TheMattBerman/landing-page-factory

Landing Page Factory Orchestrator

This is the control layer.

Use this skill when the request is about:

• building a full landing page pipeline
• creating one or more page variants
• rerunning selected stages after edits
• figuring out the right order of operations
• admin tasks around page package naming, artifact checks, cleanup, or memory logging

Do not replace the stage skills. This skill routes and supervises them.

What this skill owns

• request classification
• stage ordering
• artifact preflight checks
• hard-stop and downgrade enforcement
• page naming and variant naming
• deciding which stages must rerun after a change
• memory logging in memory/YYYY-MM-DD.md

What this skill does not own

• extraction logic
• strategy logic
• copywriting logic
• image-generation logic
• HTML build logic
• QA scoring itself

Those belong to:

• site-extract
• page-strategy
• brand-profile
• page-copy
• page-visuals
• page-build
• page-qa

Request classification

Route every request into one of these modes first.

1. Full pipeline

Examples:

• "Build me a landing page for https://example.com"
• "Run the whole thing for this brand"

Run in order:

1. site-extract
2. page-strategy
3. brand-profile
4. page-copy
5. page-visuals
6. page-build
7. page-qa

2. Variant pipeline

Examples:

• "Make a version for enterprise buyers"
• "Build three angle variants"

Use the existing brand artifacts if available. Create a new page package per variant and rerun:

1. page-strategy
2. page-copy
3. page-visuals
4. page-build
5. page-qa

Rerun site-extract or brand-profile only if source facts changed or the current brand artifacts are missing or weak.

3. Stage rerun

Examples:

• "Rebuild visuals only"
• "Run QA again"
• "Update copy after strategy changes"

Only rerun the minimum necessary downstream stages.

Dependency rules:

• strategy changed → rerun copy, visuals, build, qa
• brand profile changed → rerun copy, visuals, build, qa
• copy changed → rerun build, qa
• visuals changed → rerun build, qa
• build changed → rerun qa

4. Admin / inspection

Examples:

• "What artifacts are missing?"
• "What order should we run this in?"
• "Create a new page package"

Inspect, report, and prepare. Do not fabricate missing artifacts.

Preflight checks

Before invoking downstream work, verify required artifacts.

Brand-level:

• workspace/brand/extract.md
• workspace/brand/extract.json
• workspace/brand/profile.md when copy, visuals, or build need it
• workspace/brand/palette.json when visuals or build need it

Page-level:

• workspace/pages/[page-name]/strategy.json
• workspace/pages/[page-name]/copy.md
• workspace/pages/[page-name]/visuals/
• workspace/pages/[page-name]/index.html
• workspace/pages/[page-name]/meta.json
• workspace/pages/[page-name]/qa.md

If an artifact required for the requested stage is missing, stop and route to the earliest missing stage.

Hard stops

Never continue past these:

• no strategy.json for copy, visuals, or build
• no unique mechanism identified
• no CTA destination for lead-gen or action-driven pages
• proof too thin for the proposed certainty level
• Firecrawl expected for extraction but not configured or runtime-blocked
• Bloom expected for visuals but not configured or runtime-blocked, unless the user explicitly approves Nano Banana or manual assets

When blocked, state:

1. what is missing
2. which stage must run next
3. what can be downgraded versus what cannot

Downgrades

Continue with downgrade only when the stage skill allows it.

Typical downgrades:

• weak proof → softer claims, draft-only QA target
• partial visuals → neutral support layout
• exact-product image failure → branded environment or operator asset
• missing Bloom brand → onboard for provider context, but keep repo brand artifacts as source of truth

Naming rules

Use descriptive page package names under workspace/pages/.

Pattern:

• base run: [brand]-[core-angle]
• variant run: [brand]-[audience]-[angle]

Good examples:

• ridge-wallet-proof
• stealads-enterprise-time
• notion-agency-velocity

Avoid:

• test-page
• v2
• new-one

Variant workflow

When building multiple variants:

1. keep one shared brand source of truth in workspace/brand/
2. create a separate page package per variant
3. change audience, angle, offer framing, proof emphasis, or visual treatment deliberately
4. run full page-level downstream stages for each variant
5. compare QA verdicts and review flags across variants

Do not let variants silently overwrite each other.

