Content agents: plan once, check every post

01A content agent is a prompt pipeline

Chapter 6 has no code. It is nine Markdown files: a planning sheet, seven platform templates and a README with the voice rules. Read them as a tester and they are a pipeline you already know: an input, a spec, a generator and an oracle.

flowchart LR
  I["Idea"] --> P["00 plan"]
  P --> G{"Final gate"}
  G -->|fail| P
  G -->|pass| T["Plan + template"]
  T --> L["LLM writes"]
  L --> D["Draft"]
  D --> C{"Checklist"}
  C -->|fail| L
  C -->|pass| S["Publish"]

The README gives the one instruction the model needs, after you paste the filled plan and a template: "Write the [platform] piece using this plan and these template rules. Follow the format exactly. No banned phrases." Its summary: the template is the spec; the plan is the content.

02Hook, Story, Offer

The planning sheet reduces every post to three moves, in its own words:

Every piece of content is three moves: **stop them (Hook), take them somewhere (Story), ask for
something (Offer).** Get these three right on paper and the platform formats become copy work.
Get them wrong and no amount of editing saves the post.

Fill this out **once per idea.** Everything else in this chapter pulls from it.

Hook: six patterns, each with an example

### Hook patterns that work

| Pattern | Shape | Example |
|---------|-------|---------|
| **Blunt declarative** | A flat, confident claim that dares disagreement | "Your 5,000 tests are not an asset. They're a liability." |
| **Unpopular opinion** | Name the consensus, then break it | "Unpopular opinion: more coverage is making your suite worse." |
| **The reader's own thought** | Say the thing they're already half-thinking | "AI assistants get dumber after a few hours. You've noticed too." |
| **Cold-open story** | Drop them mid-scene | "A candidate walked into my interview yesterday and…" |
| **The reversal** | Set an expectation, snap it | "I automated everything. Then I deleted half of it." |
| **The number** | A specific, surprising figure (must be real) | "We cut flaky reruns 80% by changing one line." |

The hook rules: write five versions and keep the sharpest; no statistic as the opener unless the statistic is the shock; never a table of contents such as "In this post I'll show you"; cut every word that delays the punch.

Story: Problem, Tension, Turn, Proof

```
PROBLEM (their pain, their words): ___________________
TENSION (why the easy answer fails): _________________
TURN (your insight / method):       _________________
PROOF (number, demo, story, code):  _________________
```

One core idea per piece; side ideas become future posts. Specifics over adjectives: "Reduced runtime from 41s to 12s" beats "much faster". Add one screenshot-worthy line and one honest caveat that cuts against your own interest.

Offer: exactly one ask, lightest rung first

### Offer rules
- Match the platform: LinkedIn hides body links (put links in the first comment); Instagram
  uses comment-keyword DMs; YouTube wants a one-word comment + subscribe.
- The CTA should feel like a favor to the reader, not a favor to you.

03Live demo: plan a QA post and lint it

The plan starts filled with a sample QA topic. Change the hook pattern to load the repo's example line for it, edit any field, and watch the Final gate. Then generate a draft for LinkedIn or Instagram and edit it: the pre-publish checks rerun on every keystroke. Try the offer rungs too: the LinkedIn template wants the post to end on a specific PS question, so rungs 2 to 5 fail that one check there, while the Instagram caption takes any single CTA.

04Plan once, repurpose everywhere

The chapter README's rule: you do not write seven things, you write one Hook, Story and Offer and bend it into seven formats.

flowchart TD
  P["00 Hook, Story, Offer: plan the idea once"] --> YT["01 YouTube"]
  P --> RE["02 IG Reel"]
  P --> PO["03 IG Post"]
  P --> CA["04 Carousel"]
  P --> ME["05 Medium"]
  P --> BL["06 Blog"]
  P --> LI["07 LinkedIn"]

The repurpose map says where each part of the plan lands on each platform:

## The repurpose map

Once Hook · Story · Offer is locked, here's where each piece lands:

| Plan element | YouTube | IG Reel | IG Post | Carousel | Medium | Blog | LinkedIn |
|---|---|---|---|---|---|---|---|
| **Hook** | 0–15s hook | Slide-1 text + first spoken line | First caption line | Slide 1 | Title + bold subtitle | H1 + intro | First 2 lines |
| **Story** | Body beats | 2–4 quick beats | Caption body | Slides 2…N-1 | ## sections | ## sections | → arrow lines |
| **Proof** | On-screen demo/code | Screen-record snippet | The image | A code/stat slide | Code blocks + sources | Code + data | One stat line |
| **Offer** | One-word comment + subscribe | Comment-keyword DM | Caption CTA | Final slide CTA | Italic CTA | Inline + end CTA | First-comment link |

05Format rules you can check

Most template rules are numbers or patterns, which makes them testable. These are the ones a script can check; the rest (voice, honesty, whether the idea is worth posting) still need a person.

