aeo-daily

Local AEO daily content engine — FORKED VOICE + voice filter (v0.4.2). One command → today's pack across 5 BUSINESS channels by default (was 11). Travel pack on-demand via --voice flag. BUSINESS voice

SKILL.md

GitHub

/aeo-daily v0.6 — Local Daily AEO Content Engine

Ships 1 LinkedIn post (voice-clean) + 8-slide carousel + comment-bait reply pack + GHL CSV row every day.

Source repo: <repo>

WEDGE (locked 2026-06-08): ONE thing this engine sells — "Get Cited": make SMBs the answer ChatGPT / Perplexity / Gemini gives. Every BUSINESS post points at that. TRAVEL voice is NOT a second business — it is the proof-of-life layer of the same wedge (solo operator running an AI agency from a local cafe). Default --voice business. Travel runs 3x/week max on personal IG/TikTok only, never competing for the business CTA, with at most ONE subtle bridge line/week ("built this from a local cafe").

FUNNEL (locked 2026-06-08): content is not the product, leads are. Every business post carries a funnel block {cta_keyword, lead_magnet_ref, booking_link}. Flow: reel/post -> comment keyword (e.g. CITED) -> DM auto-reply -> free AEO Visibility Snapshot (scripts/snapshot_gen.py) -> book call. Snapshot = lead magnet + demo + close in one. CTA bank: funnel/cta-bank.json. DM script: funnel/dm-flow.md. Full spec: funnel/README.md.


