agentstackhqdotcom

# SEO Article Factory User Guide v2.3 (Final)
This guide explains how to run the 6-step prompt pack and what inputs to provide to avoid generic AI output.

Language note:
- Prompts and outputs are English-only by design.

---

## What you get (in one run)
You run Steps 0 to 6 to produce:
- One publish-ready SEO article (Markdown or HTML)
- Citations and a sourcing plan embedded in the writing
- Image placeholders with filenames, alt text, and copy-paste prompts for an external image generator
- A cabinet-style SEO audit report with scores and prioritized fixes
- A final improved article after applying the audit

---

## The workflow in plain terms
Step 0: Load the rules and workflow map. No questions are allowed here.
Step 1: You answer a structured questionnaire. The model outputs a run_config JSON.
Step 2: The model proposes 2 to 3 SEO strategy options. You pick one.
Step 3: The model does web research, then returns a final outline and image slot plan in JSON.
Step 4: The model writes the full article in plain text.
Step 5: The model audits the article and produces a clear SEO report (plain text), customized to your audit profile.
Step 6: The model applies improvements and outputs the final article plus a changes log.

---

## Why Step 1 looks long (and how to keep it fast)
Step 1 is long because it prevents three common failures:
1) Wrong intent: the article answers a different question than your audience.
2) Generic drift: the article becomes a bland overview.
3) SEO spoofing: the output looks "over-optimized" and risky.

If you want speed, you can leave most fields blank.
Defaults are applied and logged.

---

## The two big levers that reduce generic output
### 1) Angle controls (Step 1, section M)
Why angle matters:
- Without an angle, AI writing drifts into a generic overview.
- An angle forces a thesis and a consistent "spine" through the article.

What you provide:
- An angle type (recommended, choose, or custom)
- A thesis sentence (what you want the reader to conclude)
- 3 sub-questions (what the article must answer)

Examples you can copy and tweak

A) Timeline explainer (great for news and updates)
- Thesis: "In the last 30 days, X shifted because Y, and that changes Z."
- Sub-questions:
  - What happened, in what order?
  - What changed in incentives, constraints, or expectations?
  - What should readers watch next, and why?

B) Myth-busting (great when SERP is full of lazy narratives)
- Thesis: "The popular narrative misses X, and that leads readers to do Y wrong."
- Sub-questions:
  - What is the myth and why is it appealing?
  - What is verifiably true, with citations?
  - What practical conclusion follows (what should the reader do differently)?

C) Stakeholder lens (great for B2B, products, and decisions)
- Thesis: "For [role], the real outcome is X, not the obvious Y."
- Sub-questions:
  - What does this role care about most?
  - What tradeoffs exist and what are the failure modes?
  - What is the simplest decision rule a reader can use?

D) Policy or feature comparison (great for competitive SERPs)
- Thesis: "Option A vs B differs most on X, and that is where decisions are won or lost."
- Sub-questions:
  - What is the comparison frame and why it matters?
  - What are the differences that change outcomes?
  - What should readers choose in scenarios 1, 2, and 3?

Tip:
- If you want a sharp article, do not skip the thesis and sub-questions.
- If you are unsure, choose ANGLE_MODE = recommended and let the system pick one.

---

### 2) Audit personalization (Step 1, section N)
Step 5 is not a generic audit. You can tell it what to care about most.

AUDIT_FOCUS_PROFILE (pick up to 2)
- balanced: default
- rank-ready: on-page SEO and intent fit
- snippet hunter: featured snippet and PAA readiness
- trust-first: citations, claim integrity, trust cues
- conversion: CTA, persuasion, objections
- google-safe: spam risk and scaled-content risk minimization
- ai-readable: machine readability, extractable structure

AUDIT_DEPTH
- standard: fast, still useful
- deep: longer report, more evidence and edge cases

RISK_TOLERANCE
- low: conservative edits, avoids anything that looks over-optimized
- medium: normal SEO tuning
- high: aggressive but still Google-safe

Example combos
- News explainer: snippet hunter + google-safe (depth: deep, risk: low)
- Product or service article: conversion + trust-first (depth: standard, risk: medium)
- Technical guide: rank-ready + ai-readable (depth: deep, risk: medium)

---

## Freshness and news handling
If CONTENT_TIMELINESS = news-driven:
- Provide FRESHNESS_WINDOW_DAYS (for example 30)
- Provide NEWS_EVENT_DATE when possible (a date or range)
- The article will include "As of as_of_date" near the top
- Claims about events, quotes, and numbers must be cited

If you do not want recency:
- Set IGNORE_FRESHNESS = yes

---

## Links and citations (how to think about them)
EXTERNAL_LINK_POLICY
- use_only_provided: strict compliance mode, only cite your URLs
- provided_plus_research: the model can add high-quality sources if needed

Minimum external citations
- 3 is workable for evergreen how-to pages
- 5 is safer for news-driven or claim-heavy pages
- For controversial topics, set a higher minimum and use trust-first audit focus

Allowed vs disallowed sources
- Use this to prevent low-trust citations, affiliate spam, or thin blogs.