- **Banned phrases** (never use): *game-changer, leverage, dive deep, in today's fast-paced
  world, it's no secret, unleash, revolutionize, cutting-edge, synergy, supercharge,
  unlock the power.*

06The files, and how to run the routine

Everything is in chapter_06_AI_Social_Media_Content_Creation. The top-level README's requirements for this chapter: nothing but a Markdown editor and any LLM.

[paste your filled 00_Hook_Story_Offer_Planning.md]

[paste 07_LinkedIn_Post_Template.md, or another platform template]

Write the LinkedIn piece using this plan and these template rules.
Follow the format exactly. No banned phrases.

## Final gate (before you open any platform template)

- [ ] Idea fits in one sentence
- [ ] Hook stops the scroll in 3 seconds (and it's not a stat)
- [ ] Exactly one core idea
- [ ] At least one real proof point (number / code / demo)
- [ ] One screenshot-worthy line
- [ ] One honest caveat
- [ ] Exactly one primary offer
- [ ] No banned phrases

Pass all eight → go to the platform templates.

07What to watch

• There is no expected output in the repo. The chapter ships templates, not generated posts. Judge a model's draft by the checklists, never by "it reads well".
• Numbers must be real. The README: never fabricate a statistic; if a claim needs a source you do not have, cut it. The template examples ("41s to 12s", "80%") show the shape of a hook or proof, not facts to reuse.
• A model will still use banned phrases. Saying "no banned phrases" lowers the rate; it does not make it zero. A ten-line check like the demo's catches what slips through.
• Run the sanity check before shipping. Does it sound like a senior colleague over chai? Would a serious professional nod or wince? Will the audience use it, or is it clickbait? All three must pass.
• Match the offer to the platform. LinkedIn hides body links (the link goes in the first comment), Instagram uses comment-keyword DMs, YouTube asks for a one-word comment and a subscribe.
• Keep one keyword per set. A Reel, a Carousel and a Post that go out together reuse the same new comment keyword.
• The brand-voice file is elsewhere. The Testing Academy voice guide used by the n8n content workflow lives in chapter 4 (skillfile_content_generation/brand-voice.md), not in this folder.

DDrills for the chapter

Drills 1 to 6 use the templates in chapter_06_AI_Social_Media_Content_Creation. Drills 7 to 9 automate the planner on this page: every control has a data-testid (turn on Show locator badges to see them).

1. Name the hook pattern

   Match each hook to its pattern: (a) "I automated everything. Then I deleted half of it." (b) "AI assistants get dumber after a few hours. You've noticed too." (c) "Your 5,000 tests are not an asset. They're a liability."

   Expected result

   (a) The reversal, (b) The reader's own thought, (c) Blunt declarative. They are the planning sheet's own examples.
2. Lint a LinkedIn draft by hand

   A draft opens "In today's fast-paced world...", lists its points with "-" bullets, has a link in the body, ends "What do you think?" and carries 12 hashtags. List every rule it breaks.

   Expected result

   Five: a banned phrase; bullets instead of → arrows; a link in the body (it goes in the first comment); the PS must be a specific or binary question; hashtags must be 5 to 7.
3. Place the proof

   Your plan's proof is "Reduced runtime from 41s to 12s". Where does it land on a Carousel and on LinkedIn?

   Hint

   The Proof row of the repurpose map.
   Expected result

   Carousel: a code or stat slide (slide 6, "CODE / PROOF", in the carousel's fill-in template). LinkedIn: one stat line.
4. Write the blog metadata

   Write a title tag and a meta description for the keyword "flaky test quarantine".

   Expected result

   A title tag of 60 characters or fewer with the keyword near the front, and a meta description of 155 or fewer with a promise and a reason to click. The keyword must also be in the H1, the URL slug, the first 100 words and at least one H2.
5. Ship a set

   A Reel, a Carousel and a Post go out together. What must match across them?

   Expected result

   The comment keyword: the Reel checklist and the Post caption rules both say to reuse the same new keyword across the set.
6. Fix a Medium section

   A Medium draft argues its point in five bullets and has no section on limitations. Fix it.