CHANGELOG

  • v0.6.0 (2026-06-08): WEDGE + FUNNEL + engine truth-up (4-agent fanout). (1) Wedge locked = "Get Cited" AEO-for-SMBs; travel demoted to 3x/wk proof-of-life layer (see WEDGE callout above). (2) Funnel built — scripts/snapshot_gen.py (AEO Visibility Snapshot lead magnet, client-presentable cream/purple HTML), funnel/cta-bank.json (10 one-word keyword CTAs, zero "DM me"), funnel/dm-flow.md, funnel/README.md. copy.json now carries a funnel block. (3) Missing scripts now REAL — the v0.5.0 vaporware is built: config/scroll-hooks.json (150 hooks / 15 triggers, ___ blanks, _meta.angle_to_category), scripts/suggest_hooks.py, scripts/export_text.py, scripts/export_variants.py, scripts/build_master_html.py. (4) lint_voice.py hardened (v1.2.0) — now enforces ai_tell_structures (negation-parallelism / rule-of-three / balanced-pair-warn / framework-promise), scans ALL business channels (was linkedin.post only), added U+2018 + funnel.cta_keyword rule (fail under --strict, warn without), UTF-8 stdout fix. (5) Engagement loop live (true v0.6) — analytics/weekly.csv + scripts/weight_hooks.py -> config/hook-weights.json; suggest_hooks.py reorders an angle's categories by learned cta_trigger weight. (6) client-wins.json refreshed (+5 real wins through 06-07). Canonical psych-category vocab = the 15 hyphen-named categories in scroll-hooks.json (weekly.csv must use them).
  • v0.5.0 (2026-06-07): HOOK LIBRARY wired in. New config/scroll-hooks.json = 150 scroll-stopper hooks across 15 psychology triggers (curiosity-gap, loss-aversion, contrarian, identity-callout, transformation, secret, authority, urgency, shock, question, specificity, story-loop, problem-agitate, confession, benefit), each with ___ blank for any topic. New scripts/suggest_hooks.py maps weekday angle → best psych categories and prints N filled hooks (deterministic, --offset to rotate). New MANDATORY PHASE 2.5 — engine ALWAYS surfaces 8 matched hooks before writing copy; LI hook + video 0-4s beat must be built from them. Browsable viewer HOOK-LIBRARY.html (live topic-fill + filter + search + copy).
  • v0.4.2 (2026-05-20): TRAVEL SOFT-DISABLED by default + MASTER.html + image v2.
    • build_pack_v2.py new --voice business|travel|all flag (default business). Travel platforms skipped from render loop unless explicitly requested. Travel code path preserved — flip to --voice all to resume both tracks. run-aeo-daily.bat now passes --voice business for the 07:00 local-time auto-build.
    • scripts/build_master_html.py new — fans every channel pack into ONE MASTER.html (iframes + sticky TOC + angle header + GHL CSV download block). Auto-invoked by build_pack_v2.py right after INDEX.html. Voice filter respected → travel iframes simply don't exist when skipped.
    • INDEX.html template no longer renders empty "Travel Voice" section when 0 travel cards present.
    • process_image.py v2 — face-detect now runs on ANY filename NOT starting DJI_ (was: only IMG_*, missed all SCENE-* portraits). Multi-cascade union (frontal_default + frontal_alt2 + profileface with horizontal flip). CLAHE histogram equalize before detect → recovers faces in shaded/backlit outdoor frames. Headroom guard in _crop_to shifts crop DOWN when face would clip top edge (kills "forehead at hairline" bug). Per-ratio focal-Y lift — 1×1 → 0.40, 4×5 → 0.38, 9×16 → 0.32 (face lands in upper third for reels). --debug flag for verbose face-detect logging.
  • v0.4.1 (2026-05-19 PM): Round-3 production patch. (a) build_ghl_csvs.py is now schema-aware: tries nested copy[track][slug] first, falls back to legacy flat keys → 4 travel CSVs (ig_travel_1/2, tt_travel_1/2) no longer SKIPPED on extended schema. New --copy flag accepts override path. Image refs sourced from _picks_log.json (Track C source-of-truth) with avatar-dir glob fallback. (b) config/hook-bank.json::image_picker_hints now holds real filename PATTERNS (glob) + literal stems pulled from rotation pool, never invented filenames. (c) build_pack_v2._resolve_image_hint supports glob matching (DJI_*, SCENE-2026-*-street-*) in addition to exact/prefix. (d) copy.image_picks MUST reference real stems or glob patterns — see PHASE 3 note below.
  • v0.4.0 (2026-05-19): FORKED VOICE schema. copy.json now contains shared{} + business{linkedin,fbpage,ig_pro,yt,tt_pro} + travel{fb_profile,ig_travel_1,ig_travel_2,tt_travel_1,tt_travel_2,pinterest} = 11 per-platform variants written in one pass. New travel_rotation 7-day hook bank added to config/hook-bank.json (cafe-discovery / visa-life / expat-mistake / scooter-day / food-find / builder-life / week-recap-travel). lint_voice.py --voice business|travel|both enforces both tracks independently. Travel voice hard-blocks AEO/agency/client/SkynetLabs/schema/n8n/MRR/Loom/Fiverr terms + DM-me CTAs. Image picker hint per voice — business uses IMG_ portraits, travel uses DJI__ + scene_* scenes. Schema doc at config/copy-schema.json.
  • v0.3.0 (2026-05-18): brand palette swap (drop pure-black → cream base + brand purple #6b3fa0 + cyan #2dc6d6 + coral #ff6b5b + gold detail), 2→4 photo picks (portrait priority — IMG** over DJI**), photo prominence (hook bg duotone, context top strip, data 2-photo grid, CTA gold-bordered portrait), per-role slide styling (hook/context/data/confession/fix/cta classes).
  • v0.2.0 (2026-05-18): 5→8 slide carousel (research: 60% dwell gap), voice-lint script (60+ AI tells blocked), client-wins + local-anchor + banned-phrases configs, hook variants A/B per day, hashtag tail required, comment-bait reply pack (first 60 min), founder-voice + local authenticity layer, arrow glyph fix (→ not ▸).
  • v0.1.0 (2026-05-18): initial scaffold + 5-slide pipeline.

VOICE FORK — v0.4 (READ FIRST)

The engine now writes 11 per-platform variants in 2 voices in a single pass.

