Content Ideas

content-ideas

Your For You page. Scrapes every platform where your tracked creators publish, scores what's performing, and turns it into content ideas you can act on. Designed to run daily — each run creates a dated feed under $CONTENT_HOME/research/.

The output is a single self-contained HTML page (two tabs: Posts — one sortable, filterable feed merging tracked-account posts and discovered niche outliers — and Ideas) that you can open in a browser, react to, and keep. Reactions are captured for future personalization.

Resolve the skill directory

Everything this skill runs lives under its own folder. The skill installs the same way on Claude Code and Codex, so resolve SKILL_DIR against both plugin caches (and a plain repo checkout) once, before anything else:

# 1) Codex plugin cache, or a repo cloned into ~/.codex/skills/ (latest wins on upgrade).
SKILL_DIR="$(ls -d "$HOME/.codex/plugins/cache/"*/content-ideas/*/skills/content-ideas/ "$HOME/.codex/skills/"*/skills/content-ideas/ 2>/dev/null | sort -V | tail -1)"
SKILL_DIR="${SKILL_DIR%/}"

# 2) Claude Code plugin cache.
if [ -z "$SKILL_DIR" ] || [ ! -f "$SKILL_DIR/scripts/scrape.py" ]; then
  CLAUDE_ROOT="$(ls -d "$HOME/.claude/plugins/cache/content-ideas/content-ideas/"*/ 2>/dev/null | sort -V | tail -1)"
  CLAUDE_ROOT="${CLAUDE_ROOT%/}"
  [ -n "$CLAUDE_ROOT" ] && [ -f "$CLAUDE_ROOT/skills/content-ideas/scripts/scrape.py" ] && SKILL_DIR="$CLAUDE_ROOT/skills/content-ideas"
fi

# 3) Plugin root passed by the host, or a repo checkout / local dev.
if [ -z "$SKILL_DIR" ] || [ ! -f "$SKILL_DIR/scripts/scrape.py" ]; then
  for dir in "${CLAUDE_PLUGIN_ROOT:-}/skills/content-ideas" "${CLAUDE_PLUGIN_ROOT:-}" "${GEMINI_EXTENSION_DIR:-}/skills/content-ideas" "./skills/content-ideas" "."; do
    [ -n "$dir" ] && [ -f "$dir/scripts/scrape.py" ] && SKILL_DIR="$dir" && break
  done
fi

echo "$SKILL_DIR"

If you can already see this file's path, just use its directory. The two scripts you'll call are $SKILL_DIR/scripts/scrape.py and $SKILL_DIR/scripts/generate_feed.py. The renderer template is $SKILL_DIR/assets/for-you-template.html (the generator finds it automatically).

Resolve the content home

All persistent files this skill reads and writes — the brand/ profile and the dated research/ runs — live under one stable base, never the current working directory. The skill runs daily and is invoked from anywhere, so the base must be the same every time or it loses the profile and the run history. Resolve it once and capture the concrete path:

CONTENT_HOME="${CONTENT_HOME:-$HOME/Documents/Content}"
mkdir -p "$CONTENT_HOME/brand" "$CONTENT_HOME/research"
echo "$CONTENT_HOME"

Throughout this guide every brand/... and research/... path is relative to $CONTENT_HOME (so brand/profile.md means $CONTENT_HOME/brand/profile.md). Use the printed absolute path for every Read/Write of those files — the file tools don't expand shell variables, so writing a bare brand/profile.md would land it in the wrong directory. (Credentials stay separate, in ~/.config/content/.env.) The scrape/generate scripts read CONTENT_HOME themselves, so a relative research/{today} passed to them resolves here too.

Step 0: First-run setup

Run this before anything else, even if the user gave a topic. Detect first run by checking whether ~/.config/content/.env exists and contains SETUP_COMPLETE=true. Check silently. If it's already set up, skip to Step 1.

0a. Welcome + API key

Setup has three quick parts: an API key, your profile (built from your own channels), and the competitors you want to track. Only the key is required — the rest the skill bootstraps for you and you can refine any time. Nothing to install; one ScrapeCreators API key covers all four platforms — X, Instagram, TikTok, and YouTube (including transcripts).

Show this as a normal message, then call AskUserQuestion (don't repeat the welcome inside the modal):

I turn your social presence into a daily For You feed: I build a profile from your own channels, track the competitors you pick, and surface what's performing as content ideas backed by real engagement. I just need a ScrapeCreators API key (one key covers all four platforms; 100 free calls, no card).

AskUserQuestion — "Add your ScrapeCreators API key?"

