ig-feed-research

Scans the logged-in Instagram HOME feed (instagram.com — reels, carousels, and image posts together) in the real Chrome profile (Claude-in-Chrome MCP) for on-niche content ideas AND tailors the algori

SKILL.md

GitHub

Video Studio: this skill lives in engines/content-os (Content OS, inside this repo). Before running anything from it, cd engines/content-os and use that folder's copy (.claude/skills/ig-feed-research/), which has its scripts and reference files. Every path below is relative to that folder. This copy only lets Claude find the skill from the Video Studio folder.

IG Home-Feed Research — Ideas + Not-Interested Tailoring in One Loop

Scroll the real logged-in Instagram home feed (instagram.com), classify every post against the niche, keep what's replicable for the brand, and hit three dots → "Not interested" on everything that isn't. Same thesis as yt-feed-research: the dismissals aren't cleanup — they're the research engine. Every "Not interested" teaches the suggestion engine, and the feed compounds into a daily idea source.

Why the home feed and not the reels viewer: the home feed carries carousels and image posts too (a whole content lane the reels tab hides), and it's the only surface on IG web where "Not interested" exists. One surface, one loop — scan and tailor in the same pass.

How to Trigger

  • "scan my instagram feed for ideas"
  • "ig feed research"
  • "tailor my instagram feed"
  • "what's my ig pushing right now"

When to Use This vs ig-competitor-research

ig-feed-research ig-competitor-research
What the algorithm pushes at the creator right now Tracked handles' weekly best
Real Chrome, logged-in home feed Apify scraper, no browser
Discovers unknown accounts breaking out Fixed competitor list
Side effect: permanently improves the feed No side effects

Prerequisites

  • Claude-in-Chrome MCP connected to the real Chrome profile (must be logged into Instagram). Terminal Claude Code only — the desktop app spawns the MCP tab hidden and scraping silently fails.
  • Instagram UI language must be English. feed-tools.js matches English strings ('Suggested for you', 'Sponsored', 'Learn more', 'More options', 'Not interested', "Don't suggest posts from"); any other locale classifies every row as followed, misses ads, and returns no-entry on every dismissal.
  • Nothing else: no Apify, no yt-dlp. (VidIQ's VPH/multiplier badges only render in the reels viewer, not the home feed — here the signals are likes, comments, and judgment.)

Inputs

  • Idea target — default 20 saved ideas; user can override inline ("get me 30").
  • Niche filter — the classification question is: is this on-niche, or is the format directly replicable for the brand? If neither → dismiss. The KEEP/DISMISS test is derived at run time from the repo's context files, never hardcoded here. Before the first __feed() call, read these and build the filter:
    1. CLAUDE.md Creator Profile → the Niche line. This one-liner defines "on-niche". If it still contains a << placeholder, stop and tell the user: "apply your brand kit first" (brand-kit.md), then end the run.
    2. backbone/offer.md → what is sold (anything that could seed content promoting that offer is KEEP) and, when present, its "What is NOT sold" list (content whose only angle is a NOT-sold thing is DISMISS).
    3. backbone/icp.md → who is served and who is not (content that speaks to the ICP's problems is KEEP; content aimed at the not-served segments is DISMISS).
    4. competitor-list.md → every IG handle listed there is always KEEP when it shows up in the feed.
    5. knowledge/*.md when present → recent research signal on what is working in the niche. From that, state the filter in one line each before scanning: KEEP = topics matching 1–5, plus any standout carousel/image format the brand could replicate for the carousel lane; DISMISS = the categorical off-niche buckets you infer from the above (topics no one in the ICP would ask about, formats the brand cannot film or reproduce); NEUTRAL = on-niche but weak or ambiguous. Print the derived KEEP/DISMISS one-liner in the report's {{FEED_STATE}} fieldnote (the template has no separate slot for it) so the user can veto it. Example only: a creator whose Niche line says "AI leverage for creators" would KEEP AI builds and model news and DISMISS fitness, cars, and real estate.

