add-substack-article-to-repo

name

add-substack-article-to-repo

description

Mirror a Substack article into the In the Weeds public learning repo at `~/in-the-weeds-substack-articles/`. Handles two source modes - local (slug from `~/hannah-personal-agent/`) and firecrawl (URL from any Substack-hosted post). Copies article + images, generates the article-level CLAUDE.md, cascades index updates across root CLAUDE.md, series CLAUDE.md, welcome.md, CONTRIBUTORS.md, and LICENSE, then opens a clean PR. Use this skill whenever Hannah says "add this article to in-the-weeds", "publish [slug] to the repo", "mirror [URL] into the learning repo", "add the [slug] article", or runs /add-substack-article-to-repo. Also use when she pastes a Substack URL (https://...substack.com/p/... or any /p/ pattern) and asks to bring it into the repo.

Add Substack Article To Repo

End-to-end skill for getting a Substack article into the public learning repo /Users/hannahstulberg/in-the-weeds-substack-articles/. Covers everything from source discovery to a clean PR.

When the skill triggers

Hannah just published an In the Weeds article (or wants to add an external article by a collaborator) and needs to mirror it into the public repo so the AI learning assistant can teach from it. Two patterns:

• She says a slug like add the curiosity-is-the-only-prerequisite article to the repo — assume the source is local under ~/hannah-personal-agent/.
• She pastes a URL like add this article: https://www.pathtostaff.com/p/business-sense-for-engineers-the — assume the source is the live Substack post and use firecrawl-cli to scrape it.

Mode selection (Stage 0)

Inspect the input:

• Matches ^https?:// → firecrawl mode. Load references/source-mode-firecrawl.md. Derive slug from the last /p/<slug> segment of the URL.
• Otherwise → local mode. Load references/source-mode-local.md. Treat input as the slug.

Hannah can override: "use firecrawl on " forces firecrawl, "fetch from local " forces local.

Write /tmp/add-substack-article-<slug>.json:

{"state": "mode_selected", "mode": "local|firecrawl", "input": "<input>", "slug": "<slug>"}

Workflow (after mode selection)

The two source modes diverge at Stages 1–2 (discovery + metadata read), then re-converge from Stage 3 onward. Mode-specific reference files spell out Stages 1–2 in detail; Stages 3–8 below apply to both modes (Stage 7 is firecrawl-only, but Stage 8 still gates on it).

Stage 3 — Determine target paths

Read the live repo structure at runtime — do NOT assume a snapshot. Source of truth is the repo's root CLAUDE.md (/Users/hannahstulberg/in-the-weeds-substack-articles/CLAUDE.md).

Series → target folder mapping (current):

• Standalone → standalone/<folder-slug>/
• Claude Code for Everything → claude-code-for-everything/<NN>-<folder-slug>/ where NN = next zero-padded number
• Tool School → tool-school/<NN>-<folder-slug>/

To find <NN>: ls the series folder, find highest existing <NN>- prefix, add 1, zero-pad to 2 digits.

Computing <folder-slug> (do not reuse the input slug verbatim):

The input slug often carries a series prefix (cc4e-, tool-school-) inherited from ~/hannah-personal-agent/ source folders. Repo convention strips those prefixes — folder names are title-derived, short, and don't repeat the series. Compute the folder slug as follows:

1. Read the article title (from frontmatter or H1).
2. Drop any leading series-name fragment that duplicates the series itself: Claude Code for Everything: , Tool School: (case-insensitive). Also drop a parenthetical aside at the end if the title is long.
3. Lowercase, replace non-alphanumerics with hyphens, collapse runs of hyphens, strip leading/trailing hyphens.
4. Cap at ~60 chars — pick the most descriptive substring.
5. Cross-check existing sibling folders in the target series for style match.

Examples (existing repo folders, all derived from titles, none containing the series prefix):

• "How the Guy Who Built It Actually Uses It" → 02-how-the-guy-who-built-it-actually-uses-it
• "Why AI Gets Dumber The Longer You Talk To It" → 03-why-ai-gets-dumber-the-longer-you-talk-to-it
• "Keeping Up with the Claude Code Treadmill (30 Claude Code Tips & Tricks)" → 08-keeping-up-with-the-claude-code-treadmill (drops the parenthetical)

Confirm the computed folder slug with the user before creating the directory if it differs from the input slug.

If the root CLAUDE.md describes a series that isn't in the table above, follow whatever convention root CLAUDE.md documents — the live file always wins over this skill's defaults.