Stage ordering rules

Use this exact logic:

• If the user provides only a URL: start at site-extract

• If brand artifacts exist but page package does not: start at page-strategy

• If strategy.json exists and the user asks for copy: run page-copy

• If copy.md and brand artifacts exist and the user asks for visuals: run page-visuals

• If strategy, copy, and visuals exist: run page-build

• If build artifacts exist: run page-qa

Build and visual script entrypoints

For Cowork and other skill-folder runtimes, prefer the skill-local scripts:

• skills/landing-page-factory-orchestrator/scripts/page-admin.py
• skills/page-visuals/scripts/image-provider.sh
• skills/page-build/scripts/resolve-visual-assets.py
• skills/page-build/scripts/prepare-build-meta.py
• skills/page-build/scripts/select-build-images.py
• skills/page-build/scripts/html-image-context.py
• skills/page-build/scripts/build-page.py

Repo-root scripts/ entrypoints are local wrappers only.

Use the admin helper for orchestration support:

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py status --page-name [page-name] Returns current artifact status and the earliest missing stage.

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py suggest-name --brand [brand] --audience [audience] --angle [angle] Returns a clean, non-colliding page package name.

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py compare-variants --page-name [page-a] --page-name [page-b] Returns a side-by-side JSON summary of artifact status, readiness score, readiness notes, headline, page type, CTA destination, image count, review flags, and QA verdict when present.

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py compare-variants --prefix [brand-prefix] --format markdown Returns a human-readable markdown comparison table for Cowork reviews, ranked by readiness.

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py compare-variants --prefix [brand-prefix] --format markdown --output workspace/pages/[report-name].md Writes the comparison report to disk.

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py compare-variants --prefix [brand-prefix] Compares all page packages that match the prefix.

• python3 skills/landing-page-factory-orchestrator/scripts/page-admin.py log-memory --title "[title]" --page-name [page-name] --stages "[stages]" --verdict "[verdict]" Appends a short run log to memory/YYYY-MM-DD.md.

Use the runner when you want the orchestrator to actually execute the scripted parts of the pipeline:

• python3 skills/landing-page-factory-orchestrator/scripts/run-pipeline.py --url https://example.com --page-name example-proof --format markdown Runs extraction if needed, then stops at the next skill-authored stage with exact handoff instructions.

• python3 skills/landing-page-factory-orchestrator/scripts/run-pipeline.py --page-name example-proof --format markdown Resumes an existing page package and continues through the build chain when the required artifacts already exist.

• python3 skills/landing-page-factory-orchestrator/scripts/run-pipeline.py --url https://example.com --brand Example --audience enterprise --angle time-saving --format markdown Creates a non-colliding variant page name and runs until the next manual stage.

The runner is intentionally hybrid:

• it executes scripted stages directly
• it stops cleanly at page-strategy, brand-profile, page-copy, and most page-visuals work when authored artifacts are still missing
• it resumes automatically once those artifacts exist

Extraction policy:

• prefer Firecrawl automatically when FIRECRAWL_API_KEY is available
• if Firecrawl is not configured, basic extraction is allowed but should be treated as a downgrade
• for public-release-quality runs, recommend adding Firecrawl rather than silently acting like the extraction quality is equivalent

Provider enforcement:

• do not allow site-extract to quietly collapse into WebFetch/manual scraping when Firecrawl is the intended path
• do not allow page-visuals to quietly collapse into default built-in image generation when Bloom is the intended path
• if the preferred provider is blocked, missing, or not configured, pause and ask the operator which downgrade is acceptable
• only continue with a downgrade when the user explicitly approves it

Cowork-specific behavior:

• if Cowork blocks api.firecrawl.dev, say so directly and recommend either allowlisting the domain or running the local extraction preflight
• if Cowork cannot reach Bloom, say so directly and ask whether to configure Bloom, use Nano Banana, or proceed with manual assets
• do not let “it kind of worked with WebFetch/default images” count as a successful provider run

Do not claim full automation for stages that still depend on skill-authored judgment.

Memory logging

After any meaningful run, append a short note to memory/YYYY-MM-DD.md with:

• brand or page name
• stages completed
• QA verdict if available
• blockers or downgrades
• notable lessons

Keep it terse. One short section per run is enough.

Output expectations

When orchestrating, always report:

• chosen mode: full pipeline, variant, stage rerun, or admin
• page package name(s)
• stages run or next stages required
• blockers, downgrades, or review flags

For variant runs, include a flat comparison:

• variant name
• audience / angle
• current artifact status
• QA verdict if present