(No concrete list ships: derive KEEP/DISMISS on the first run from the Niche line in CLAUDE.md, backbone/icp.md, backbone/offer.md and competitor-list.md, state it in the report, and pin it here once the creator confirms it.)

Hard Safety Rules (IG action-blocks bots — YouTube doesn't care, Instagram does)

  • Read + dismiss only. Never like, follow, comment, save, or share from this loop.
  • One dismissal at a time, with the waits baked into the helpers. Never fire dismissals in batches.
  • Cap ~40 dismissals per session, max 2 sessions a day. Tailoring compounds across sessions anyway.
  • If any "Try again later" / "We limit how often..." toast appears: stop dismissing immediately, finish the pass read-only, flag it in the report.

The Loop

read CLAUDE.md Niche line + backbone/offer.md + backbone/icp.md + competitor-list.md → derive KEEP/DISMISS (Inputs)
  → navigate instagram.com → inject scripts/feed-tools.js → __feed() → classify every row
  → KEEP: save to idea list (reels AND carousels AND posts)
  → DISMISS (suggested, off-niche/not replicable): __feedDismiss(ix,'post')
      · categorically wrong account → __feedDismiss(ix,'account')
  → NEUTRAL / followed-but-off-niche / AD: no action (flag repeat offenders in the report)
  → __more() → __feed(<lastSeenIx+1>) → classify the new tail → repeat
  → until idea target hit or feed stalls → fill assets/report-template.html → research/ → open it

Step 1 — Open the home feed

tabs_context_mcp {createIfEmpty:true}, navigate to https://www.instagram.com/, wait ~4s. If a login wall renders, stop and tell the user.

Step 2 — Inject helpers + extract

Paste the entire scripts/feed-tools.js into one javascript_tool call (re-inject after every navigation — page loads wipe window.*). __feed(from) caches one row per loaded article on window.__rows and returns only the rows from index from onward (__feed() = all rows, first pass only): ix | SUGGESTED/followed/AD | user | reel/carousel/post | likes | caption. Tool results truncate around 1.5KB, so after the first pass always read the tail, never the whole list. The feed lazy-loads a few articles at a time — __more() scrolls one step and reports the article count, so a scan pass is: __more() → __feed(<lastSeenIx+1>) → classify the tail. Batch 2–3 [__more, __feed(N)] pairs per browser_batch call when no dismissals are pending.

Safety-filter trap (same as YT): never return raw hrefs, query strings, or innerHTML from javascript_tool — the harness blocks the whole result ([BLOCKED: Cookie/query string data]). Cleaned text only; the helpers already enforce this.

Step 3 — Classify every row

The test: on-niche, or replicable for the brand? (the KEEP/DISMISS filter derived in Inputs)

  • KEEP → idea list: username, kind (reel/carousel/post), likes, comments, hook/caption line, one-line concept, remix angle. Carousels are first-class keeps — they feed the carousel-generator lane directly.
  • DISMISS → __feedDismiss(ix,'post') for a SUGGESTED row that fails the test; (ix,'account') when the account will never be on-niche (its whole catalog sits in one of the derived DISMISS buckets). The account tier clicks "Don't suggest posts from <user>" on the collapsed card — IG confirms with "We won't suggest posts from X", and an Undo stays available to the creator.
  • Followed rows: IG offers no "Not interested" on followed accounts (their menu shows Unfollow — never touch it). The helper returns no-entry harmlessly if tried; don't try. Flag repeat off-niche offenders in the report instead — unfollowing is the creator's call.
  • AD rows: no action, ever. The extractor flags them AD (Sponsored/Ad/Learn more text, no "Suggested for you" marker) and __feedDismiss refuses them with ix:ad.
  • NEUTRAL (in-niche but weak, ambiguous): no action. Don't punish the algorithm for almost-right guesses.

__feedDismiss return codes: ix:ok (post-level landed), ix:ok-account (account block landed too), ix:already (card was already collapsed), ix:no-item (no article at that index), ix:ad (refused), ix:no-btn (no three-dots control), ix:no-entry (menu had no "Not interested", followed post).

Outlier judgment: no VidIQ badges on this surface — use likes/comments vs how recognizable the account is. Big engagement from an unknown account = the signal; flag small sample when it rides on a tiny account.

