start-the-roast

name

start-the-roast

description

Run a full funnel roast: collect a sales-funnel URL and a Meta or Facebook ads deeplink, fetch the ad, walk the live funnel with Playwright, apply the matching marketing rubric at each hop, and write the report card. Use when starting a roast, roasting a funnel, or producing a roast report from an ad plus a landing page. Do not use for an isolated Ad Library fetch, an isolated walk, Niko, or npm run score.

Start the Roast

Single-agent workflow. Disk is the handoff. Chat output is the workdir path and the report path.

flowchart LR
  gather["Collect inputs"] --> meta["get-meta-ad"]
  meta --> funnel["funnel.json"]
  funnel --> creative["ad-creative invert"]
  creative --> walk["walk-the-funnel"]
  walk --> hops["Per-hop cro"]
  hops --> mock["copywriting + image"]
  mock --> report["write-the-roast"]

The chart is the map. The steps below are the driving instructions.

Step 1 — Collect inputs

Resolve two values from the invocation:

1. Funnel — an absolute HTTP(S) URL for the landing page or offer.
2. Ads deeplink — an absolute HTTP(S) URL to Meta or Facebook ads. A library URL with a non-empty id query parameter is one ad. A URL without id is a set.

If either value is missing or is not an absolute HTTP(S) URL, ask for the missing piece and stop. Do not create a run directory. Do not fetch ads. Do not walk. Do not write a report.

Step 2 — Establish workdir

Create the run directory. It persists if later steps only retrieve part of the ads.

Default path

.rmf/<timestamp>/ in the operator's current working directory. <timestamp> is UTC at seconds precision with colons replaced by hyphens so the name is one path segment (example: 2026-05-12T18-30-45). Create .rmf/ if it does not exist.

Override

If the invocation includes --workdir <path>, use that path instead of the default. Create it if it does not exist.

Writable check

Create a small file in the workdir and delete it. If that write fails, stop. Name the path and the failure reason. Do not continue without a writable workdir.

Step 3 — Read the schema

Read references/funnel-json.md. That file is the shape of funnel.json. Do not write funnel.json until this read is done.

Step 4 — Fetch ads

Invoke ../get-meta-ad/SKILL.md with:

1. The ads deeplink from Step 1
2. The workdir from Step 2

Do not open, fetch, or parse the Ad Library yourself. That skill writes <workdir>/ads.json and any media files.

When it finishes, continue to Step 5. A missing ads.json or an empty ads array is not a failure.

Step 5 — Write funnel.json

Write <workdir>/funnel.json to the schema from Step 3.

• funnel.url is the funnel from Step 1.
• source.ads_deeplink is the deeplink from Step 1, stored as given.
• ads is the ads array from <workdir>/ads.json. Use [] when that file is missing or ads is empty.
• For each media object, include path only after confirming that file exists under the workdir. Omit path when the file is missing.

Do not invent ads, copy, or media paths.

When funnel.json is on disk, continue to Step 6.

Step 6 — Invert ad creative

Read references/hop-skills.md. Do not write any analysis/ file until this read is done.

Apply the retrieved-ad row. Write <workdir>/analysis/creative.md.

If ads is empty, write the miss. Do not invent an ad. Continue to Step 7 either way.

Do not browse. Do not call npm run score.

Step 7 — Walk the funnel

Invoke ../walk-the-funnel/SKILL.md with:

1. The funnel URL from Step 1
2. --out <workdir>
3. --run-id walk

Do not drive the browser yourself. Do not type email, address, or card data. Do not submit payment.

If the walk fails, print the error and stop. Do not invent hops. Do not write hop analysis. Do not write a report.

When it finishes, read <workdir>/walk/steps.json from disk. Do not use chat paste as the hop list.

If steps.json is missing, stop. Name the missing path.

steps.json on disk is necessary, not sufficient. For each hop, open the screenshot at the path in that file (relative to <workdir>/walk/). A hop that is not the intended page is a walk failure — bot wall, challenge, interstitial, blank chrome, leftover previous page. The URL looking right does not save it. Cloudflare is one example, not the definition.

Inspect screenshots only. Do not open bundle.json.

If any hop is not the intended page: recover or bail. Retry the walk, wait, or stop. Do not type email, address, or card data. Do not submit payment. Do not invent hops. Telling the operator that analysis cannot be completed is an acceptable end state, and a better one than judging the wrong page. Print the error and stop. Do not write hop analysis. Do not write a report.

Continue to Step 8 only when every hop is the intended page.

Step 8 — Analyze each hop

Read references/hop-skills.md if that read is not already done.

For each hop in walk/steps.json, apply the mapped row and write <workdir>/analysis/<hop>.md.

Cite workdir-relative artifact paths (walk/artifacts/…). Do not invent pages. Do not browse. Do not drive the browser. Do not call npm run score.

When every hop in steps.json has an analysis file, continue to Step 9.

Step 9 — Draft the LP mock

Apply the after-landing row from hop-skills.md. Write <workdir>/analysis/lp-mock.md. Write <workdir>/media/lp-mock.png only when an image file actually lands.

A missing image is not a failure. Continue to Step 10.

Step 10 — Write the report

Invoke ../write-the-roast/SKILL.md with the workdir from Step 2.

Do not assemble the report yourself. Do not call npm run score.

When it finishes, continue to Step 11. If <workdir>/report.html is missing, stop. Name the missing path.

Step 11 — Print the paths

Print:

1. The absolute path of the workdir
2. The absolute path of <workdir>/report.html

Those two paths are the deliverable.

Do not paste funnel.json, steps.json, or report.html into chat.