Extract Links

Extract links (Codex variant)

Annotate every URL in a notes file with its content. By default (light) writes a one-line gist inline next to each URL — fast triage, nothing written to disk. With --full, pulls full content (YouTube subtitles, Telegram post text, HTML articles) into a shared sibling extracted/<note-basename>/ directory, replacing each URL with a local pointer (multiple notes in one directory consolidate under a single extracted/ parent).

This is the Codex CLI variant. Behaviourally identical to the Claude Code variant — only the user-interaction idiom differs (numbered list TUI prompt instead of AskUserQuestion radio UI). All shared resources (roles/, scripts/) come from the Claude variant tree via install-time symlinks.

Letter = spirit. If a rule blocks you from reaching the goal it was written for, the rule is wrong, not the goal. Don't look for a wording loophole — ask what the rule is protecting, and protect that.

Usage

/extract-links <note.md> [--force] [--full | --light]

--force re-processes URLs even if they're already annotated (default: skip already-annotated).

Light is the default: a one-line summary of each link's gist written inline next to the URL, no extracted/ tree — fast triage for URL-heavy notes. Pass --full to extract full content into extracted/ with local pointers (use it when /cleanup / /blueprint need the actual text offline). --light is accepted explicitly too.

Weaknesses and when NOT to use

• Does not work with private/auth resources. Private Telegram channels, paywalled articles, logged-in-only pages — out of scope. Those URLs return an error in the final report.
• Depends on external tools (yt-dlp, pandoc). If they're not installed, the skill prompts the user (numbered-list TUI prompt) to install. Never auto-installs without explicit OK. If the user declines — the matching URLs get an error. (Telegram embed-scrape needs only curl, which is almost always present.)
• Overkill for 1-2 URLs. Copy-paste is faster than the pipeline. Use it only with 3+ URLs.
• Long YouTube videos (>2h, ~30k words). Extract will succeed, but downstream work (cleanup) may choke on the volume. Pre-trim manually if needed.
• JS-heavy SPA sites. extract-html.sh uses curl — JS is not executed. You'll get the page skeleton without content. Use it for blog posts, articles, docs, NOT for interactive web apps.
• Heuristic reference-detection is imperfect. Notes often contain URLs that aren't content to extract — API doc landing pages, GitHub repo roots, citation-style references, tool homepages. The skill flags them via heuristics in step 1 and surfaces a triage prompt in step 2 (default = skip). What looks like a "tool homepage" might be content the user wants (e.g. a project's blog) — the call is always handed to the user, never silently dropped.
• Light mode is best-effort metadata, not content. --light writes a one-line gist (from title / description / Telegram post-preview) next to each URL — fast, no extracted/ tree. It does NOT bring content offline; for JS-SPA / paywalled / private URLs the metadata is thin or missing (the summary must say so, never fabricate). Use --full (default) when downstream /cleanup / /blueprint need the actual content.
• Non-interactive mode (codex exec). This skill needs user input for the dependency-install prompt (step 3) and reference triage (step 2). If invoked from codex exec without TTY, fail with an explicit error rather than silently auto-installing or auto-skipping. User must run from codex TUI.

How to do it wrong vs right

Dependency management

❌ Wrong: yt-dlp missing → automatically pip install --user yt-dlp without asking.

• User might not want Python user-site packages.
• User might be on a shared system where pip is restricted.
• The skill turned itself into an installer.

✅ Right: Detect missing → print numbered options in TUI and wait for reply:

yt-dlp not found. How would you like to proceed?
  1. Install yt-dlp now (recommended)
  2. Skip YouTube URLs (they will be reported as errors)
  3. Abort
> _

Accept 1, 2, 3, or text alias. Re-prompt on ambiguous reply. User-explicit only.

URL annotation

❌ Wrong: Replace the URL in the note entirely: [YouTube video](./extracted/youtube-abc/subtitles.en.txt) — original URL is lost.

• Auditability: no way to tell where the content came from.
• Re-extract is impossible without the original.

✅ Right: URL preserved, local pointer appended next to it:

See https://youtube.com/watch?v=abc → [./extracted/note/youtube-abc/subtitles.en.txt](./extracted/note/youtube-abc/subtitles.en.txt)

URL noise (citations, API refs, repo roots)

❌ Wrong: Note has 12 URLs — extract all 12, including 4 that are bare GitHub repo roots, an API docs landing page, and citation [1] [2] [3] references at the end.

• 4 useless extracts pollute extracted/, each with their own slug subfolder.
• Downstream /cleanup has to navigate around noise files that contain only navigation HTML.
• The user's actual content (the other 8) gets diluted; signal-to-noise drops.

✅ Right: Classify URLs in step 1. Detected reference URLs (bare hosts, docs.*, GitHub repo roots, package-registry landings, anchor-only fragments) are surfaced together in a single triage prompt (step 2): "These N URLs look like references, not content — skip all? Extract all anyway? Pick which?" Default = skip all. Skipped URLs appear in the final report so the user can audit the decision.

