Objective

Turn the attached transcript into one markdown file readable on mobile: a guide that beats the source, a table of every reusable prompt in it (if any), and a catalog of every resource it names.

Input gate

The transcript must be an attached file. Read it from /mnt/user-data/uploads/ and report its line count as your fattached as af output. If nothing is there, stop and say so — do not proceed from project files or memory. If the transcript was pasted into the message instead of attached, write it to a file first, then continue.

Skills

Primary: transcript-superguide. Follow its Phase 0 classifier and its own composition rule — route gap-fill to transcript-deep-dive only if Phase 0 returns a shallow solo technical tutorial. Any other classification: fill gaps inside the superguide and do not invoke deep-dive. Never run both over the same material.

Seed the inventory with python3 /mnt/skills/user/transcript-superguide/scripts/extract_resources.py <transcript-path>. It's stdlib-only and needs the transcript on disk. If it fails, extract by hand and say so in one line.

Size gates — obey these, they're what makes this finish
Catalog at most [N, default 25] resources, chosen by usefulness to the reader, not order of mention. Everything beyond that goes in a names-only overflow list at the bottom.
At most one web lookup per catalogued resource, and at most [M, default 20] lookups total for the whole run. When the budget is gone, stop searching and mark the rest [VERIFY: what to check].
No second research sweep. No "let me also confirm."
Catalog schema — overrides the skill's built-in seven-column table

Two tables, because a nine-column table is unreadable on a phone.

Scan table (read on the phone): | Name | Category | Have/Installed?* |

Detail table (read on a laptop or in the file viewer): | Name | Category | Source/Vendor URL | Repo/Archive URL | Description | Best Use Case | Currently Have/Installed* | Price/Cost | Published Date |

Column rules:

Source/Vendor URL — the official page for the thing. If it's open-source/free/no vendor, mark N/A (OSS) or N/A (free).
Repo/Archive URL — canonical repository, source, or original location, else N/A.
Description — one sentence on what it is, not what the speaker thought of it.
Best Use Case — the situation where it beats the alternative. Not a restatement of Description.
Price/Cost — current price or terms, or Free. Never "varies."
Published Date — initial public release/publication. If only a latest-version date is findable, give it and label it latest version.

*Have/Installed column — conditional. Include this column pair only if the topic is software, tools, skills, or connectors this session could plausibly check. For non-digital topics (recipes, techniques, physical products, places, people, books), drop the column pair entirely rather than forcing an artificial "installed" status onto it.

When it applies, this session cannot see your local devices or accounts. Do not try, and do not write "unknown" for everything either. Answer from what this session can actually observe, with exactly these four values:

Installed (skill) — appears in this session's own available-skills list.
Connected (connector) — appears in this session's connected MCP/connector list.
Not detected — neither, and this session can't check your machine.
N/A — not the kind of thing that gets installed (a website, an article, a technique, a physical resource).

Never infer installation from a name sounding familiar. A wrong "Installed" here is the worst single output this prompt can produce.

Prompt-library table (skip entirely if the transcript has no reusable prompts)

Prompts the transcript demonstrates, quotes, or describes well enough to reconstruct: | Prompt Name | Purpose | Origin | Prompt Text | Reuse Notes |

Origin is one of verbatim, reconstructed, or written by me. Reconstructed and written prompts must be labeled in the table itself, not only in surrounding prose. If the transcript contains no reusable prompts, say that in one line instead of manufacturing rows.

Output

One markdown file, written to /mnt/user-data/outputs/ and presented as a file so it's reachable on mobile. Name it [topic]-superguide_MMDDYY.md (append your agent code per ClearPath if the surface you're running from has one). If a same-purpose file exists, write a new dated one and say whether yours supersedes or extends it. Never overwrite.

Section order: classification and a one-paragraph verdict on what the source gets right and wrong · scan table · the superguide body (gaps closed inline, no visible move tags) · prompt-library table (if applicable) · detail catalog table · overflow list · a short endnote for anything still carrying [VERIFY].

Keep the chat reply under four sentences. The file is the deliverable; the reply is a pointer to it.

What wrong looks like — each of these satisfies every instruction above and still fails
A faithful summary. The output beats the source or it wasn't worth the run.
A URL, price, or date filled from memory and left untagged. A confident wrong price is worse than [VERIFY].
"Installed (skill)" written for something not in this session's actual skill list — or an Installed column forced onto a non-digital topic where it doesn't apply.
A reconstructed prompt presented as verbatim.
Rows padded with things the transcript merely gestures at, to make the table look complete.
Both skills run end to end over the same material, doubling length and adding no executable step.
A visible [CORRECT] / [FILL] / etc. tag, or a prominent "Honest Gaps" section, anywhere in the delivered file — gaps get closed during research and folded into plain prose; only a genuinely unresolvable item earns the short closing endnote.
The search budget blown past because a few cells were still [VERIFY]. The budget is the instruction; the gaps are the expected output.
A file written but not presented. On mobile that file is unreachable.
Ending the chat reply on what else could be researched, instead of on the file.
Done means

The file is written and presented. Every volatile cell is sourced-and-dated or [VERIFY: ...]. Every catalogued resource has a correct Have/Installed value, or the column is absent because the topic doesn't call for it. Resources beyond the cap appear in the overflow list. The chat reply is four sentences or fewer.