Compute:

• target_folder = <series-folder>/<folder-slug>/ (Standalone) or <series-folder>/<NN>-<folder-slug>/
• target_article = <target-folder>/article.md
• target_claude_md = <target-folder>/CLAUDE.md
• target_images = <target-folder>/images/

Stage 4 — Branch + copy

Sequential — branch must exist before any writes.

git pull origin main
git checkout -b add/<slug>
mkdir -p <target-folder>/images

Local mode:

cp <source-article-md> <target-folder>/article.md
[ -d <source-images> ] && cp -r <source-images>/* <target-folder>/images/

Firecrawl mode:

mv .firecrawl/<slug>/<scraped>.md <target-folder>/article.md
# For each image URL in the scraped content, download to images/
# Then rewrite article body so image refs point at images/<filename>

Sanitize article.md (both modes, after copy):

Strip working-doc artifacts and convert source draft format to the public-repo format. Apply BEFORE Stage 5 so line numbers in the generated CLAUDE.md match the final file.

1. Top-of-file Google Docs draft block. Hannah's drafts often include **Formatted Google Docs:** followed by a blank line + bullet links to docs.google.com (e.g., v2 draft, v2 push 2026-04-29) + a blank line. Delete the entire block.

2. Convert H1 + **Subtitle:** to YAML frontmatter. Source drafts open with # <Title> then a blank line then **Subtitle:** *<subtitle>*. The public-repo format is YAML frontmatter at the top of article.md. Use this canonical schema (described fully in /Users/hannahstulberg/in-the-weeds-substack-articles/.claude/article-frontmatter-schema.md):

   ---
   title: "<title>"
   subtitle: "<subtitle without surrounding asterisks>"
   authors:
     - name: "Hannah Stulberg"
       publication:
         name: "In the Weeds"
         url: "https://hannahstulberg.substack.com/"
   published: YYYY-MM-DD               # today's date for unpublished articles
   series: "Claude Code for Everything"  # OMIT entirely for standalone
   series_number: 8                      # OMIT entirely for standalone
   original_url: "https://..."           # canonical Substack/blog URL of the post
   license: "CC BY-NC 4.0"
   media:                                # OMIT unless companion media exists
     youtube: "https://..."
   ---
   

After sanitize, re-run grep -n "^## \|^### " <target>/article.md so Stage 5 uses the correct line numbers.

Update temp file: {"state": "files_copied", "branch": "add/<slug>", "files": [...]}.

Stage 5 — Generate article-level CLAUDE.md

Format MUST match the existing repo pattern. Two canonical examples Claude can read for reference:

• standalone/stop-typing-start-talking/CLAUDE.md (4-column with all rows filled with Substack links)
• claude-code-for-everything/05-claudemd-files/CLAUDE.md (4-column with mixed [Link] and — rows)

Required output structure:

# <Article Title>

**Author:** <names> | **Published:** <YYYY-MM-DD>

## Article Map

| Section | Line | Coverage | Substack |
|---------|------|----------|----------|
| <heading> | <line> | <coverage> | — |
| ... | ... | ... | ... |

Notes:

• For featured-contributor articles (e.g., Aakash's Team OS): **Author:** Aakash Gupta (featuring Hannah Stulberg) | **Published:** ....
• One row per H2 (^## ) and H3 (^### ) heading. H1 is the title (already in the file header), don't include it in the map.
• Line = grep -n "^## \|^### " <target>/article.md returns the line numbers.
• Coverage = a 1-phrase summary auto-drafted by reading 2-3 lines after each heading. Keep under 15 words. No filler.
• Substack = — for every row in local mode. In firecrawl mode, Stage 7 replaces these with verified URLs.

Stage 6 — Cascade updates

Batch ALL of these edits in ONE message (independent files). Read the live root CLAUDE.md first to identify exact insertion points; do not assume line numbers.

1. Root CLAUDE.md — Article index table

Insert one new row in the article index. Format mirrors existing rows:

| <#> | <Title> | <Series> | <Author(s)> | <key topics, comma-separated> | `<path>` |

<#> = next sequential article number across the whole repo (count existing rows + 1).

2. Root CLAUDE.md — Topic quick lookup table

Auto-draft topic entries by reading the article's H2/H3 headings. For each meaningful heading (skip purely decorative ones like "Title & subtitle"), insert:

| <Topic phrase> | <Article short name> | <Section heading> |

Insert at the bottom of the existing Topic quick lookup table. Don't try to alphabetize — append.