• Open scrapecreators.com to grab a free key
• I'll paste a key now
• Skip for now

If they pick "Open scrapecreators.com", run open https://scrapecreators.com, then ask them to paste the key. When the user pastes a key, write ~/.config/content/.env (create dirs; append, don't clobber other keys):

SCRAPECREATORS_API_KEY={key}
SETUP_COMPLETE=true

If they skip, write only SETUP_COMPLETE=true.

0b. Manual alternative

If they'd rather configure by hand, tell them to add those two lines to ~/.config/content/.env. Offer to write the file if they paste the key here.

0c. Build your brand profile

This is what personalizes everything: ideas get framed against your niche, pillars, and goal, and checked against what you've already posted. Build it from the user's own presence rather than a long questionnaire.

Ask for their own channels (AskUserQuestion: "Set up your profile now?" → I'll share my handles / Skip — I'll add it later). When they share handles — free-form across any platforms (@me on X, a YouTube channel, a TikTok, etc.) — normalize them into the {platform: [handle]} shape and scrape them like competitors, but over a much wider window (--days 90, the max) so you characterize their work from a full quarter, not just recent posts:

python3 "$SKILL_DIR/scripts/scrape.py" \
  '{"x": ["me"], "youtube": ["@mychannel"]}' \
  --pillars "" --days 90

From the returned posts (plus comments/transcripts), draft the profile:

• Niche, Audience, Voice Notes — infer from recurring topics, framing, tone.
• Content Pillars — the 3–5 themes their posts actually cluster into. These drive --pillars on every future run, so get them right.
• My Social Profiles — handle, follower count, bio, and a one-line content- style note per platform, taken from the scrape.
• Target Platforms / Research Channels — the platforms they're active on.
• Search Terms — concrete keywords from their top topics.

Two things you can't scrape — ask (AskUserQuestion), then fold the answers in:

• Content Goal — why they post (lead gen / awareness / growth / thought leadership / selling…), where they drive traffic, and what they're promoting.
• Pillar confirmation — show the 3–5 pillars you inferred and let them edit or confirm before writing.

Write brand/profile.md per the schema in FILE-SCHEMAS.md. If the scrape returned enough of their own posts, also write an initial brand/my-content.md (performance summary, what's working, topics covered, and audience requests distilled from their comments) — this powers anti-cannibalization and the "your audience is asking for" banner from day one.

If they skipped (or there's no API key yet to scrape with), don't block: build a minimal brand/profile.md from a 2–3 question Q&A (niche, rough pillars, goal), note that re-running setup with a key auto-enriches it, and move on.

0d. Track competitors

Ask who they want to track (AskUserQuestion: list them now / skip and use an example). If they list handles, create brand/tracked-accounts/{platform}.md files per the schema in the plugin's FILE-SCHEMAS.md. If they skip, run a small example so they see the shape, and tell them they can add real competitors later.

End of first-run setup. Then continue with the user's original request.

Step 1: Load context

1a. Ingest the previous run's feedback into taste memory

Before anything else, fold the last run's reactions into your memory — this is what makes each run better than the one before. List the dated subfolders of $CONTENT_HOME/research/ (YYYY-MM-DD) and take the most recent one. If it has a feedback.json, read it and distill each entry in reviews[] (▲ "more like this" / ▼ "less" / a note) into the generalizable taste signal, not the one-off:

• "▲ on three contrarian takes in the user's niche" → "gravitates toward contrarian takes"; "▼ on listicles" → "listicle formats don't land." A note often states the reason directly — use it.
• Record these to your project memory (the auto-memory you maintain) as the user's content taste — the same place 1b recalls from. Update an existing taste note rather than duplicating it; let a single ▼ inform, not override, an established preference. Don't record one-off reactions with no pattern, anything already obvious from brand/profile.md, or post/run specifics (those live in research/). Taste only.

If there's no prior dated folder, no feedback.json, or no reactions in it, skip silently. If auto-memory isn't available in this environment, skip too — the reactions stay in feedback.json for whenever it is. (The current run's reactions are ingested by the next run, the same way — there's no end-of-run distillation step.)

1b. Recall taste and load brand context

Read whatever brand context exists (all optional — degrade gracefully):

• brand/profile.md — niche, pillars, search terms, content goal, audience
• brand/tracked-accounts/*.md — tracked creators per platform
• brand/my-content.md — the user's own content performance + audience requests

Recall the user's content taste from your memory. This skill stores an evolving taste profile in your project memory (the auto-memory you maintain). Before generating ideas, recall what you know about what this user gravitates toward — preferred topics, formats, angles, creators they keep saving, and what doesn't land for them. If relevant taste signals are already surfaced in context, use them; if not and memory is available, look for taste notes tagged for this skill. This is the single most important personalization input: engagement metrics measure what audiences like, taste memory measures what this user likes. If auto-memory isn't available, fall back to engagement signals alone (and to brand/my-content.md if present).

