Fable 5 Reception, Scraping Stacks, Dev Toolkits — AI Daily Jul 02

469 messages · 77 active members

469
messages
77
active members
@samb69, @arielletolome, @jasonakatiff
top contributors

Overview

Day two of Fable 5's redeployment brought a sharply mixed reception. Enthusiasts like @filiuser, @iggot, and @samb69 reported one-shot Shopify checkout clones, a 2.7M-fact production platform built in a day, and Fable 5 low-effort outrunning Opus 4.8 max at ~30% fewer tokens with ~4x better context compaction. Skeptics including @sav310 and @samtome called it hot garbage on redesigns, slower on some tasks, and prone to verbose non-answers on yes/no prompts. The bigger operational headache was silent auto-switching to Opus 4.8 mid-task — even on innocuous SEO writing — which torches sessions; toggling `Switch models when a message is flagged` to false in `/config` was shared as a partial fix. Token burn is real (@namtalks hit 7% of a 5hr window on one prompt), and Anthropic appears to be throttling capacity until GPT-5.6 lands. Tooling talk centered on scraping stacks and multi-agent orchestration. The group mapped Apify (site-specific actors), Firecrawl CLI (lowest token overhead for static pages), ScrapingBee, Brightdata, ScrapeCreators, and Cloudflare's new /crawl endpoint. On the harness side, @samb69 moved off cmux to Ghostty + herdr and shared the omp.sh forked harness, @csoares12 explored multiple Hermes profiles on one VPS, and @rmktg explained the profiles-as-specialists pattern (Creatives → Mediabuying → Research). @geilt delivered a masterclass on scaling AI dev teams via standardized toolkits, frontmatter-tagged markdown docs, AI-only doc folders, and unified credentials — culminating in a public `toolkit-lite` GitHub release. On the applied side, @arielletolome reported ~$29.75 CAC from an AI creative system, shipped Yapper ad outputs and an Offer Cracker that mines Reddit and competitor ads to rank affiliate angles, and flagged Meta suspending custom apps (pushing users to system users or Composio/Hermes MCPs). @jasonakatiff introduced Goverly.ai, a Telegram-orchestrated agentic govtech platform, while @zippi101 one-shotted a browser game and @iammikevineyard announced a weekend Fable game jam.

Topics

Reactions split sharply — @filiuser one-shot a working Shopify checkout clone and @GuruTime built a 2.7M-fact production platform in a day, while @sav310 and @samtome reported worse output than 4.8 on redesigns. Multiple users hit unexpected auto-switches to Opus 4.8 mid-task; toggling `Switch models when a message is flagged` to false in `/config` prevents silent fallbacks (Fable will refuse instead). Token consumption is brutal — @namtalks burned 7% of his 5hr window on a single prompt — with Anthropic seemingly throttling capacity.

Kicked off by @fmill1 noting Claude Code refuses to scrape, the group mapped preferred stacks: Apify for site-specific actors (TikTok, Reddit, Amazon), Firecrawl CLI for low-token static-page pulls, ScrapingBee as a cost-competitive alternative, plus Brightdata, ScrapeCreators, and Cloudflare's new /crawl endpoint. @arielletolome flagged Grok MCP/CLI for Reddit thread scraping.

@samb69 moved off cmux to Ghostty + herdr and shared the omp.sh harness with IDE wiring. @csoares12 and @rmktg discussed running multiple Hermes profiles as specialists (Creatives / Mediabuying / Research) with sub-agents per profile. @geilt released a public `toolkit-lite` repo demonstrating shared docs/toolkit/ops repos, frontmatter-indexed markdown, AI-only doc folders, and unified credential management. @fuckyesiwannatalkbusiness swapped Codex's context compressor for Gemini 3.5 Flash for a measurable speed win.

@arielletolome reported ~$29.75 CAC on a live account using a structured AI creative system, and shipped raw and v2 Yapper ad outputs (45s and 60s) built largely one-prompt with n8n enforcing script accuracy. Her Offer Cracker takes any affiliate link and ranks angles by potential using Reddit research plus historical competitor ads. Meta is suspending custom apps, pushing users to system users or MCPs like Composio/Hermes for ad uploads.

@jasonakatiff introduced Goverly.ai — an agentic defence/govtech platform with a Telegram-managed n8n overseer routing to sub-agents. @zippi101 one-shotted a browser game (Hivebreak) with Fable, and @iammikevineyard announced a weekend Fable game jam. Side threads covered enterprise AI token waste and the IP surrender risk of feeding proprietary logic to foundation model providers (Anthropic's Slack integration cited as the most blatant vector).

Key Takeaways

  • Toggle `Switch models when a message is flagged` to false in Claude Code `/config` to stop silent Opus 4.8 fallbacks — Fable will refuse instead of downgrading and torching your session.
  • Fable 5 compacts context ~4x better than Opus and burns ~30% fewer tokens on hard planning, but individual prompts can consume 5–7% of a 5hr cap — front-load it on audits and architecture.
  • Firecrawl CLI is the lowest-token option for static-page scraping inside Claude Code; use Apify actors for site-specific jobs (TikTok, Reddit, Amazon) and test Cloudflare's new /crawl endpoint.
  • Treat docs as first-class infra: frontmatter-tagged markdown in dedicated repos with an AI-only subfolder outperforms code-as-documentation, enables cleaner RAG, and lets agents and humans hit the same known-good state.
  • Run multiple Hermes profiles as specialists (Creatives / Mediabuying / Research) rather than separate installs — different models per profile, easy switching, sub-agents spawn per profile.

Hot Threads

@fmill1started

What are people using to scrape? Claude Code refuses

22 replies11 participants
@geiltstarted

Centralized docs, toolkits, and standardized AI dev team infrastructure

18 replies5 participants
@sav310started

You guys really all that hyped on Fable? Hot garbage on my redesign

15 replies8 participants

Linked Items