BUSINESS voice (5 channels: linkedin, fb_page, ig_pro, yt, tt_pro)

  • Voice spine = locked LinkedIn template (contrarian hook → reframe → arrow lines → vulnerability beat → CTA).
  • Anchor: local-life PUNCH (one specific detail dropped mid-post).
  • Client win mandatory (shared.client_win_ref from config/client-wins.json).
  • Hashtags: business pool (#aeo #linkedinmarketing #agencyowner #automation #brandtracking #claudecode #n8n #b2b).
  • Hook bank: config/hook-bank.json → rotation[<weekday>].
  • Image picker: IMG_* portraits (face-on-camera).
  • Lint: python scripts/lint_voice.py --date <DATE> --voice business --strict.

TRAVEL voice (6 channels: fb_profile, ig_travel_1, ig_travel_2, tt_travel_1, tt_travel_2, pinterest)

  • NO AEO, NO agency, NO client pitches, NO SkynetLabs, NO schema/citation/n8n/MRR/B2B talk.
  • Self-deprecating, observational. Local daily-life (cafes, scooters, beach, expat scene, food, local-currency prices).
  • Anchor: local-life DRIFT (one detail per post, plus 1 named place + 1 real currency amount).
  • Story anchor mandatory (travel.fb_profile.story_anchor) = small embarrassing/funny moment.
  • Hashtags: travel pool (#travel #digitalnomad #nomadlife #expatlife #cafehopping #scooterlife #slowtravel). Blocklist: any business tag.
  • Hook bank: config/hook-bank.json → travel_rotation[<weekday>].
  • Image picker: DJI** + scene**.jpg + IMG_4*** scenes (NOT portraits).
  • Lint: python scripts/lint_voice.py --date <DATE> --voice travel --strict.
  • Banned CTA patterns: "DM me", "I'll reply with X", "follow for more", "save this".

Travel rotation (locked v0.4)

Day Angle Sample hook
Mon cafe-discovery "New cafe in town. Flat white was cheap. Wifi held for 2 hours."
Tue visa-life "Visa run on Tuesday. Agent quoted the 60-day rate. Standard."
Wed expat-mistake "Rented a scooter without checking the brakes. Found out at the first downhill."
Thu scooter-day "Cross-town on the scooter today. 90 minutes if you skip the toll."
Fri food-find "Local plate for a few bucks at a spot behind the apartment. Extra sauce added."
Sat builder-life "Saturday looks like: wifi-cafe morning, beach walk noon, local dinner."
Sun week-recap-travel "Week recap: 1 visa run, 2 new cafes, 1 wipeout learning to surf."

Lint both tracks in one call

python scripts/lint_voice.py --date <DATE> --voice both --strict

VOICE — LOCKED (the author's reference posts, do not flatten)

Reference POST 1 — "Freelancer → Operator" reframe (his top-engaging)

Stop calling yourself a freelancer. Words shape ceilings. Freelancer says: "I trade hours for cash." Operator says: "I run systems that deliver outcomes." The market pays operators 10x more. Same skills. Different label. Different price. → Stop saying "rate" → start saying "engagement" → Stop saying "available hours" → start saying "client capacity" → Stop sending invoices → start sending retainer agreements → Stop chasing leads → start filtering inbound What's on your LinkedIn headline today? Calling myself out first — mine literally said "freelancer" 12 months ago.

Reference POST 2 — "Claude Routines killed my n8n flows"

I spent 2 years building automations on n8n. Last week, Claude Routines replaced three of them in under 10 minutes. No nodes. No webhooks. No 2 AM error emails. Just a text box and a schedule. Here's one I'm running daily at 9 AM: → Pulls last 24-48 hrs of news across AI automation, n8n, Claude API, Next.js, Meta Ads → Filters for what's actually relevant to my agency and content pipeline → Writes 30-45 sec vertical video script — hook, body, CTA — ready to film I wake up. Script is waiting. I shoot. I post. A genuine ask to the builders here: What's the most creative Routine you've deployed?