If there are no tracked accounts and no topic filter, ask for handles or a topic before scraping.

1c. Refresh your own content (my-content.md)

Before generating ideas, bring brand/my-content.md up to date — this is the per-run counterpart to the one-time build in Step 0c, and it's what keeps anti-cannibalization and the "your audience is asking for" banner honest as the user keeps posting. (my-content.md is declared updated each run in FILE-SCHEMAS.md; this is the step that does it.)

Take the user's own handles from the ## My Social Profiles section of the brand/profile.md you just loaded, normalize them into the {platform: [handle]} shape, and re-scrape them over a window wide enough to catch their own cadence (--days 30 — a creator's own posts are sparser than the merged competitor feed, but keep it "recent," not the 90-day profile build from Step 0c):

python3 "$SKILL_DIR/scripts/scrape.py" \
  '{"x": ["me"], "youtube": ["@mychannel"]}' \
  --pillars "<pillars from profile.md>" --days 30

The scraper already pulls comments on the top posts, so the returned data carries the audience replies you need. Rewrite brand/my-content.md from it per the schema in FILE-SCHEMAS.md (performance summary, what's working / not, topics covered, and audience requests distilled from the comments) — it's replaced, not appended. Use this fresh version, not the copy you read in 1b, for the rest of the run.

Best-effort — never block the feed. If profile.md has no own handles (the user skipped profile setup), or the scrape returns nothing or errors, keep the existing my-content.md and continue. This refresh is an enrichment, not a gate.

Step 2: Create the daily run folder

List existing dated subfolders of $CONTENT_HOME/research/ (YYYY-MM-DD). The most recent one that is not today is the last-run date — pass it as --since in Step 3 so the scrape only keeps posts on/after that day. If there are no prior dated folders, there's no --since.

Either way, the scraper enforces a recency window so the daily feed never surfaces stale posts: by default it keeps only the last 7 days (--days). --since can only narrow that window, never widen it — so first runs and long-gap runs are both bounded to a week by default. (The script's hard cap is 90 days; for the daily feed keep it tight — a month at most. The 90-day window is for one-off profile builds in Step 0c, not the daily feed.)

Create $CONTENT_HOME/research/{today}/.

If $CONTENT_HOME/research/{today}/feed-data.json already exists, ask whether to:

• Refresh — re-pull and rebuild (reuse the same --since / --days)
• Expand — widen the window: drop --since and/or raise --days (keep the feed within ~30 days) when the user wants more than the last week
• View — just (re)open the existing feed (skip to Step 6)

Step 3: Scrape competitors

Build a JSON object mapping each platform to its tracked handles. Pass content pillars (from brand/profile.md, or the user's niche/topic) via --pillars so the script scores relevance, and the last-run date via --since. Leave --days at its default (7) unless the user asks for a wider window, then raise it (max 31).

python3 "$SKILL_DIR/scripts/scrape.py" \
  '{"x": ["h1","h2"], "instagram": ["h3"], "youtube": ["@h4"]}' \
  --pillars "<the user's content pillars>" \
  --since 2026-04-15 \
  --days 7

Tell the user this takes a few minutes; progress streams to stderr. The script fetches all accounts in parallel, drops anything outside the recency window, scores engagement and relevance, flags outliers, and pulls comments/transcripts on top posts. It returns:

{ "results": { "x": { "h1": [ {post}, ... ] } }, "errors": [] }

Each post has text, url, author, date, platform, engagement, score (weighted), relevance (0–1 vs pillars), baseline (Nx the account average), outlier (bool), and — on top posts — comments / transcript.

On errors: report which accounts failed and proceed with what came back.

Ad-hoc: fetch specific posts by URL

When the user hands you specific post URLs (a competitor's viral post, a link they saw), use URL mode instead of profile mode. It returns a flat [post] array with the same shape:

python3 "$SKILL_DIR/scripts/scrape.py" urls "https://x.com/u/status/1" "https://www.tiktok.com/@u/video/2" --pillars "..."

Step 4: Review the scored data

The script pre-computes score, baseline, relevance, and outlier. Identify the top-performing posts and the topics/themes/angles driving engagement — especially high-relevance ones. This is the raw material for the Ideas tab.

Step 5: Build the feed

Two tabs. Everything shown has proven engagement. Build a FEED_DATA object and write it (Step 6). Field-by-field structure is in the plugin's FILE-SCHEMAS.md (feed-data.json).

