antisearch-automation

Mill for antisearch.ai organic clips + publish (synthetic-oracle). Not the product repo.

A queue row goes in and a 9:16 video of about 5–8s + one metadata row come out. The clip recreates the antisearch.ai screen as a cold monolith: a blinking cursor, a bare phrase typed in (thing, dry shampoo), a still pause while the oracle thinks (the typed input and the caret, nothing added), then one lowercase verdict fetched live from POST https://antisearch.ai/verdict {"category": "thing"}. Nothing else: no captions, no logos, no motion design. Captions are GTM's job later (the queue's caption column).

Every row, log line and draft is tagged capture=synthetic-oracle. The pixels are synthetic (ffmpeg, not a screen recording); the verdict is real.

queue/sample_queue.csv ──produce──▶ POST antisearch.ai/verdict ──▶ ffmpeg monolith
                                                                    │
            out/clips/<clip_id>.mp4 / .png / .json  ◀───────────────┘
            out/clips.csv  (metadata row per clip)
                          │
                       publish (dry-run default)
                          ├─ x         v2 media upload + POST /2/tweets
                          ├─ linkedin  Videos API + Posts API
                          ├─ tiktok    draft json only (stub)
                          └─ instagram draft json only (stub)
            out/publish.csv, out/dry_run/<platform>/*.json, out/drafts/<platform>/*.json
            out/logs/mill.jsonl

Quick start

make install        # pip install -e '.[dev]'  (needs ffmpeg + a sans font, see SETUP.md)
make test           # unit + ffmpeg tests, no network, no tokens (live-oracle test opt-in: MILL_TEST_NETWORK=1)
make dry-run        # proof: fresh out-dry-run/, live oracle, publish dry-run, verify, contact sheets
make dry-run-offline  # same with the fixture oracle (no network)
make doctor         # tools, font, live oracle ping, which credentials are missing

Credentials: see SETUP.md, a single checklist of everything to add before --live. Nothing blocks without them.

Own publisher (mill post)

One adapter per network behind post(media, caption) -> url (src/antisearch_mill/publisher/). Dry-run unless --live.

• Routes: native (each network's official API) everywhere by default: Bluesky, Instagram Reels, LinkedIn Page, TikTok. PUBLISHER_BRIDGE=buffer is a removable bridge through Buffer free while a native gate is pending. Threads, Reddit and Mastodon are built too.
• Cadence: per-network slots in America/New_York on fixed :07 minutes, ≥90 min apart. The day starts 09:00 ET; ≤5 per network per day, one per run. PUBLISHER_CADENCE=carlos|soul|reach switches presets.
• Fill: queue rows first, then (optional) one trend row per network per day, then queue/pools/; unowned words open each wall and come back one in four.
• Lanes: mill winners (inputs likely to crown an entity; the only text is its human-verified @handle, ≤2 a day), mill trends (a timely noun input from free sources, strictly filtered) and mill replies (Bluesky, a verdict-only answer to "what's the … X?", ≤3 a day). See docs/TREND_AND_REPLY_LANES.md.
• Guard: one clip → one network; an input never repeats; inputs English and never serial; caption = empty / bare input / @handle (Bluesky: empty or a verified domain handle); live only for networks in PUBLISHER_LIVE_NETWORKS.
• One honest run: only --live asks the real oracle, once per input; the verdict is frozen in verdicts.csv next to the post log. Dry-runs never call it.
• Post log: a pending row is written before every API call, so nothing double-posts.
• Scheduling: .github/workflows/publisher.yml (manual + cron, off by default) runs it without the Grok Bot box.

mill networks                        # routes, slots (ET + Madrid), credentials, pool runway
mill due                             # exit 0 if a live network has a slot open now (cron pre-check)
mill post --queue queue/bluesky_trial.csv --queue queue/tiktok_trial.csv   # dry-run plans, no oracle call
mill post --queue queue/bluesky_trial.csv --live --state state/posts.csv

Setup, exact env vars and the honest limits per network: docs/PUBLISHER_SETUP.md. mill publish (below) is the older X/LinkedIn path and stays as it was.

Manual one-off posts (mill manual-post)

.github/workflows/manual-post.yml (Run workflow; dry-run unless live) posts the rows of a reviewed file, queue/manual/*.csv with id,category,network,due_et, through the normal publisher and Buffer path. The row goes out once even if its network is paused, only with a verdict already frozen in publisher-state/verdicts.csv, scheduled for due_et in New York (or the same time the next day if passed), and its posts.csv row records caption and format.

settle it thread list (mill settle-find, settle-shoot)

Founder, 2026-09-29: Carlos posts on Reddit himself, so settle it is a TEXT-ONLY list. There are no images, GIFs, video or captions; the old media were removed and are not to be regenerated. Everything is scripted, with no LLM and no agent, and nothing is ever posted.