Voice spine (every post)

  1. Contrarian hook — 1 line, ≤210 chars, ≤10 words ideal (hits 2.3x engagement per 2026 LI data).
  2. Reframe — 2-3 short lines, fragments OK.
  3. Arrow lines (→) — 3-5 items (keep glyph per voice memory — do NOT use ▸ ▪ • or em-dash).
  4. Personal vulnerability beat — admit specific past mistake, dated ("12 months ago" / "last quarter" / "the dental call last week").
  5. CTA question + compile-promise — "drop X below, I'll reply with Y."
  6. Hashtag tail — 5-8 tags (#aeo #linkedinmarketing etc).

Anti-AI authenticity layer (NEW v0.2 — required)

  • 1 anchor per post — a named local cafe, a ride-hail detail, a local-currency amount, a power-cut, a surf/commute time, or another concrete detail from your city.
  • 1 real client win referenced — slug from config/client-wins.json (dental-flagship / real-estate-voice / saas-email / healthcare-workflow / security-av / travel-client-site).
  • Allowed local code-switch sparingly: 1-2 colloquial words from your own dialect (max 1 per post, italicized).
  • Mid-post drift allowed — "Anyway." / "Scratch that, actually —" / trail-off ellipsis (... straight, not …).

BANNED PATTERNS (lint_voice.py hard-fails)

Full list: config/banned-phrases.json. Highlights:

Phrases: delve, moreover, navigate, leverage, robust, seamless, holistic, transformative, groundbreaking, pivotal, evolving landscape, in essence, at its core, the real question is, embark on, foster, harness, "let that sink in," "read that again."

Openers BANNED: "Unpopular opinion:" / "Hot take:" / "Plot twist:" / "Real talk:"

Closers BANNED: "What's your take?" / "Agree?" / "Save this post" / "Bookmark this"

Structures BANNED:

  • Negation parallelism — "It's not X, it's Y"
  • Forced rule-of-three — "innovation, inspiration, insight"
  • Balanced sentence pairs — "Talent gets you in. Discipline keeps you there."
  • Unnamed framework promise — "Here's the 3-step framework" (OK if named after real client)

Encoding BANNED: em-dash (—), en-dash (–), curly quotes (" " '), unicode ellipsis (…). Use straight ASCII.

Travel cliches BANNED: paradise, island vibes, manifesting, sunset views, nomad lifestyle.


EXECUTION FLOW (one session, ~6 min)

PHASE 1 — Fetch + dedup feeds (30s)

cd "<repo>" && \
python scripts/fetch_feeds.py --date $(date +%Y-%m-%d) --top 30

20 RSS, 48h SQLite dedup, writes outputs/<DATE>/today_items.json.

PHASE 2 — Pick today's angle (5s)

python scripts/pick_angle.py --date $(date +%Y-%m-%d)

Weekday → angle.json w/ 2 hook variants + top 5 candidates.

PHASE 2.5 — Surface scroll-stopper hooks (MANDATORY, 3s)

Always show the user matched psychology hooks BEFORE writing copy. Pull from the 150-hook library, filled with today's topic noun.

python scripts/suggest_hooks.py --angle <angle_slug> --topic "<topic noun>" --n 8
  • <angle_slug> = angle.json::angle_slug. The script maps angle → best psych categories via config/scroll-hooks.json::_meta.angle_to_category.
  • <topic noun> = the concrete subject of today's post (the niche/service/pain, e.g. "editing video in the terminal", "AEO for clinics").
  • Print the 8 hooks to the user. The LinkedIn hook + the video 0-4s beat MUST be built from one of these (paraphrase tightly to keep voice — never paste verbatim).
  • Full library viewer for browsing/filtering: HOOK-LIBRARY.html (live topic-fill box + search + copy). Dump all 150 grouped: python scripts/suggest_hooks.py --all --topic "<noun>".
  • Hook rule: first line ONLY. Pay off the open loop within the next 2 lines or it trains distrust. ASCII only (no em-dash / curly quotes).

PHASE 3 — Claude writes copy.json (~5 min, v0.4 FORKED schema)

Read: angle.json, config/client-wins.json, config/local-anchors.json, config/banned-phrases.json, config/hook-bank.json (read BOTH rotation for business AND travel_rotation for travel), config/copy-schema.json.