Step 4 — Advance and repeat

__more() then __feed(<lastSeenIx+1>); classify only the new tail (indexes are stable — dismissed posts collapse in place, scrolling appends). Dismissals go one per javascript_tool call, never batched with other dismissals. Expect 20 keeps across a few hundred scanned posts on an untuned feed — and don't grind past ~60–80 articles to force the target; report what you have and note that tailoring compounds across sessions.

Pagination stall: the home feed can stop serving articles entirely (stuck at a handful even at page bottom) — that's IG throttling the feed API. Reload once and re-inject; if it persists, end the session early rather than hammering it, and say so in the report.

Step 5 — Report

Fill assets/report-template.html (this skill's folder); do NOT design from scratch. The template is the locked preset (same editorial dark "Algorithm Field Report" as yt-feed-research, text-only cards with linked usernames, kind tags reel/carousel/post, likes/comments badges). The HTML comment at the top documents every {{PLACEHOLDER}} and the card partial to duplicate. Write the filled file to research/IG-Feed-Research_YYYY-MM-DD_HHMMSS.html (timestamped so a second same-day session never overwrites the first) and open it in the default browser (macOS open, Linux xdg-open, Windows Git Bash start "" <file>). Never delete prior reports. Sections, in order (the summary strip is the header stat row + feed-state note):

  1. Summary strip — posts scanned, ideas saved (by kind: reels/carousels/posts), dismissals applied (post vs account level), the derived KEEP/DISMISS one-liner used this session (inside {{FEED_STATE}}), feed-state before/after one-liner, any rate-limit flag.
  2. 🚀 Top outliers — biggest engagement-vs-account-size keeps, one-line remix-angle note each.
  3. 📈 Strong picks — the rest of the keeps, badges only.
  4. 🎠 Carousel steals — carousel/image formats worth cloning via carousel-generator.
  5. 🤔 Borderline — logged but not counted toward the target.
  6. 🧹 Tailoring log — every dismissal (username + concept + post/account level) so the creator can veto via IG's Undo, plus flagged followed accounts.

Cards: linked username (instagram.com/<user>/), hook/caption excerpt, likes/comments badges, kind tag. No inline thumbnails — IG media URLs are signed, expire, and trip the safety filter. For visuals/transcripts of the winners, hand the usernames to ig-competitor-research or transcribe-url.

Gotchas (all observed live 2026-07-12 — don't rediscover)

  • "Not interested" exists ONLY on SUGGESTED home-feed posts. Followed posts and the reels viewer don't have it (reels menu = Report/Share/Copy/Embed only — that's why this skill doesn't scroll the reels tab; that viewer also hides carousels).
  • JS scrolling is the hardened advance path here (__more() sets scrollTop directly) — the home feed is a normal document scroll. Real wheel scrolls get eaten by carousel posts under the cursor; window.scrollTo with smooth behavior can silently no-op.
  • Menus are NOT role="dialog" — entries are found by scanning visible childless text leaves outside <article>. The account-level "Don't suggest posts from X" renders inside the collapsed article, not in an overlay.
  • JS clicks work for menus and dialog entries (no cold-start trap on IG, unlike YT). Close a stuck menu with a real computer Escape; the JS-dispatched one is unreliable.
  • Dialog text uses curly apostrophes ("Don't suggest…") that clean() strips — helpers match with don.?t regexes; preserve that when editing.
  • Ads render as articles without the "Suggested for you" marker but with "Sponsored"/"Ad"/"Learn more" text — never dismiss ads (opens ad-preference dialogs); scroll past. __feed flags them AD and __feedDismiss refuses them (ix:ad).
  • Like counts sometimes render as "Liked by X and N others" — the extractor matches both forms; "?" likes on followed posts is normal (counts hidden).
  • If Claude-in-Chrome errors "not connected", retry once — transient drops mid-session are normal and the in-flight page JS still completed.
  • First run of any variant is a calibration run: verify extraction on a few articles before the first dismissal, patch feed-tools.js against the live DOM, and log fixes here.