Stacks, how-tos, and rankings
for actually using AI in your business.

Why these picks

• 010% take rate vs Substack's 10%. Ad network + Boosts pay for the subscription most weeks.
• 02Better tone matching than GPT-5. Longer instruction adherence than Gemini 3.
• 03Database queries no one else does as well. Free tier is enough for a one-person shop.
• 04Local-first transcription — audio never leaves the laptop. Doesn't feel like a SaaS demo.
• 05Direct Stripe Payment Links in the footer. No platform fee, no upsell flow. Covers hosting most months.

We'd switch if…

• Beehiiv raised Scale past $75/mo — evaluate Buttondown ($29/mo) or Kit ($89/mo).
• Sonnet lost the tone edge to Gemini 3 or a future GPT.
• Notion kept adding AI features we don't want. Obsidian is the likely replacement.
• Granola changed the local-first model.

What we tried and dropped

• Substack — 10% take rate, social feed pollution, discovery engine wants your readers more than you do.
• ConvertKit / Kit — fine product, pricing scales faster than Beehiiv at our band.
• Mailchimp — not even close.
• Ghost — beautiful, but the CMS-first design adds friction we don't need.

Why these picks

• 011M context at standard pricing. Tool use that doesn't hallucinate arguments. Best instruction-following for multi-step work.
• 02Durable steps that retry from where they failed with cached outputs. Replaces a queue + state machine + cron.
• 03Managed Chromium with stealth (FingerprintJS >80% human). Plain Playwright is the boring answer that doesn't break.
• 04Postgres + auth + storage + edge in one place at $25/mo. Cold-starts in 200–500ms.
• 05Every step traced — tokens in/out, cost, latency, the prompt. Without this you have no idea why an agent burned $40 overnight.

We'd switch if…

• Anthropic raised Sonnet past $5/$20. GPT-5.4 is close on price but tool use is noisier.
• Inngest lost durable-step semantics. Trigger.dev is cheaper at entry but less mature.
• Browserbase raised prices materially. Bug0 Browsers ($0.15/hr) is the credible alternative.
• Supabase kept pushing self-hosted complexity. Neon is on the table for bursty workloads.
• Langfuse changed unit-pricing math after the ClickHouse acquisition.

What we tried and dropped

• LangChain / LangGraph — abstraction tax we didn't need. Anthropic SDK + Inngest does the same with fewer footguns.
• AWS Step Functions — every line becomes IaC. Fine for big orgs, terrible for shipping.
• Temporal — over-engineered for what we do. Inngest is what Temporal would be if it shipped this decade for normal teams.
• CrewAI / AutoGen — multi-agent frameworks that solve a problem we don't have.

Why these picks

• 01Current quality leader for cinematic shots with native audio. Veo 3.1 Lite ($0.05/sec) is the volume tier.
• 02Cleanest creative pipeline — motion brush, director mode, Aleph editing, frame interpolation. Best for character consistency.
• 0370+ cinematic camera presets nobody else has solved. Also the easiest way to access Veo / Kling / Seedance under one sub.
• 04Best AI voice on the market. Flash and Turbo cost half the credits per character — the plan stretches twice as far.
• 05One-time $295 license. Runs locally on a decent GPU. Free updates for 12 months, yours forever after.

We'd switch if…

• Veo 3.1 prices climb. Kling 3.0 is the volume fallback — Sora is no longer credible after the OpenAI shutdown.
• Runway loses the editing pipeline edge. Nothing competes today on the integrated motion-brush + Aleph workflow.
• Higgsfield changes credit math on Kling again — they've quietly devalued credits twice in 12 months.
• ElevenLabs raised entry above $30. PlayHT and Hume are credible alternatives.
• Topaz moves to a subscription. We checked — still one-time. If that flips we move to a stable open-source upscaler.

What we tried and dropped

• Pika — fine for goofy social content, not for client work.
• Luma Ray2 — was a contender mid-2025, fell behind.
• Sora 2 — off the board entirely. App shut April 2026, API ends Sept 2026.
• Adobe Firefly Video — feature parity coming, cost-per-second uncompetitive.

Why these picks

• 01Handles 30–50% of inbound at most B2C companies. Per-resolution pricing means cost scales with success — honest, can spike.
• 02API-first, Slack-native, AI on all tiers. GraphQL API is fully public and unrate-limited.
• 03Purpose-built for Slack Connect channels with B2B customers. Account health scores make it more than triage.
• 04Not a tool you buy — a workflow you build. 50 labeled examples, one prompt, <$5/mo at high volume.
• 05Interactive decision-tree help docs for the top 5 ticket-generating issues per quarter.

We'd switch if…

• Fin raised per-resolution past $1.50. Today's math is fine; can spike fast.
• Plain loses API-first ethos or rate-limits the public API.
• Pylon gets acquired and the Slack-native edge softens.
• Anthropic Sonnet loses the classification edge to Haiku 4.5 at $1/$5 — we'd save 5×.
• Stonly keeps raising prices. Cheaper guided-flow tools exist; editor and analytics are the moat.

What we tried and dropped

• Zendesk — fine for enterprises, expensive and rigid for everyone else.
• Freshdesk — ok for small teams, lost the AI race.
• Help Scout — beloved by some, didn't ship AI features fast enough.
• Custom GPTs / OpenAI Assistants — works as a prototype, falls apart at production volume.