The settle-pack workflow runs every 2 days at 19:00 Madrid (or by hand) and writes publisher-state/settle/threads-YYYY-MM-DD.md, with up to 20 rows of thread | input | output.

1. Find: live "what's the best X?" threads from Reddit RSS and Bluesky search, under 48 h old. Soul's skips are applied by keyword (grief, health, legal or money, minors, private people, politics, NSFW). Each question is reduced to its bare noun phrase by fixed rules: "What's the best 80s horror movie?" becomes 80s horror movie. If no clean noun is left (a "way to…" or "one" question, or "pizza or pasta"), the thread is skipped.
2. Shoot: ONE live run per row through antisearch.ai in headless Chrome. The User-Agent ends antisearch-internal/settle-pack and the request carries X-Antisearch-Internal: settle-pack. The output is frozen in verdicts.csv and the call is logged in settle/internal-calls.csv. A no (or no verdict) drops the row. An input frozen before is never asked again.

Locally, with a publisher-state checkout in ./state:

mill settle-find --state state/posts.csv      # the list, outputs pending
pip install -e '.[settle]'                    # once: Playwright for shoot
mill settle-shoot --state state/posts.csv     # one live run per pending row

Metrics (mill metrics)

.github/workflows/metrics.yml runs every 3 hours (and on demand) and appends per-post numbers to the publisher-state branch, read-only: it never logs in, drives a browser, or posts.

• metrics/posts_metrics.csv: ts_utc, network, post_id, url, input, verdict, slug, posted_at, format, views, likes, comments, shares, saves, source, join. slug is the posts.csv job id; format (video, text, image) comes from posts.csv, blank legacy rows being videos.
• metrics/accounts.csv: ts_utc, network, handle, followers, likes_total, posts, source, note.
• metrics/summary.md: top posts by views and by engagement rate per network, last 24h and 7d, follower deltas, and the X format A/B (video vs text) over 7 days.
• Sources: Buffer post metrics once a day when the key exposes them (trusted only once Buffer has refreshed them; its default 0s are ignored); TikTok's public profile, embed profile, and video pages; X's public embed endpoint (likes, replies). Instagram's public pages have no counts when logged out, so Instagram depends on Buffer. X followers and X views are not public and stay blank.
• A number that no source reported is blank, never 0. A network with published posts that returns nothing turns the run red, with the reason in the job summary.
• Vars: BUFFER_ORG_ID (optional), METRICS_TIKTOK_HANDLE (default antisearch).

CLI

mill (or PYTHONPATH=src python3 -m antisearch_mill):

command	what it does
mill produce [--queue q.csv] [--oracle live|fixture] [--limit N] [--only id] [--force]	ask the oracle, render, validate, write metadata
mill publish [--live] [--platform x] [--clip id]	publish produced clips; dry-run unless --live
mill status	clips and the latest publish state per platform
mill verify	re-probe every clip against the contract (size, length, codec, sha256, tag)
mill sheet	contact-sheet PNGs into out/proof/
mill doctor [--network]	environment + credential check (never prints values)
mill oracle "dry shampoo"	one verdict, as JSON

Global: --out DIR (default $MILL_OUT_DIR or out/), --env-file .env.

Queue

queue/sample_queue.csv:

column
id	stable job id, [a-z0-9_-], unique. Produce skips ids that already have a clip
category	the bare phrase that gets "typed" and POSTed as-is; lowercased, ≤55 chars (the product's input maxlength). See Input shape
platforms	x|linkedin|tiktok|instagram (empty = all)
caption	optional post text (GTM). Empty = video-only post
notes	free text, ignored

Input shape

"best" is the oracle's hidden premise, never the user's input. The user types only the thing and the oracle returns the single best value for it: thing → love. The clip, the queue and the POST body carry the bare phrase; nothing in the pipeline prepends, appends or templates "best" into it.

The category field/column name is historical; its value is any bare input.