Pick once per day:

  • 1 source from candidates (most agency-relevant) → business block only.
  • 1 business hook variant from hook-bank.rotation[<weekday>] (or tight paraphrase).
  • 1 travel hook variant from hook-bank.travel_rotation[<weekday>].
  • 1 client_win_slug (shared.client_win_ref).
  • 1 local anchor (shared.local_anchor) — STYLED DIFFERENTLY in each track (business = punch / travel = drift).
  • 1 vulnerability_beat (shared.vulnerability_beat) — used by business only.
  • 1 story_anchor (travel.fb_profile.story_anchor) — used by travel only.

Write outputs/<DATE>/copy.json with the forked schema:

{
  "date": "2026-05-19",
  "angle_slug": "prompt-exposure",
  "weekday": "tuesday",
  "source_url": "https://...",
  "source_title": "...",
  "hook_variant_chosen": 2,
  "shared": {
    "news_anchor": "...",
    "client_win_ref": "dental-flagship",
    "local_anchor": "Coffee at the corner cafe, a few bucks.",
    "vulnerability_beat": "I used to think AEO was schema + FAQ. Cost me 2 closes in Q1."
  },
  "business": {
    "linkedin": {"post": "<900-1200ch w/ 8-role spine>", "slides": [...8 slides...], "hashtags": [...5-8...], "comment_bait": [...3...]},
    "fb_page": {"cards": [...3...], "post": "<200-800ch>", "hashtags": [...3-5...]},
    "ig_pro": {"slides": [...8...], "caption": "...", "hashtags": [...5-8...], "reel_script": "30-45s"},
    "yt": {"thumb": {"headline":"...","sub":"..."}, "shorts_script": "45-60s", "title": "<=70ch", "desc": "..."},
    "tt_pro": {"cover": {"headline":"...","sub":"..."}, "script": "30-45s", "caption": "...", "hashtags": [...3-5...]}
  },
  "travel": {
    "angle_slug": "visa-life",
    "hook_variant_chosen": 1,
    "story_anchor": "Stood in the immigration line 90 min. Agent took 20 USD over WA.",
    "fb_profile": {"cards": [...3...], "post": "<150-700ch>", "story_anchor": "..."},
    "ig_travel_1": {"slides": [...5...], "caption": "...", "hashtags": [...5-8...], "reel_script": "15-30s"},
    "ig_travel_2": {"slides": [...5...], "caption": "...", "hashtags": [...5-8...]},
    "tt_travel_1": {"script": "...", "caption": "...", "hashtags": [...3-5...]},
    "tt_travel_2": {"script": "...", "caption": "...", "hashtags": [...3-5...]},
    "pinterest": {"pin": {"headline":"...","overlay":"...","size":"1000x1500"}, "desc": "...", "board": "Travel", "hashtags": [...8-12...]}
  }
}

Full schema spec: config/copy-schema.json.

image_picks block — MUST use real filenames OR glob patterns (v0.4.1)

Authoring image_picks is the #1 cause of "PENDING_IMAGES" CSV rows. The hint MUST be one of:

  1. A literal stem present in the rotation pool — verify before writing:

    python "<your-images>/_pick-next.py" --count 400 | grep -E "^DJI_|^SCENE-|^IMG_"
    

    Examples that resolve today: IMG_4105, SCENE-2026-05-07-garden-variant, DJI_20260518_0023.

  2. A glob pattern with * or ? — build_pack_v2._resolve_image_hint runs fnmatch against every candidate stem (case-insensitive). The FIRST match wins, which means the rotation-log order decides which actual file you get.

    • Travel scenes: DJI_*, SCENE-2026-*-street-*, SCENE-2026-*-garden-*.
    • Business portraits: IMG_*, SCENE-*-rooftop-portrait-*, SCENE-*-veranda-*.
  3. NEVER invent filenames like scene_street_01.jpg or scene_beach_low_tide unless they actually exist on disk. They will not resolve, drop to fallback (whatever-track-appropriate is available), and you lose authorial control over which image pairs with which slide.