Multi-URL note

❌ Wrong: Note has 10 URLs — processed 7, 3 failed — report "extract done" without mentioning the errors.

• The user doesn't know about the 3 broken URLs.
• Downstream works from a holey map.

✅ Right: Final report enumerates: 7 extracted, 3 errors (with reason per URL). The note only annotates successful ones. Errored URLs are left un-annotated — the user decides whether to retry or accept the gap.

Roles

roles/interactive-prompt.md — prompt template for other-type URLs (not YouTube, not Telegram). Substitution: {url_short} (first 60 chars of the URL). Originally written for AskUserQuestion semantics; in Codex you render it as a numbered list in the TUI and accept a numeric or text reply.

Scripts in scripts/ are building blocks; the skill calls them via Bash (no Codex-specific changes — same shell scripts as Claude variant):

Mode routing

Default is Light (the ## Light mode section below) — fast inline gists, nothing written to disk. Pass --full for the full extraction pipeline (steps 1-7 here). Both modes start from the same step-1 URL detection.

Full mode (--full) — step by step

1. Read the note, find URLs. Regex https?://[^\s)]+ (with trailing-punctuation strip). Classify each URL:

   • youtube — youtube.com/watch?v=* or youtu.be/* (specific video).
   • telegram — t.me/<channel>/<post-id> (specific public post).
   • reference — likely not content, by any of:
     • bare host (no path, or only /),
     • docs.* subdomain, or path starts with /docs/, /reference/, /api/, /library/,
     • GitHub repo root (github.com/<owner>/<repo> with no further segments — /blob/, /issues/N, /pull/N, /releases/... are content, not reference),
     • package registry landings (npmjs.com/package/*, pypi.org/project/*, crates.io/crates/*, rubygems.org/gems/*),
     • anchor-only or fragment-only URLs (#section-id).
   • other — everything else (article, blog post, generic HTML).

2. Triage references with the user (if any detected). Show a numbered list in the TUI:

   N URLs look like references, not content:
     a) https://github.com/owner/repo  (GitHub repo root)
     b) https://docs.python.org/        (docs landing)
     ...
   How to handle?
     1. Skip all references (recommended — go into final report as skipped(reference))
     2. Extract all anyway (re-classify each to `other` and run interactive prompt)
     3. Pick which to extract (1-at-a-time prompts per URL)
   > _
   

   Default = 1 if user just hits enter. Accept 1 / 2 / 3 or text alias. For option 3, iterate per-URL prompts.

   If there are no reference URLs — skip this step silently.

3. Probe dependencies. bash scripts/install-deps.sh. If a tool required for the URL types is missing — TUI numbered prompt (install / skip / abort). NEVER auto-install without explicit OK.

4. For each URL — extract. Output root is <note-dir>/extracted/<note-basename>/ — one shared extracted/ per directory, per-note subfolder inside. For multi-note runs in the same directory, the parent is consolidated automatically.

   • YouTube → bash scripts/extract-youtube.sh <url> <note-dir>/extracted/<note-basename>/<slug>/
   • Telegram → bash scripts/extract-telegram.sh <url> <note-dir>/extracted/<note-basename>/<slug>/
   • Other → render roles/interactive-prompt.md as numbered TUI prompt; accept reply. By choice: readable HTML / skip / custom command.
   • Slug: <type>-<short-id> (e.g. youtube-dQw4w9WgXcQ, telegram-channel-123, html-blog-example-com). Max 50 chars.
   • Errors (404, private, fetch fail) — log to <note-dir>/extracted/<note-basename>/.errors.log, do not annotate the URL, keep going.

5. Annotate the note. For each SUCCESSFULLY extracted URL — append → [<local-path>](<local-path>) right after the URL in the note. <local-path> is relative to the note: ./extracted/<note-basename>/<slug>/.... Original URL preserved. Use the Edit tool, not Write. If a URL is already annotated (→ [./...] directly after) — skip unless --force. Triage-skipped URLs are NOT annotated.

6. Update .gitignore. Add extracted/ to .gitignore at git root (if git is initialized). Idempotent — only append if the line isn't already there:

   grep -qxF 'extracted/' .gitignore 2>/dev/null || echo 'extracted/' >> .gitignore
   

7. Final report + commit. One line per URL with state extracted / error / skipped(reference) / skipped(user) plus aggregate metrics. Auto-commit extract-links: <N> URLs from <note> (only processed files, not the whole branch).

