout only.
  3. Spends the word budget in order: the direct answer, readers' questions, important facts, gaps, related searches, background. Aims for the low end.
  4. Quick Answer of 40 to 60 words. Paragraphs of 2 sentences at most. Reading grade 6 to 8.
  5. At least 3 facts with the source named in the sentence, and about 8 links to other pages.
  6. Unlike the 20-stage draft, it fills every required setting now, even though titles and dates are finished later in stage 7.
  7. It prefers one FAQ in the settings box from the start, and already follows the punctuation rules (later checks still look).
  8. Self-check for 6 common mistakes, then fix and check again.

📤 Comes out

  • The complete first draft, with all required settings

🚦 What it checks, and what stops it

  • Same as the 20-stage draft: no copied text, the right layout, tables within budget.

🧰 Tools it uses

  • The AI writer only (no searches)
3
Stage 3 of 8 · Four separate editing passes

Edit and normalize 🧽

🧽 Edit🤖 The AI writer does it

The AI rereads its draft four times, each time with just one job: sound human, answer clearly, remove long dashes, and keep one FAQ.

Does these 20-stage jobs: 4 6 15 16

Four 20-stage jobs, done right after the draft and before linking: 4 (sound human), 6 (clear answers), 15 (dashes) and 16 (one FAQ). Each is still its own separate pass. Stage 15 runs again later, in stages 7 and 8.

📥 Goes in

  • The first draft

⚙️ What happens, step by step

  1. Pass 1, naturalness (old stage 4): finds textbook wording, jargon, filler and piles of facts, improves only weak parts, adds bridges between ideas, and ends sections on a takeaway.
  2. Pass 2, answer clarity (old stage 6): each section opening answers right away; fuzzy words are made clear; sentences with sources stay frozen.
  3. Pass 3, punctuation (old stage 15): every extra-long dash becomes a comma, a period or a better sentence.
  4. Pass 4, FAQ (old stage 16): keeps the FAQ in the settings box, moves extra questions in, removes repeats, keeps sources clickable.
  5. The four passes are never merged into one big rewrite. Each is logged on its own.
  6. Why so early? To settle the words and the FAQ before linking and the big measured checks.

📤 Comes out

  • The edited article, with one FAQ
  • A separate note for each pass

🚦 What it checks, and what stops it

  • No helper script. The measured checks that catch problems come in stage 5.

🧰 Tools it uses

  • The AI writer only
4
Stage 4 of 8 · Connect the topic

Connect articles 🔗

🔗 Links🤖⚙️ The AI writes, a script checks

The AI adds links from the new article to related pages, and asks up to five older articles to add one link back.

Does these 20-stage jobs: 3 18

Stage 3 (links out of the article) plus the neighbor-link half of stage 18 (links into it). Stage 18's measuring half moves to stage 5.

📥 Goes in

  • The edited article
  • The content map and link list

⚙️ What happens, step by step

  1. Family links first: big topic pages, middle pages and articles must link to each other in the main text.
  2. Then useful links up to about 8 different pages, same topic first. Links to another big topic need a written reason.
  3. Links to planned pages wait as plain text until those pages are published.
  4. Natural link text of 2 to 5 words, never "check out our guide".
  5. Neighbor links (from old stage 18): look at up to 10 older articles, pick at most 5, add one link each, changing only one sentence, after checking the file has not changed.
  6. The helper checks every link is recorded and every family link is in place.

📤 Comes out

  • About 8 useful links out
  • Up to 5 older articles linking back

🚦 What it checks, and what stops it

  • Blocks: missing family links, unrecorded links, wrong map pages. Warns: fewer than 8.

🧰 Tools it uses

  • The AI writer
  • A helper script that checks the links
5
Stage 5 of 8 · Check and repair

Review and repair 🧪

🧪 Check🤖⚙️ The AI writes, a script checks

The AI checks the article like a strict teacher with a ruler: it counts the words, measures how easy it is to read, makes sure every important topic is there, and fixes problems a limited number of times.