Pattern cheat-sheet lives in config/hook-bank.json::image_picker_hints. Add new globs there as new naming families enter the pool. NEVER edit process_image.py for this — the resolver is in build_pack_v2.py.

PHASE 3.5 — Voice lint, BOTH tracks (3s, HARD GATE)

python scripts/lint_voice.py --date $(date +%Y-%m-%d) --voice both --strict

Business fails on ANY: banned phrase, encoding tell, hook>210 chars, hashtag count off, missing vulnerability beat, missing anchor, missing client_win_ref, wrong slide count/roles.

Travel fails on ANY: banned business term (aeo/agency/client/SkynetLabs/n8n/MRR/Loom/Fiverr/schema/citation), blocked travel hashtag, missing local-currency anchor, missing named place, missing story_anchor, business-CTA pattern ("DM me" / "I'll reply with"), encoding tell.

If lint fails → fix copy.json + re-run BEFORE Phase 4.

PHASE 4 — Export TEXT-ONLY pack (5s) — DEFAULT since 2026-05-31

Preferred: 3 variants per platform (story / contrarian / confession-style angles for A/B). Author outputs/<DATE>/variants.json (see shape in scripts/export_variants.py header — platforms[].variants[] each = {style, angle, body, hashtags}), then:

python scripts/export_variants.py --date $(date +%Y-%m-%d)

Emits VARIANTS.html (bright theme, 3 labelled copy-boxes per platform + Copy button) + VARIANTS.txt. 11 platforms x 3 = 33 ready posts. Pick one variant per platform, attach own image, post.

Voice rules for variants (research-backed, 2026): open IN-SCENE (a moment, not a thesis) · uneven rhythm (short. fragment. one longer line) · KILL rule-of-three / "it's not X it's Y" / inflated words / em-dashes · end on a LOW-LOAD CTA that drives comments (binary "chase or wait?", "tell me I'm wrong", fill-in-blank, oddly-specific question) NOT "what's your take?" · LI 900-1500 chars, IG hook <125 chars, FB/TT short.

Video = ONE shared, SHOOT-READY script, NOT per-platform. video_scripts.{business,travel}[] each = {style, angle, shots:[{t, line, angle, delivery, broll}], caption, hashtags}. The shots array is a director's breakdown: each beat tells you the exact LINE to say, the CAMERA angle/framing, the DELIVERY cue, and the ON-SCREEN/B-roll text. Rotate camera angles across beats (front-on tight -> push-in/lean -> 3/4 angle -> back to front-on for CTA) so the video isn't static. One shoot -> post to YouTube Short + TikTok + IG Reel + FB Reel. Do NOT write separate YT/TikTok/IG/FB scripts.

  • business video scripts MUST be driven by the LATEST industry/service news — pull a real, dated item from PHASE 1's today_items.json (AI agents / automation / AEO / the niches we sell), attribute the source claim (no fabricated stats), then tie it to a SMB service angle + comment CTA. One real news anchor per variant.
  • travel video scripts stay personal lifestyle (local daily-life), NOT industry news.

Single-copy fallback (1 per platform, from copy.json):

python scripts/export_text.py --date $(date +%Y-%m-%d)
  • Reads copy.json, emits COPY.html + COPY.txt.
  • NO images, NO _avatars, NO picker, NO html2canvas. You select + push your own images.

Image carousels are DEPRECATED (the old build_pack_v2.py render still works if ever needed, but the renderer had a silent blank-text bug; text export is the supported path). To resume rendered images: python scripts/build_pack_v2.py --date <DATE>.

PHASE 5 — (optional) GHL CSV row (5s)

python scripts/append_ghl.py --date $(date +%Y-%m-%d) --slot 10:00

6-col canonical CSV → tomorrow 10:00 local time. imageUrls left blank; paste manually-chosen image URLs before import. Skip entirely if posting by hand.