• Never contains "best", "top", "greatest", "what's the" or any superlative ("finest", "ultimate", "most", "biggest", "#1", "of all time"...), anywhere in the phrase (founder, 2026-09-29: this is the root error). The premise is invisible; the user reads the pattern of excellence from the results. Every lane runs bare_input.superlative_problem before an input is used, and both oracles refuse a ranked input.
• Never a question ("what is the best…", trailing "?"). Another person's question is reduced to its bare noun phrase first: "What's the best horror movie of the 80s?" becomes 80s horror movie.
• No "-est" superlative and no "#1" in inputs, captions or tags (Soul, 2026-09-29).
• Never an ask ("recommend…", "give me…", "find me…").
• Never a ranking ("top 3…", "…ranked").
• ≤ 55 characters. Constraints are fine as plain phrases: restaurant in barcelona, gift for a new mom.

Correct: thing · dry shampoo · restaurant in barcelona · pizza · federico valverde · running shoes · emotion · gift for a new mom

Wrong: best running shoes · best dry shampoo · best software · best country for having fun · what is the best espresso machine?

tests/test_input_shape.py enforces this on every queue row, every fixture key, and every code-formatted example in README/SETUP.

Clip contract

• 1080×1920 (9:16), 30 fps, H.264 High yuv420p + silent AAC stereo track, +faststart
• about 5–8s. The old ≤4s cap is lifted (founder, 2026-09-25). LinkedIn and Reels still reject clips under 3s; the visible wait is why these run longer
• the product CSS evaluated on a 360×640 CSS-px phone at 3×: #0A0A0A background, #E8E4D8 text, 18px→54px sans (Inter), centered at 38% height, 0.6ch × 1.05em block cursor blinking at 1.05s, dim antisearch.ai footer
• the blink phase restarts on every keystroke and when the pause begins
• timeline: 0.3s idle cursor → jittered typing (≤1.5s, seeded by category, so deterministic) → 0.22s hold of the completed input → a still pause (the typed input and a blinking caret only; no spinner, no dots, no "thinking" text, even if antisearch.ai shows one) → verdict holds to the end (≥1s). Short clips are padded on the verdict hold so the total is at least 5s. The pause length is the oracle call's wall-clock latency, clamped to 1.2–3.5s. If that latency was not recorded, the pause is 2.0s
• the pause does not copy the site's loading chrome. The site fetched 2026-09-25 hides the typed line on submit. The word clip keeps the input anyway. A later spinner, dots, or status word on the site is not drawn
• PUBLISHER_CLIP_STYLE=app withdraws that word-stays lock (2026-09-26, Soul + Carlos) and records the live screen: keyboard in, human typing, the finished word held ≥0.5s, Enter clears the word and slides the keyboard out while the line moves once to 38% of the full height, the caret blinks alone for the real oracle latency, the verdict holds ≥1.5s. Target ≤4s, hard cap 5s; a long latency is not cut. Unset or word is the default above and is what already-scheduled posts keep using
• PUBLISHER_CLIP_AUDIO=off (default) keeps the silent AAC track. pulse muxes a procedural felt-piano bed into Instagram and TikTok only: the same C3 note on each caret blink during the real pause, dry, no pedal, short room. The tail is silent before the clip ends. PUBLISHER_CLIP_AUDIO_VARIANT defaults to c (blinks, then silence). a resolves to an open fifth with octave one register below, b to the same note an octave lower. X, LinkedIn, and Bluesky stay silent
• the template (founder, 2026-09-26). Deterministic code only: no model at render time, no licensed audio or samples. Every flag defaults off, so scheduled posts do not change:
  • PUBLISHER_CLIP_FRAME=phone plays the app screen inside a generic phone drawn in code (thin bezels, rounded corners, a centered punch-hole camera, side buttons; no logo, no notch, no status bar, no captions). PUBLISHER_CLIP_PHONE_SCALE is the phone height over the video height (default 0.80, allowed 0.60–0.95). PUBLISHER_CLIP_PHONE_BG is a calm surface lighter than the phone, so it reads at a glance: stone (mid warm grey, default), bone (off-white), or sage (muted grey-green). black and gradient (very dark grey) are kept but the phone disappears on them. The phone is a near-black slab with a graphite rim and a black bezel; the app screen is unchanged. none is the default
  • PUBLISHER_CLIP_AUDIO=heartbeat (Instagram and TikTok only): a synthesized mechanical key click on the frame each letter appears, a heavier Enter, then a low lub-dub on every caret blink of the wait. The verdict is silence (variant c, locked). Seeded from the input, so the same input gives the same audio. Sample peak -3 dBFS, true peak under -1 dBTP. PUBLISHER_CLIP_AUDIO_BED=on adds a very quiet low drone (about 20 LU under the beat) under the wait only; off is the default
  • PUBLISHER_CLIP_AUDIO=keys (Instagram and TikTok only; founder, second pass): the only sound is typing and Enter, one click per letter on the frame it appears, about -22.5 dBFS peak per key, Enter about -20.8 dBFS. Exact silence from the first wait frame through the verdict. PUBLISHER_CLIP_KEYS_SOUND (founder, 2026-09-27: a phone tap, not a laptop key): tap (default: a dry 1.6 kHz tock, about 7 ms), tick (a fine 3.2 kHz tick) or glass (two pure partials, no noise). The earlier laptop keys soft and muted still work. Enter is the same voice, a little lower and longer. Every voice is original synthesis in audio.py: no recording, sample pack, or third-party sound, so no licence applies. heartbeat and pulse are kept for back-compat
  • PUBLISHER_CLIP_TYPING sets the fixed-wait typing rhythm: human (default: 70 + 550/len ms per key with ±25% jitter and a beat after each space, so thing types in about 0.7s, and the old 1.5s cap, so long inputs keep today's length), steady (an even curve, capped at 1.6s) or legacy (every key about 65 ms). One frame per key is the floor
  • with the phone frame, the keys, or the heartbeat, the wait is fixed: WAIT_BLINKS = 2 caret blinks of 1.05s (blink, blink, verdict), 63 frames (2.1s) at 30 fps, in every clip. It overrides the oracle latency. Only the typing and the verdict can vary with length. Segment frames are in the clip's sidecar JSON (timeline.segments_frames)
• verdict is lowercased and whitespace-collapsed; wraps like the product if long
• the verdict frame is the cover
• renders are byte-identical for the same (category, verdict, pause, style, audio, frame)
• mp4 metadata: title=<verdict>, comment=capture=synthetic-oracle;oracle=<live|fixture>;category=<...>