Does these 20-stage jobs: 5 7 18

Stage 5 (the big quality check), stage 7 (important "things") and the measuring half of stage 18. Here, clear answers (old 6) already happened in stage 3.

📥 Goes in

  • The linked article
  • The page as a reader sees it
  • The research brief

⚙️ What happens, step by step

  1. It starts your site on your Mac and measures the page as a reader sees it.
  2. The QA checklist (old stage 5): 19 areas from length to trust. Links are judged by the link rules from stage 4, so a useful link to another topic is never removed.
  3. If QA adds links, the link checker runs again.
  4. The entity audit (old stage 7): TextRazor counts important things; tiny one-sentence fixes; 1 main and 3 to 6 other things labeled with Wikipedia links.
  5. The final measurements (old stage 18): words, reading ease and entities, with one extra entity fix allowed after trimming.
  6. Fix limits: up to 6 trims (at least 40 words each), 3 grows, 2 reading-ease fixes, 2 entity fixes. Problems left over are written down, never hidden.
  7. Measurements are always run fresh. Old results are never reused just because they look similar.

📤 Comes out

  • A repaired article
  • 1 main + 3 to 6 important things labeled
  • Honest notes on anything left

🚦 What it checks, and what stops it

  • Blocks: paragraphs over 2 sentences, table count or placement, pillar FAQ rules.

🧰 Tools it uses

  • Local preview of your site
  • TextRazor
  • Measuring scripts
6
Stage 6 of 8 · Complete image work

Finish images 🖼️

🖼️ Images🤖⚙️ The AI writes, a script checks

Now that the words are final, the AI decides which pictures the article needs and where they go, and an AI picture service paints them.

Does these 20-stage jobs: 9 10 11 14 19

All five picture jobs, done together after the words are final: 9 (plan), 10 (hero), 11 (place), 14 (count check; the shot-list file is optional) and 19 (make). Images off skips this whole stage.

📥 Goes in

  • The settled, checked article

⚙️ What happens, step by step

  1. Plan (old stage 9): about 1 body picture per 350 words, photo descriptions, alt text, sizes 1200 x 630 and 800 x 500.
  2. Hero (old stage 10): the cover picture's path and description go into the settings.
  3. Place (old stage 11): safe spots between paragraphs, about 330 to 350 words apart, checked as a whole before inserting.
  4. Recheck (old stage 14): the picture count is always rechecked against the current article. The shot-list file itself is optional ("optional export not needed").
  5. Make (old stage 19): fal.ai paints each picture one at a time, never paying twice; cropped, saved as WebP quality 82; reviewed by the AI.
  6. Why so late? So pictures are planned on the final words, not on a draft that will change.

📤 Comes out

  • Finished pictures, placed in the article
  • The picture plan

🚦 What it checks, and what stops it

  • Blocks a bad placement plan or a duplicate file name. Missing pictures fail the final check.

🧰 Tools it uses

  • fal.ai image generation
7
Stage 7 of 8 · Finish the page

Assemble the page 🏷️

🏷️ Page parts🤖⚙️ The AI writes, a script checks

The AI adds the finishing labels: the date, the title and description people see on Google, one more dash clean-up, and the table of contents.

Does these 20-stage jobs: 8 13 15 17

Stage 8 (date), stage 13 (titles), stage 15 again (dashes in everything written since stage 3) and one table-of-contents build that stands in for both 12 and 17. A test proves one build gives the exact same result as two.

📥 Goes in

  • The illustrated article

⚙️ What happens, step by step

  1. Date (old stage 8): "last updated" is set; the original publish date is kept.
  2. Titles (old stage 13): Google title 53 to 59 letters with the keyword first, headline 50 to 70, description 140 to 160. Checked for length and never copied.
  3. Dashes again (old stage 15): every sentence written after stage 3 gets checked: fixes, neighbor sentences, picture descriptions, titles, Quick Answer and FAQ answers.
  4. Contents list (old stages 12 + 17): built once, at the very end, instead of twice. Read time at 238 words a minute.
  5. Any later change to headings or the FAQ means building the contents list again.