3. Root CLAUDE.md — Key metaphors table

Only add rows IF the article uses an explicit metaphor pattern. Heuristic: scan the article body for phrases like:

• is like a <noun>
• the <noun> metaphor
• think of it like <noun>
• imagine a <noun>

If found, draft a row:

| **<Metaphor name>** | <What it represents> | <Article name>, "<Section>" |

If no clear metaphor, skip this entirely. Don't force it.

4. Root CLAUDE.md — About / Contributors list

Inside the "About" section near the bottom, the contributor bullet list. If the article introduces a NEW contributor (one not already listed), add:

- [<Full Name>](<their-url>) - <Role>, <Article title>

If the contributor is already listed, no edit here.

5. Series-level CLAUDE.md

Pick the right one based on series:

• standalone/CLAUDE.md
• claude-code-for-everything/CLAUDE.md
• tool-school/CLAUDE.md

Insert one new row in the series articles table. Format mirrors existing rows in that file (read the live file for the exact column schema).

6. .claude/welcome.md

Two edits:

• Bump the article count number by 1 (currently says "12 published articles" or similar — find and increment).
• If new contributor: add their name to the credit list line. Format follows existing pattern.

7. CONTRIBUTORS.md

Only edit if the article introduces a NEW contributor. Append a new bio block following the existing pattern:

## <Full Name>

<1-2 sentence bio>

- [<Publication> (<Substack/blog>)](<url>)
- [LinkedIn](<linkedin-url>)
- [<Other relevant link>](<url>)

**Articles in this repo:** <Article title> (<role: co-author / collaborator / author featuring Hannah>)

If the bio info isn't available from the source metadata, draft a stub like <TODO: bio> and surface it in the PR body for Hannah to fill before merging.

For repeat contributors: update their existing bio block's **Articles in this repo:** line to include the new article. Don't add a new bio block.

8. LICENSE

Add per-article copyright entry under "Individual Article Attribution:" based on relationship.

Relationship	Template
Hannah + co-author (e.g., GitHub 101 with Sidwyn)	"<Title>" Copyright (c) <year> Hannah Stulberg and <Co-author Name> Co-authored. Included with co-author's permission.
Hannah + collaborator originally elsewhere (e.g., Joel)	"<Title>" Copyright (c) <year> Hannah Stulberg and <Name> Originally published on <Publication>. Included with permission.
External author featuring Hannah (e.g., Aakash)	"<Title>" Copyright (c) <year> Hannah Stulberg Written by <Name> (<Publication>), based on Hannah Stulberg's original content and expertise. Originally published on <Publication>. Included with permission.
Hannah only	No LICENSE entry — covered by the catch-all "All other articles" line.

For repeat contributors (e.g., Sidwyn already has GitHub 101 entry, now adding a second article): add a brand-NEW entry for the new article. Each article gets its own copyright line. Don't consolidate or modify the existing entry.

If the relationship is ambiguous, draft the most permissive template (Co-authored) and call it out in the PR body for Hannah to refine before merging.

After all edits: {"state": "cascade_done", "files_modified": [...]} in temp file.

Stage 7 — Substack URL backfill (firecrawl mode ONLY)

Skip this stage entirely in local mode (post may not be live yet — Hannah will run the companion skill later). In local mode, write {"substack_backfill": {"skipped": true, "reason": "local mode"}} to the temp file so the Stage 8 pre-commit gate knows it wasn't forgotten.

In firecrawl mode the live URL is the input — fill the per-row Substack column inline before commit. This stage is mandatory in firecrawl mode. Skipping it ships a CLAUDE.md with — placeholders that should have been real anchor URLs.

Invoke the companion skill via an explicit Skill tool call (not prose intent — actually call it):

Skill(skill="add-substack-urls-to-article", args="slug=<slug> url=<input-url> subroutine=true")

The companion runs Stages 0–3 (resolve target → fetch post → match headings → edit CLAUDE.md) but skips its own branch/commit/push/PR. All edits stay on the same add/<slug> branch.

After the companion returns, write to temp file:

{"substack_backfill": {"rows_filled": N, "rows_left_as_dash": M, "skipped": false}}

If companion fails (404, anchor map empty, etc.), write:

{"substack_backfill": {"skipped": false, "failed": true, "error": "<reason>", "rows_left_as_dash": <total>}}

Then continue to Stage 8 — Hannah can re-run the companion skill manually later. Don't abort the pipeline.