LinkedIn, TikTok and Instagram get the 1080×1920 master. X is plain text only (founder, 2026-09-29): the lowercase input on line 1, the lowercase verdict on line 2, and the winner's verified @handle on line 3 when one applies. There is no media on X, two posts a day.

Metadata row (out/clips.csv)

clip_id, job_id, category, verdict, capture, oracle_source, oracle_url, oracle_latency_ms, oracle_fetched_at, produced_at, clip_path, poster_path, duration_s, width, height, fps, size_bytes, sha256, font, render_style, mill_version, platforms, caption

clip_id = <job_id>-<sha8(category, verdict, oracle_source, render_style, pause)>. The oracle is not deterministic (the same category can return a different pick on a later call), so the row freezes the verdict that was actually shown.

Oracle

Retry policy mirrors the product frontend: one 20s budget, ≤2 attempts, retry only on 5xx / network / timeout, never on 4xx. On 429 the mill waits Retry-After (or 60s) once, then gives up on that job. Calls are spaced ≥1.5s. A failed verdict fails that job only; it is logged (oracle.failed) and the run continues.

--oracle fixture (fixtures/oracle_fixtures.json) exists for tests and no-network runs. Fixture verdicts are marked oracle_source=fixture and the publisher refuses to post them live.

Publish modes

mode	x / linkedin	tiktok / instagram
dry-run (default)	full upload flow runs against a simulated API; every request (auth redacted) is written to out/dry_run/<platform>/<clip_id>.json; dry_run_ok	draft json, draft_logged
--live	real API; needs credentials, else skipped_missing_credentials; idempotent per (clip, platform)	still draft json only

Publish statuses: dry_run_ok, published, draft_logged, skipped_missing_credentials, skipped_not_live_oracle, skipped_already_published, invalid_clip, failed.

Layout

src/antisearch_mill/
  oracle.py      live + fixture verdict oracles
  timeline.py    pure frame timeline (typing → pause → verdict)
  render.py      product-CSS layout → ffmpeg drawtext/drawbox graph, probe, poster, sheets
  fonts.py       font resolution + HarfBuzz-shaped metrics (Pillow/raqm)
  produce.py     queue → verdict → clip → metadata row
  ledger.py      queue CSV, clips.csv, publish.csv, mill.jsonl
  publish/       x.py, linkedin.py, stubs.py (tiktok, instagram), http.py (live/dry-run transport)
  cli.py
tests/           unit, ffmpeg render, sandboxed X/LinkedIn flows, CLI e2e
.github/workflows/ci.yml    tests + offline dry-run on every PR
.github/workflows/mill.yml  manual mill run (dry-run unless `live` ticked)