   Expected result

   Rewrite the bullets as paragraphs (bullets only for tutorial step headers: "Bullets are where arguments go to die") and add ## The Honest Caveats with 3 to 5 limitations, at least one against your own interest.
7. Pass the gate, then the checksPlaywright

   Write a Playwright test: assert the sample plan's gate score, tick the one-idea box, generate the LinkedIn draft and assert the checks.

   Hint

   Locators: ct-gate-score, ct-gate-one-idea (data-state), ct-one-idea (a checkbox), ct-generate, ct-draft, ct-lint-score.
   Expected result

   The gate goes from 7 / 8 to 8 / 8; the draft starts with "Your 5,000 tests are not an asset." on its own line; the checks show 8 / 8.
8. Open with a statisticPlaywright

   Pick the hook pattern "The number" and assert what the gate says.

   Hint

   ct-pattern value number fills ct-hook; the row is ct-gate-hook.
   Expected result

   The hook becomes "We cut flaky reruns 80% by changing one line.", ct-gate-hook has data-state="fail" and the gate drops to 6 / 8 (the one-idea box is still unticked).
9. Break the voice rulesPlaywright

   Generate the LinkedIn draft, append the line "- This is a game-changer." and assert the checks.

   Hint

   Append with fill(await draft.inputValue() + ...). Rows: ct-lint-banned, ct-lint-arrows.
   Expected result

   Both rows fail (the banned row names game-changer) and ct-lint-score drops to 6 / 8.

SSolutions: the planner test and the template rules

The Playwright spec for the planner, then the template blocks the demo's rules come from, verbatim from the repo.

import { test, expect } from '@playwright/test';

const URL = 'https://app.thetestingacademy.com/ai/blueprint/learn/content-agents.html';

test('the sample plan passes the gate and its LinkedIn draft passes every check', async ({ page }) => {
  await page.goto(URL);
  await expect(page.getByTestId('ct-gate-score')).toHaveText('7 / 8');
  await expect(page.getByTestId('ct-gate-one-idea')).toHaveAttribute('data-state', 'fail');

  await page.getByTestId('ct-one-idea').check();
  await expect(page.getByTestId('ct-gate-score')).toHaveText('8 / 8');

  await page.getByTestId('ct-generate').click();
  await expect(page.getByTestId('ct-draft')).toHaveValue(/^Your 5,000 tests are not an asset\.\n/);
  await expect(page.getByTestId('ct-lint-score')).toHaveText('8 / 8');
});

test('a statistic as the hook fails the final gate', async ({ page }) => {
  await page.goto(URL);
  await page.getByTestId('ct-pattern').selectOption('number');
  await expect(page.getByTestId('ct-hook')).toHaveValue('We cut flaky reruns 80% by changing one line.');
  await expect(page.getByTestId('ct-gate-hook')).toHaveAttribute('data-state', 'fail');
  await expect(page.getByTestId('ct-gate-score')).toHaveText('6 / 8');
});

test('a banned phrase and a bullet fail the LinkedIn checks', async ({ page }) => {
  await page.goto(URL);
  await page.getByTestId('ct-generate').click();
  const draft = page.getByTestId('ct-draft');
  await draft.fill((await draft.inputValue()) + '\n- This is a game-changer.');

  await expect(page.getByTestId('ct-lint-banned')).toHaveAttribute('data-state', 'fail');
  await expect(page.getByTestId('ct-lint-banned')).toContainText('game-changer');
  await expect(page.getByTestId('ct-lint-arrows')).toHaveAttribute('data-state', 'fail');
  await expect(page.getByTestId('ct-lint-score')).toHaveText('6 / 8');
});

## The one-line idea

> Write your idea as a single sentence. If you can't, it's not ready.

```
IDEA: ________________________________________________
WHO IT'S FOR: ________________________________________
WHY NOW: _____________________________________________  (what makes it timely / urgent)
```

```
PRIMARY OFFER (exactly one): _________________________
CTA WORDING (specific, not generic): _________________
LEAD MAGNET (if any): ________________________________
```

## The repurpose map

Once Hook · Story · Offer is locked, here's where each piece lands:

| Plan element | YouTube | IG Reel | IG Post | Carousel | Medium | Blog | LinkedIn |
|---|---|---|---|---|---|---|---|
| **Hook** | 0–15s hook | Slide-1 text + first spoken line | First caption line | Slide 1 | Title + bold subtitle | H1 + intro | First 2 lines |
| **Story** | Body beats | 2–4 quick beats | Caption body | Slides 2…N-1 | ## sections | ## sections | → arrow lines |
| **Proof** | On-screen demo/code | Screen-record snippet | The image | A code/stat slide | Code blocks + sources | Code + data | One stat line |
| **Offer** | One-word comment + subscribe | Comment-keyword DM | Caption CTA | Final slide CTA | Italic CTA | Inline + end CTA | First-comment link |

## Final gate (before you open any platform template)

- [ ] Idea fits in one sentence
- [ ] Hook stops the scroll in 3 seconds (and it's not a stat)
- [ ] Exactly one core idea
- [ ] At least one real proof point (number / code / demo)
- [ ] One screenshot-worthy line
- [ ] One honest caveat
- [ ] Exactly one primary offer
- [ ] No banned phrases

Pass all eight → go to the platform templates.