📤 Comes out

  • Dates
  • Google title, headline and description
  • The table of contents and read time

🚦 What it checks, and what stops it

  • Blocks titles outside the letter ranges, or a Google title equal to the headline.

🧰 Tools it uses

  • The AI writer
  • Helper scripts for dates, title lengths and the contents list
8
Stage 8 of 8 · Verify before saving

Verify the result ✅

🧪 Check🤖⚙️ The AI writes, a script checks

The AI builds the real website, looks at the page on a computer and a phone, runs the automatic checklist, and only then saves the article as "ready".

Does these 20-stage jobs: 20

The same job as stage 20, with the same full scope, plus one more dash clean-up after any final fixes. Then the article is saved as ready.

📥 Goes in

  • The finished article
  • Every changed file

⚙️ What happens, step by step

  1. Builds your real website and checks every link in it. A failed build fails the article.
  2. Looks at the article and the edited neighbors at computer and phone sizes.
  3. The script checks everything (links, files, pictures, Amazon codes, settings) and gives a grade: validated, validated with warnings, unverified or blocked.
  4. Up to 2 final fixes, and any repaired sentences get one more dash clean-up.
  5. Saves through MTC as a "ready" bundle on your site's GitHub, then confirms MTC says "saved" with the right pipeline and version.
  6. Publishing is a separate step you choose later.

📤 Comes out

  • A final grade
  • The article saved as ready on GitHub

🚦 What it checks, and what stops it

  • Only validated or validated with warnings can be saved.

🧰 Tools it uses

  • Your site's real build
  • Link checks over the internet
  • Amazon (product links)
  • MTC

Where every 20-stage job went

20-stage #JobDone in 8-stage stepWhat changes
1Research and entity brief1. Research and plansame job
2SEO optimized article write2. Write the draftsame job
3Internal link fix4. Connect articlesmoved after the four editing passes
4Humanization rewrite3. Edit and normalizemoved right after the draft
5QA check5. Review and repairnow after the clear-answers pass
6AI SEO optimization3. Edit and normalizemoved before linking and QA
7Entity audit and frontmatter5. Review and repairsame job
8Last updated date7. Assemble the pagemoved after the pictures
9Image prompts6. Finish imagessame job
10Image formatting6. Finish imagessame job
11Image placement6. Finish imagessame job
12Table of contents7. Assemble the pagemerged: the contents list is built once
13SEO title, H1 and meta description7. Assemble the pagemoved after the pictures
14Image prompt export6. Finish imagesthe shot-list file is optional; the count recheck always runs
15Dash punctuation rewrite3. Edit and normalize 7. Assemble the pagedone twice: in stage 3, and again in stage 7 (plus after final fixes)
16Duplicate FAQ merge3. Edit and normalizemoved right after the draft
17Regenerate TOC7. Assemble the pagethe one final contents build
18Inbound internal links4. Connect articles 5. Review and repairsplit: neighbor links in stage 4, measurements in stage 5
19Image generation6. Finish imagessame job
20Final validation8. Verify the resultsame job

Version history, in plain words

VersionWhat changed
v1The 8-stage pipeline is born. Both pipelines start recording which pipeline and version wrote each article.
v2Two-sentence paragraphs, easier reading (grade 6 to 8), US research. Pillar pages get their own pictures.
v3Body pictures spaced 330 to 350 words apart.
v4New pillar web addresses come from their keyword.
v5Picture placement cares about balance and meaning more than exact spacing.
v6A friendlier, reader-first voice.
v7Pillar pages get a Quick Answer box.
v8Length and links measured against the top 3 real competitors; about 8 useful links.
v9Even more conversational: one reader, short paragraphs, varied rhythm.
v10The top-3 guard: ask before rewriting a page that already ranks well.
v113 averaged searches, big sites excluded, a table budget.
v12Pillar pages get a click-to-open FAQ.
v13Section flow: every section reads as one smooth explanation.
v148-stage only: dash clean-up repeated in stages 7 and 8, picture count always rechecked, length checks for pillars. The images-off twin arrives.
v15Records who wrote the article and which writing style it followed.
v16Quick Write can link to pages on a domain you name.