---

## Images: what the system does and does not do
It does:
- Insert image placeholders
- Suggest filenames, sizes, alt text
- Provide a copy-paste prompt you can use in an external image generator

It does not:
- Generate images
- Fetch images
- Guarantee image rights

Good default styles:
- diagrams, timelines, maps, abstract editorial illustrations
Less safe styles by default:
- photorealistic faces of real people, campaign posters, branded assets

---

## Minimal "fast run" template for Step 1
If you are in a rush, fill only:
- PRIMARY_QUERY
- BUSINESS_GOAL
- LENGTH
- FORMAT
- CONTENT_TIMELINESS
- SOURCING_MODE
- EXTERNAL_LINK_POLICY + minimum citations
- ANGLE thesis + 3 sub-questions
- AUDIT_FOCUS_PROFILE

Everything else can be left blank and defaults will be logged.

---

## What to do if the model goes off-script
If the model asks for constraints in Step 0:
- It violated the bootstrap rule.
- Start a new chat and paste Prompt 0 again.

If the model outputs JSON in Steps 4-6:
- Remind it: "Plain text only in Steps 4-6."

If the model ignores the angle and becomes generic:
- Tighten the thesis sentence and make sub-questions more specific.
- Increase Technicality from 1 to 2 if the audience is advanced.
- Add 2 to 3 "MUST cover" bullets in Targeting.

---


---

## Customization Playbook (where to put very specific requests)
Use this map when you need something unusually specific.

A) Non-negotiable constraints (scope, forbidden claims, compliance)
- Put these in Step 1:
  - Niche boundaries (MUST cover / MUST NOT cover)
  - Claims that must be sourced / must be avoided
  - Disallowed sources (domains or types)
  - Profanity allowed? (yes/no)
  - Forbidden visuals/topics (images)

B) Site-specific requirements (internal links, brand CTAs)
- Put these in Step 1:
  - INTERNAL_URLS (exact URLs + note for each)
  - Conclusion CTA text + CTA URL
- Tip: include anchor guidance in your notes, like:
  "Use varied anchors, avoid exact-match repetition."

C) Make the article less generic (high leverage)
- Put these in Step 1:
  - ANGLE thesis (one sentence)
  - 3 angle sub-questions
  - Add 2-5 MUST cover bullets under Niche boundaries

D) Want a specific structure (sections, blocks, patterns)
- Put these in Step 1:
  - LENGTH, FORMAT, FAQ block, TOC
- Then enforce in Step 3:
  - Adjust the outline_pack (H2/H3) to your exact structure before writing

E) Want a specific writing voice beyond the personas
- Put this in Step 1:
  - Writer persona = USER_DEFINED
  - Paste your custom writer spec (bulleted rules)

F) Want Step 5 to audit a specific outcome
- Put this in Step 1:
  - AUDIT_FOCUS_PROFILE (pick up to 2)
  - AUDIT_DEPTH and RISK_TOLERANCE

G) Last-mile overrides (after Step 5)
- Put this in Step 6:
  - Any new constraints override Step 5 recommendations.
  - If you provide no overrides, Step 6 applies Step 5 by default.

---

## Advanced inputs (how to write constraints that the model can follow)
Prefer short, testable statements.

Good MUST cover bullets:
- "Explain X in 2-3 sentences, then give 1 real-world example."
- "Include a short timeline of events with dates (cited)."
- "Define key term Y in one crisp paragraph, then use it consistently."

Good MUST NOT cover bullets:
- "No speculation about motives."
- "No legal advice."
- "No partisan framing."

Good 'claims that must be avoided':
- "Avoid: 'everyone knows', 'it is obvious that', 'studies show' without citations."
- "Avoid absolute language unless sourced: 'always', 'never', 'proves'."

---

## Angle sharpening checklist (quick)
If your output feels generic, tighten these three items in Step 1:
1) Thesis: make it falsifiable (it could be wrong).
2) Sub-questions: make them specific and non-overlapping.
3) MUST cover: add 2-5 bullets that force concrete detail (dates, mechanisms, tradeoffs).


---

## Special formatting needs (where to specify)
- FORMAT (markdown or html): Step 1
- If you need WordPress-friendly blocks:
  - Put a note in Step 1 under Structure controls, for example:
    "Prefer short paragraphs, scannable H2s, and optional callout blocks."
- If you need schema-friendly sections (FAQ, HowTo steps):
  - Enable FAQ in Step 1
  - Choose AUDIT_FOCUS_PROFILE = snippet hunter in Step 1


## Glossary
- Angle thesis: the one-sentence spine of the article.
- Information gain: what the page offers that competitors do not.
- Snippet readiness: formatting and clarity that makes a page eligible for featured snippets or PAA.
- AI readability: how easy it is for AI systems to parse and reuse content accurately.