### Hook patterns
- "It is happening in [field] right now."
- "Unpopular opinion."
- A question they already think ("AI assistants get dumber after a few hours. Noticed?")
- An interview/story cold-open ("A candidate came to my interview yesterday…")
- A blunt reversal ("Your 5,000 tests are not an asset. They're a liability.")

## How to actually publish it (this is half the result)

1. **No links in the post body.** LinkedIn suppresses reach on posts with outbound body links.

## Pre-publish checklist
- [ ] Image stops the scroll and states the idea in 3 seconds
- [ ] Headline on image ≤ 8 words, readable at 30% zoom
- [ ] Caption first line works as a standalone hook
- [ ] Exactly one idea, one CTA
- [ ] Screenshot-worthy line present
- [ ] 5–10 relevant hashtags
- [ ] Handle/watermark on image, bottom-right
- [ ] No banned phrases, no fabricated numbers

## Pre-publish checklist
- [ ] On-screen text hook readable on mute in frame 1
- [ ] Spoken hook lands inside 3 seconds
- [ ] Exactly one micro-idea, 2–4 beats
- [ ] Subtitles burned in / enabled
- [ ] Payoff line is quotable
- [ ] One comment-keyword CTA (matches the [Carousel](04_Instagram_Carousel_Template.md)/[Post](03_Instagram_Post_Template.md) keyword if part of a set)
- [ ] 9:16, subtitles on, cover frame readable
- [ ] No banned phrases, no fabricated numbers

## Design spec (consistent set, all slides)

```
- Ratio: 4:5 (1080×1350) per slide
- Background: deep navy #0a0a1a + subtle grid texture
- Colors: gold/amber #f59e0b headers · red #ef4444 problems/warnings ·
          green #22c55e solutions · white body
- Fonts: monospace (JetBrains Mono) for code/numbers · bold sans (Inter) for labels
- Flat vector, no 3D, readable at 30% zoom
- Handle/watermark: small, muted gray, bottom-right
- Tooling: Ideogram / DAL·E for text-heavy slides, or build in Figma/Canva for perfect text.
  If a generator garbles text after 2–3 tries, switch to Figma/Canva.
```

## Pre-publish checklist
- [ ] Slide 1 hook stops the scroll on its own
- [ ] One idea per slide, 7–9 slides
- [ ] At least one proof slide (code snippet or real number)
- [ ] Final slide = payoff thesis + comment-keyword CTA
- [ ] NEW, unused 1-word ManyChat keyword + auto-DM copy written
- [ ] Consistent design across all slides, watermark present
- [ ] Readable at 30% zoom
- [ ] No banned phrases, no fabricated numbers

## Pre-publish checklist
- [ ] Title 8–15 words, specific (number or transformation)
- [ ] Bold subtitle contains full thesis + 3–5 real numbers + the promise, stands alone
- [ ] Opens with a moment/contrarian line, not a stat
- [ ] Paragraphs, not bullets (except tutorial step-headers)
- [ ] Every statistic has a source; nothing fabricated
- [ ] "## The Honest Caveats" present, with at least one self-critical point
- [ ] "## The Bottom Line" present
- [ ] Italic CTA at the very end + tags line
- [ ] No banned phrases

## On-page SEO checklist
- [ ] Primary keyword in: title tag, H1, URL slug, first 100 words, ≥1 H2
- [ ] Title tag ≤ 60 chars, meta description ≤ 155 chars with a clear promise
- [ ] One H1 only; logical H2/H3 hierarchy
- [ ] At least one image with descriptive `alt` text
- [ ] Internal links (2–4) + outbound links to credible sources (1–2)
- [ ] Key Takeaways block (snippet bait) + FAQ (long-tail + schema)
- [ ] Readable: short paragraphs, real examples, no keyword stuffing
- [ ] Mobile-readable, fast-loading images
- [ ] No banned phrases, no fabricated numbers