Tab 1 — Posts. One flat posts[] array merging two sources into a single sortable, filterable feed (the page handles sorting and grouping client-side — do not pre-sort or pre-group):

• Tracked-account posts — every post from tracked accounts (no engagement gate). Set performance / performanceDirection vs the account baseline (e.g. "+210% vs baseline", "up").

• Discovered niche outliers — statistical outliers (outlier: true, z-score 2+, or baseline 2x+). Set zScore and a why line.

  Per post, regardless of source, provide: a 1–3 sentence text summary, url, handle + displayName (creator filter), platform, an engagement object, a hook callout when notable, and the two fields that make the feed work — timestamp (ISO 8601, drives Recent sort + relative time) and sortValue (numeric total engagement/reach, drives the default Popular sort). A post is flagged as an outlier (intensity-scaled badge + accent bar) whenever it has a zScore or performanceDirection: "up" — so a tracked post that beat its baseline shows as an outlier too.

Tab 2 — Ideas. The one place you editorialize (label it as AI suggestion). Generate up to 10 ideas, each with: a specific differentiated angle, real evidence from competitor performance, and clear differentiation from what competitors already covered.

For the generative craft — turning a topic into a differentiated angle, writing hooks, classifying funnel stage (TOFU/MOFU/BOFU), aligning CTAs, repurposing across platforms, and producing a full brief — read references/content-strategy.md. The short version to keep in mind while building this tab:

• Make YOUR version, never repackage. A good angle answers at least one of: what do you know the original creator doesn't (expertise), what have you done the audience hasn't seen (access), or where do you disagree (contrarian)?
• Anti-cannibalization. When brand/my-content.md exists, don't re-pitch a topic the user already covered unless the angle has a genuine differentiator (more depth, different format, an update, a response to feedback). Note prior coverage explicitly.
• Own-audience demand wins. Requests from the user's own audience (brand/my-content.md) outrank competitor signals — foreground them in the brief's "why now."
• Taste memory biases selection. An idea that aligns with the taste signals you recalled in Step 1 (topics/formats/angles this user gravitates toward) is a stronger pick than one justified by engagement alone — and worth calling out ("this fits a pattern you keep coming back to"). Conversely, deprioritize anything that matches a recorded "doesn't land" signal.

Step 6: Write and open the feed

Write the feed data to $CONTENT_HOME/research/{today}/feed-data.json — a JSON object with keys meta, posts, ideas (see FILE-SCHEMAS.md). Do not write HTML yourself; the generator embeds this JSON into the template.

Then render it. Default to the live server (lets the user react to items, which saves to feedback.json for future personalization):

python3 "$SKILL_DIR/scripts/generate_feed.py" "$CONTENT_HOME/research/{today}"

This starts a local server and automatically opens the feed in the user's default browser. Still hand the user the http://localhost:<port> URL the command prints, so they can reopen it if the tab closes. (Pass --no-browser to suppress the auto-open; the URL is printed either way.) The command runs in the foreground until the user stops it with Ctrl+C, so run it in the background if you need to keep working.

In a headless/no-display environment, write a self-contained file instead and point the user at it (the page lets them download their reactions):

python3 "$SKILL_DIR/scripts/generate_feed.py" "$CONTENT_HOME/research/{today}" --static
# → $CONTENT_HOME/research/{today}/for-you.html

Then present a short text summary (post count, how many are outliers, a couple of standout posts) and the page location.

Step 7: Offer next steps

The user reacts to the feed in the browser; their reactions save to research/{today}/feedback.json on their own — automatically in server mode, or via the page's download button in static mode. There's no "done" signal and nothing for you to read or distill now: the file just accumulates reactions, and the next run folds them into taste memory at Step 1a. This keeps the workflow simple and, crucially, captures reactions the user makes after this conversation has ended.

Offer to: dig deeper on any idea, add/remove tracked accounts, or rerun with a different topic focus.

Notes

• Reactions / feedback → taste memory. The feed page lets the user mark items (▲ more like this / ▼ less / a note) across both tabs. In server mode these save to research/{date}/feedback.json automatically as the user clicks; in static mode the user downloads that file into the run folder. The file is just an accumulating list of reactions — no status, no submit step. The next run reads the previous run's feedback.json at Step 1a and distills it into your project memory so future runs are personalized — there is no taste file; taste lives in auto-memory.
• No API key = no run. Both profile and URL mode require SCRAPECREATORS_API_KEY — every platform, including YouTube transcripts, goes through ScrapeCreators. If the key is missing, the script returns an error; stop and show setup instructions rather than inventing data.