Light mode (default)

The default (or explicit --light). Same URL detection as Full step 1, then for every URL (reference-triage is relaxed — a one-liner on a repo root or docs page is cheap and useful, so summarise all, don't skip references):

1. Probe deps (light). Only curl (almost always present) is required; yt-dlp is needed only if YouTube URLs exist (pandoc is NOT — light mode never converts full HTML). Same TUI install gate as Full step 3, but skip it when there are no YouTube URLs.
2. Fetch metadata. bash scripts/summarize-url.sh <url> — prints labelled metadata (TITLE / UPLOADER / DURATION / TEXT / DESC) or an ERROR: <reason> line. No full-content download, nothing written to disk.
3. Condense to one line. From that metadata, write a single plain-language sentence — what the link is and why it's likely here — aim for ≤ ~140 chars. If the script returned ERROR: (private / JS-SPA / fetch failed) or only thin metadata, say so honestly (e.g. _(Telegram post — preview unavailable)_). Never invent content you didn't fetch.
4. Annotate inline. Append → _<summary>_ directly after the URL (italic, no link). Original URL preserved. Idempotent: skip URLs already annotated with → _…_ unless --force.
5. Report + commit. One line per URL: summarised / error (reason) + aggregate. Commit extract-links: summarise <N> URLs from <note>. No extracted/ tree, no .gitignore change.

Light mode is for orientation/triage of URL-heavy notes (e.g. a backlog). It does NOT bring content offline — switch to Full mode when /cleanup / /blueprint need the actual text.

Light mode — wrong vs right

❌ Wrong: summarize-url.sh returns ERROR: telegram post not accessible; you write a plausible-sounding summary from the channel name anyway.

• That's fabricating content you never fetched — the whole point of a summary is that it's grounded.

✅ Right: Write → _(Telegram post — preview unavailable: private or deleted)_ and count it as error in the report. The user decides whether to open it manually.

Outputs

Per processed note (anchored at <note-dir>/extracted/<note-basename>/):

• <slug>/ — extracted content (per URL):
  • YouTube: subtitles.en.txt, subtitles.ru.txt (if available), metadata.json
  • Telegram: post.md, media/urls.txt + optional media files
  • HTML: content.md, metadata.json
• .errors.log — per-URL errors (if any)
• <note> — modified: a local pointer link is appended next to each URL

Multi-note layout in one directory (single shared parent):

<dir>/note-1.md
<dir>/note-2.md
<dir>/extracted/
  note-1/
    youtube-abc/...
    telegram-xyz/...
  note-2/
    html-blog-example/...

Git:

• .gitignore — extracted/ added
• Commit: extract-links: <N> URLs from <note> (encompasses note edit + .gitignore; the contents of extracted/ are gitignored)

Light mode (--light): no extracted/ tree and no .gitignore change — only the note is modified, with → _<summary>_ appended next to each URL. Commit: extract-links: summarise <N> URLs from <note>.

Connections to other skills

• Input: any markdown file with URLs. Usually a note or plan before /cleanup or manual review.
• Output: annotated note with local pointers. Suitable for /cleanup, manual analysis, plain reading.
• Does not call other skills automatically. After step 7: Extracted N URLs from <note>. Run /cleanup next if needed. — a soft hint for the typical pipeline, not a forced chain.

Rules

Commonality

The note is a shared artifact. After extract, it'll be read by the user, downstream skills, and future sessions. If you skipped URLs or lost error info, everything downstream works from a holey map. Not "processed most of them" — that's "missed some", and it has to be reported explicitly.

Prior commitment

In step 1 you committed to classifying every URL (not just youtube/telegram/other — also reference). In step 2 you committed to surfacing every reference URL to the user before deciding its fate. In step 4 you committed to processing EVERY remaining URL — either extracted, or explicitly errored with reason. In step 5 you committed to preserving the original URL and appending the pointer (not replacing). In step 6 — idempotent gitignore_add. Skipping any step withdraws the basis for trusting the final report.

Authority

The skill exists precisely because extracting a dozen URLs by hand is slow and error-prone. If you silently skip "obviously irrelevant" URLs without prompting the user, you are judging content you haven't seen — that's the user's role, not yours. The reference heuristic in step 1 is a PROPOSAL, not a verdict; the triage prompt in step 2 is where the actual skip decision lands.

Self-check before delivering the result

Would this result pass review by a senior engineer? Concretely:

• Were all URLs from the note processed — extracted, errored (with reason), or skipped via explicit user decision (triage)?
• Were reference-classed URLs surfaced via numbered TUI prompt in step 2 — not silently dropped?
• Is the original note annotated correctly — original URL preserved, pointer appended only for extracted ones, triage-skipped URLs left bare?
• Is the extracted/<note-basename>/ layout consistent (slug naming, metadata.json per extract; one shared extracted/ parent per directory)?
• Is .gitignore updated idempotently — no duplicate lines?
• Does the commit message reflect actual work — extract-links: N URLs from <note>?
• Does the final report enumerate every URL with state (extracted / error / skipped(reference) / skipped(user))?
• No leaked secrets in extracted/ (auth tokens from API responses, personal data)?
• (Light mode) Every URL got either a one-line summary or an explicit error; thin/failed fetches are flagged honestly (not fabricated); no extracted/ tree or .gitignore change was made?

If "no" on any item — redo, don't ship.