Instagram Saves → mindspace

Two phases, deliberately separate: ingest is mechanical (scripts), triage is judgment (you). Never let a script author an idea card — project, tags, and framing are decisions, not fields.

Why the export, not a cookie daemon

Instagram has no official saved-posts API. Every "auto-sync my saves" tool drives the private web API with a session cookie lifted from DevTools, which violates Instagram's ToS and risks the account — instagrapi's maintainers call it "fragile in production." The official export carries the same saves with no such risk. The trade is a manual request every week or two instead of a cron job. That trade was made deliberately; do not quietly "improve" this by adding cookie auth.

Phase 1 — Ingest

Requesting the export (the user does this; ~15 min to 48 h to arrive): Instagram → Accounts Center → Your information and permissions → Download your information → JSON format → Saved posts. The download link expires after 4 days.

Then:

# 1. Parse new saves out of the export (zip or unzipped dir)
python3 ~/.claude/skills/ig-saves/scripts/parse_export.py ~/Downloads/instagram-export.zip \
  > /tmp/saves.jsonl

# 2. Enrich each with caption + transcript (rate-limited on purpose)
bash ~/.claude/skills/ig-saves/scripts/enrich.sh --limit 20 < /tmp/saves.jsonl

Staged files land in ~/.cache/ig-saves/pending/<shortcode>.md — frontmatter plus caption and transcript. State lives in ~/.cache/ig-saves/state.json; a save already staged is skipped, and one already processed is never re-emitted.

Posts that are private, deleted, or rate-limited are staged with status: unavailable rather than silently dropped. Report them; don't retry in a loop.

Classification — the two axes

Every save gets both. idea_type says what it is (mindspace's existing enum); disposition says what happens to it. Disposition is the axis that decides where a card lands and when it is done.

Precedence ladder

Many saves match several shapes, so order decides. First match wins:

1. personal — non-work topic, whatever collection Instagram put it in. Instagram collections are assigned loosely; a hand-massage reel filed under "Interesting knowledge" is still personal. Judge the content, never the collection name.
2. expired — names a superseded tool or model, or saved_at is older than 18 months and the content is a tool list. Evergreen technique never expires on age alone.
3. adopt — the transcript is self-contained and it changes how you work.
4. evaluate — names a specific installable tool, repo, or MCP server.
5. queue — describes a concrete buildable artifact.
6. intel — pricing, offers, positioning, business model.
7. reference — the value is visual.
8. reconstruct — gated, and the payload is genuinely absent from the transcript.

Never trade a DM for the answer

Roughly a quarter of work-flavored saves are lead magnets: "comment X and I'll send it." Do not DM for these, and do not create a chase list.

reconstruct means: capture the claim the reel makes, then run it through our own research and evaluation. The reel is a prompt for investigation, not a source. The card records what was claimed, what we verified independently, and the verdict. If a technique is worth having, we can derive it — and we end up understanding it rather than owning someone's PDF.

reconstruct sits last on the ladder on purpose. A gated reel that explains its technique fully in the transcript is adopt or evaluate; the gate is irrelevant when the substance is already there. Only reels where the payload truly never arrives fall through.

Personal subcategories

personal saves carry a subcategory drawn from the topic, not from the Instagram collection name (those are inconsistent — "food", "Food", and "Food N Stuff" are the same thing). Normalize to: food, drinks, home, garden, travel, health, relationships, gifts, pets, games, kids, style, finance, career, misc.

finance and career cover personal money and personal job-search material — debt mechanics, asset protection, job-market reality. They are personal even when Instagram filed them under a business collection, and they are not intel: intel is only for material that informs a Summit or Rare Find offer.

Thin transcripts

About a quarter of reels carry their meaning on screen rather than in speech. Two distinct failure shapes, both measured on a real batch (35 of 124):

• Thin — transcript under 30 words. The caption usually carries the meaning.
• Music-only — whisper faithfully transcribes the backing track, so the card reads as song lyrics. Detect it: three or more of yeah|oh oh|la la|baby|woah in a transcript under 120 words. The transcript is worse than useless here because it looks like content while saying nothing about the reel.

Never classify a music-only reel from its transcript. Re-run it with --mode balanced and read the frames, or leave it unclassified. Anything heading for reference needs the frames pass regardless — that disposition is defined by visual value, so a transcript-only reference card is empty by construction.

Dedupe on content, not just shortcode

The same reel gets reposted by different accounts, so state.json (keyed on shortcode) will not catch it. Before creating cards, fingerprint the first ~25 words of each transcript and collapse matches — one card, with the other URLs listed as alternate sources. Measured rate on the first batch: 1 duplicate pair in 124, which scales to a meaningful number across ~1,100 saves.

Phase 2 — Triage into mindspace

Read ~/mindspace/AGENTS.md and _meta/schema.md first — they are authoritative and they change. Then for each staged file:

1. Read the caption and transcript. For a visual/design save, or when the transcript is thin, read the frames from its manifest: path — the reel's look is often the point, and text alone will miss it.
2. Decide project, idea_type, tags, priority, and suggested_action. Ask when a save is genuinely ambiguous rather than guessing a project.
3. Write the card to ~/mindspace/inbox/ as YYYY-MM-DD-HHMM-{slug}.md, matching the schema exactly: source: social, source_platform: instagram, source_author: <handle>, source_url: <post url>.
4. Put the faithful source material under ## Context / source — quote the caption and the transcript lines that carry the idea. That section is evidence, not summary.
5. Update _meta/index.json. Commit with a capture: prefix. No emojis.
6. Close the loop — never hand-edit the state JSON:

python3 ~/.claude/skills/ig-saves/scripts/mark_processed.py <shortcode> [<shortcode> ...]
python3 ~/.claude/skills/ig-saves/scripts/mark_processed.py --list   # what is still staged

priority: high auto-creates a Google Calendar event — be deliberate.

Phase 3 — Content angles (on request only)

When the user asks for angles on a card, append a block to the existing card rather than rewriting it, mirroring the <!-- research: ... --> convention already in use:

<!-- angles: 2026-08-23T14:00:00-04:00 -->
## Content angles
### Hooks
- three variations, each a different emotional entry point
### Outline
Hook -> key points -> CTA
### Platform notes
- Instagram / TikTok / YouTube differences worth the reshoot
<!-- /angles -->

Ground every angle in what the source actually did — cite timestamps from the transcript. Generic hook templates are worthless; the value is in why this post worked. Do not invent engagement numbers.

Do not

• Do not add cookie-based or private-API fetching (see above)
• Do not write directly to library/ — captures land in inbox/ and get triaged
• Do not let a script generate card summaries, tags, or projects
• Do not batch more than ~20 posts per run; the sleep between fetches is deliberate

Dependencies

python3, yt-dlp, and the watch-video skill (which brings ffmpeg + whisper).