Stage 8 — Pre-commit gate + commit + push + PR

Pre-commit gate (MANDATORY before git add):

1. Read the temp file. Confirm substack_backfill key exists. If missing in firecrawl mode → ABORT with: "Stage 7 not run. Invoke /add-substack-urls-to-article before proceeding."
2. Grep the target article-level CLAUDE.md for unfilled rows: grep -c '| —$' <target>/CLAUDE.md. In firecrawl mode, if any non-intro rows are still — AND substack_backfill.skipped == false AND substack_backfill.failed != true → ABORT with the row count and the message: "Stage 7 ran but rows still empty. Re-check companion output."
3. Only proceed to git add once the gate passes.

Then sequential:

git add <specific-files>  # never use -A or .
git commit -m "Add new article to repo"
git push -u origin add/<slug>
gh pr create --title "Add new article to repo" --body "<body>"

Sequential:

git add <specific-files>  # never use -A or .
git commit -m "Add new article to repo"
git push -u origin add/<slug>
gh pr create --title "Add new article to repo" --body "<body>"

Public-repo rules apply (per commit-push-pr conventions):

• Title ≤70 chars, vanilla, no article name / series number / contributor name.
• Body: 1-2 sentences. Don't list the article title, series number, or any specifics. List counts only ("8 files updated", "1 new contributor added").

If LICENSE was updated with a "draft template" (ambiguous relationship), call that out in the PR body so Hannah refines before merging.

If Stage 7 fired but had partial failures, list count of unfilled rows in PR body.

Final temp file: {"state": "done", "pr_url": "..."}.

Crash recovery

Single temp file: /tmp/add-substack-article-<slug>.json. Updated after every stage. Schema:

{
  "state": "mode_selected | source_found | metadata_read | files_copied | article_claude_md_generated | cascade_done | substack_backfill_done | committed | pushed | done",
  "slug": "<slug>",
  "mode": "local | firecrawl",
  "input": "<original input>",
  "source_path": "/abs/path",
  "source_article_md": "/abs/path",
  "title": "<title>",
  "series": "Standalone | Claude Code for Everything | Tool School",
  "series_number": 8,
  "target_folder": "/abs/path",
  "target_files_to_modify": ["/abs/path/...", ...],
  "contributors": ["..."],
   
   
   
   
    N  M

A fresh session can read this file and resume from the last completed stage. On resumption: read the temp file first, check current branch + git state, jump to the next stage past state. Don't redo completed stages (e.g., don't re-copy files if state >= "files_copied").

Source-of-truth rules

• Title, series, date, contributors [local mode]: read from source CLAUDE.md first (<source>/CLAUDE.md), manifest.yaml fallback. If both missing, ask user.
• Title, date, author [firecrawl mode]: extract from scraped markdown (H1 + byline + date). Ask user only for the field that fails extraction, not all three.
• Slug [local mode]: strip [DD-MM-YYYY]- date prefix from published folder names; otherwise use the unpublished folder name as-is.
• Slug [firecrawl mode]: the last /p/<slug> segment of the URL.
• Series [local mode]: as specified in source CLAUDE.md / manifest.yaml.
• Series [firecrawl mode]: default to Standalone for external posts. User can override at invocation.
• Article number in series: scan target repo's series folder for highest <NN>- prefix; new = highest + 1, zero-padded. Standalone has no number prefix.
• Substack column: — in local mode. Filled inline by Stage 7 in firecrawl mode.
• Topic + metaphor entries: auto-draft, no pause. Hannah reviews in the PR.

Things this skill must NOT do

• Don't ask the user mid-run for topic / metaphor / coverage entries. Auto-draft, ship to PR, let Hannah edit there.
• Don't git add -A or git add .. Always specific files.
• Don't put article titles, series numbers, or contributor names in the commit message or PR title (public repo).
• Don't overwrite an existing target folder. If <target-folder> already exists, abort with a clear message naming the collision.
• Don't proceed if the source can't be located. Ask user; don't guess.
• Don't fabricate a Substack URL. In local mode, leave column as —. Only firecrawl mode (where URL is the input) fills it.

References

• references/source-mode-local.md — loaded only when input is a slug. Personal-agent folder discovery, source CLAUDE.md / manifest.yaml metadata extraction, version-suffix selection (prefer -final.md, else highest -v<N>.md).
• references/source-mode-firecrawl.md — loaded only when input is a URL. Firecrawl CLI invocation per ~/.claude/docs/firecrawl.md, scrape format expectations, image download + rewriting, metadata extraction from scraped content.