PHASE 6 — Tell the user what to do

Print:

OPEN: outputs/<DATE>/COPY.html

1. 📋 Each platform = a Copy button. Click → paste into the native composer.
2. 🖼  Pick + attach your own image(s) per post.
3. 📋 LinkedIn: fire the 3 comment-baits 1 per 15-20 min in first 60 min.
4. Done.

TOOL-STACK WARNING (research-backed)

  • NEVER post via Taplio — 30% reach penalty + 55% engagement penalty. LinkedIn algo flags AI-template patterns from Taplio.
  • Safe: native LI scheduler, Buffer, GHL scheduler.
  • Analytics-only (zero algo risk): AuthoredUp (hook preview), Shield Analytics (retro).
  • AVOID: Hypefury for LI (it's Twitter-first, weaker LI signal).

OUTPUT CONVENTION

outputs/YYYY-MM-DD/
├── today_items.json     # 30 deduped RSS items
├── angle.json           # angle + 2 hook variants + 5 candidates
├── copy.json            # Claude-written, voice-linted
├── index.html           # 8-slide carousel + post + reply pack
├── _avatars/            # 2 picked photos
├── _picked_images.json  # rotation audit
└── csv/ghl-linkedin.csv # tomorrow 10:00 local time

RULES (REPRIORITIZED v0.2)

  1. Voice spine = locked (reference POST 1 + POST 2). Don't flatten.
  2. Anchor required — 1 real local detail per post.
  3. Client win required — slug from client-wins.json. No invented stats.
  4. Vulnerability beat required — dated, specific, uncomfortable.
  5. Hook ≤210 chars — LI "See more" truncate point.
  6. Hashtags 5-8 — no more, no less.
  7. Comment-bait pack — 3 pre-staged replies. Fire 1 per 15-20 min first hour.
  8. lint_voice.py --strict must pass before render.
  9. Rotation log sacred — always --mark photos.
  10. No Taplio. Ever.

ANTI-PATTERNS

  • Don't paste hook from hook_variants verbatim — paraphrase tightly to keep voice unique.
  • Don't reuse a client win slug 2 days in a row.
  • Don't drop the → arrow glyph for ▸ • or em-dash — voice memory locked.
  • Don't fill vulnerability_beat with generic "I used to think AI was overhyped" — needs date, person, currency, or specific fail.
  • Don't omit the comment-bait pack — that's where 70% of reach lives.

TROUBLESHOOTING

Symptom Fix
feedparser missing pip install -r requirements.txt
0 items after dedup rm cache/seen.sqlite
Image picker fails check _USAGE_LOG.json exists in <your-images>
html2canvas blank hard-refresh; check _avatars/ populated
lint_voice.py FAILS read each FAIL line; fix copy.json; re-run
Hook too long use only first variant from angle.json, strip trailing clause
Anchor missing inject 1 line from local-anchors.json into Slide 2 + LI post body

ROADMAP

  • v0.3 — wire /reel-studio mograph mode for daily 30s video off same angle.
  • v0.4 — auto-push PNGs to GitHub carousels-png/<DATE>/ for ManyChat asset library.
  • v0.5 — Task Scheduler invocation: claude -p /aeo-daily daily 7am local time.
  • v0.6 — DONE (2026-06-08): engagement-loop live via weight_hooks.py + hook-weights.json overlay read by suggest_hooks.py.
  • v0.7 — auto-generate 2 hook variants via Claude inside Phase 3 (not pre-baked).
  • v0.8 — wire build_master_html.py auto-invoke into build_pack_v2.py; add --voice flag to build_pack_v2.py (still missing — renderer always emits all channels).
  • v0.9 — live-query the AEO Visibility Snapshot (auto-run ChatGPT/Perplexity/Gemini for the prospect instead of manual fill).

Last updated: 2026-06-08 (v0.6.0 — WEDGE "Get Cited" + FUNNEL (snapshot lead magnet, cta-bank, dm-flow) + 6 missing scripts built + lint v1.2.0 hardened across all channels + engagement hook-weighting loop live + client-wins refreshed)