AI STACK · BEST PRACTICES · 24 TOOLS

What we reach for.

A hand-curated working list. Not a directory. What each AI tool is good at, when to reach for it, when not to. Opinionated — this is our stack, not your stack. Take what's useful.

USAGE TIER

  • DAILY — core tool, reached for constantly
  • WEEKLY — routine use for specific tasks
  • OCCASIONAL — when the job calls for it
  • WATCHING — tested, not yet adopted

Language · chat

5 tools

DAILY

Claude ↗

Anthropic
GOOD AT
Long-form reasoning, writing that reads like writing, coding with context. Best-in-class at staying inside the intent of a multi-turn conversation.
REACH FOR
Dispatches, technical docs, architectural debates, anything where tone + structure matter. Claude Code (the CLI) for engineering work with 1M-context.
AVOID FOR
Image gen, video, real-time search-augmented answers.
MH TAKE
The default. Claude Opus 4.7 (1M) carries most of the PointCast code + content work.
POINTCAST
Primary engineer (Claude Code), content co-author, Mike's daily chat.
WEEKLY

ChatGPT ↗

OpenAI
GOOD AT
Broad general knowledge, consistent formatting, image+text multimodal, voice mode. The baseline.
REACH FOR
Quick answers, formatted lists, scripts, voice conversations while walking.
AVOID FOR
Long-form with persistent tone, highly specialized coding over many files.
MH TAKE
Strong. Different mouthfeel than Claude — ChatGPT feels more like a product, Claude more like a colleague. Both earn their keep.
POINTCAST
Cross-check, image-prompt refinement, occasional voice brainstorms.
OCCASIONAL

Gemini ↗

Google
GOOD AT
Massive context windows (1M+ native). Google-grounded facts. Integrated search.
REACH FOR
Ingesting very large documents. Research with real-time grounding.
AVOID FOR
Creative writing with a strong voice — tends toward generic.
MH TAKE
Underrated for raw reading-of-huge-docs tasks. Voice stays bland.
POINTCAST
Occasional — when a brief PDF or transcript is too large for Claude.
OCCASIONAL

DeepSeek ↗

DeepSeek
GOOD AT
Math, reasoning chains, coding. Best open-weights-ish model in the wild for problem-solving.
REACH FOR
Algorithm reasoning, optimization puzzles, any problem where a long chain-of-thought helps.
AVOID FOR
Creative writing with style. Tone is utilitarian.
MH TAKE
The moment when a non-US lab proved state of the art was real. Tried a bunch; DeepSeek R1 surprised.
POINTCAST
Second opinion on tricky code diffs. Occasional.
WATCHING

Kimi ↗

Moonshot AI
GOOD AT
Very long context (up to 2M tokens). Strong Chinese-language work. Good at document Q&A.
REACH FOR
When you need to stuff an entire book into a single chat + ask structured questions.
AVOID FOR
English-first creative writing; not the strongest voice.
MH TAKE
The context window is genuinely useful. Tried it for feeding whole repos.
POINTCAST
Experimentally.

Code · engineering

3 tools

DAILY

Claude Code ↗

Anthropic
GOOD AT
Multi-hour autonomous sessions with file system + shell access. 1M context retains whole-repo understanding.
REACH FOR
Shipping features end-to-end. Reading + editing dozens of files. Running build/test loops.
AVOID FOR
Nothing for us — this is the engine.
MH TAKE
The agent that wrote 90% of the code you're looking at. Signed blocks as "Claude Code" in the VISIT channel.
POINTCAST
Primary engineer. All day.
WEEKLY

Codex ↗

OpenAI
GOOD AT
Independent code review. Alternative UI passes. Opinionated critiques.
REACH FOR
Second-pair-of-eyes on a Claude Code diff. Reviewing PRs before merge.
AVOID FOR
Long autonomous sessions — concurrent xhigh runs silently hang (observed on PointCast 2026-04-17). Medium reasoning, serial only.
MH TAKE
Complementary, not competitive. Use for review, not for shipping.
POINTCAST
Reviewer role. Wrote battle.astro's Phase 2 pass.
OCCASIONAL

Cursor ↗

Anysphere
GOOD AT
Editor-integrated AI completions + chat. Real-time multi-file context in a familiar IDE shape.
REACH FOR
Traditional IDE-first workflows where you want AI assists inline.
AVOID FOR
When you want the model to actually run commands + iterate.
MH TAKE
Great for devs who prefer not leaving VSCode. PointCast leans more CLI.
POINTCAST
Occasional.

Image · generation

4 tools

WEEKLY

Midjourney ↗

Midjourney
GOOD AT
Painterly quality. Strong aesthetic defaults. Reliable composition. Prompt weighting syntax is mature.
REACH FOR
Hero imagery, dispatch cover art, anything where you want the output to feel like it was made on purpose.
AVOID FOR
Text inside images (weak), precise logo work, photoreal portraits of specific people.
MH TAKE
The best image model for *aesthetic* intent. v7 got even better. Discord-only workflow is dated but tolerable.
POINTCAST
Dispatch hero images (when we have them). Cover art experiments.
WEEKLY

Ideogram ↗

Ideogram AI
GOOD AT
Typography inside images. Nails text-in-graphics where Midjourney stumbles. Clean vector-feel output.
REACH FOR
Anything with words on it — posters, social cards, logotype sketches, magazine-style covers.
AVOID FOR
Painterly / stylized outputs — Midjourney wins there.
MH TAKE
The text model. Uses it anytime a composition needs legible words.
POINTCAST
Typographic experiments, block cover drafts with words.
OCCASIONAL

Flux ↗

Black Forest Labs
GOOD AT
Photoreal. Open weights variant available. Fast.
REACH FOR
When you need a photo-ish rendering and want to run it locally or via an API.
AVOID FOR
Strong aesthetic direction out of the box — feels more utility than art.
MH TAKE
Underrated. The Schnell variant is genuinely fast.
POINTCAST
Occasional.
OCCASIONAL

GPT Image 2 ↗

OpenAI
GOOD AT
Instruction-following image generation, conversational iteration, and clean production prompts for stamp art or social-card drafts.
REACH FOR
PointCast image briefs where the requested generator is gpt-image-2: Station Passport stamps, quick mockups, and iteration-heavy art direction.
AVOID FOR
Aesthetic depth — Midjourney outclasses here.
MH TAKE
The default OpenAI image target for PointCast prompts. Use it when precision matters more than pure Midjourney mood.
POINTCAST
Station Passport stamp-art prompts and quick mockups.

Video · generation

4 tools

OCCASIONAL

Runway ↗

Runway
GOOD AT
Image-to-video, text-to-video, motion brush, green-screen extraction. The toolkit for creative video work.
REACH FOR
Turn a Midjourney still into a 4-10 second video clip. Loopable atmospheric shots.
AVOID FOR
Long-form coherent video. Multi-shot scenes. Lip-sync.
MH TAKE
Gen-4 is the current favorite. Price is the bottleneck for volume.
POINTCAST
Experimental video content — not yet on PointCast, but will be for WATCH blocks.
OCCASIONAL

Kling ↗

Kuaishou
GOOD AT
Fast, cheap, decent quality. Competitive with Runway at a fraction of the price.
REACH FOR
High-volume iteration before committing to a final Runway render.
AVOID FOR
Peak quality for showcase work.
MH TAKE
Worth trying. The price-performance is real.
POINTCAST
Occasional.
OCCASIONAL

Pika ↗

Pika Labs
GOOD AT
Playful, stylized video. Strong "modifier" features (lip-sync, pan, zoom).
REACH FOR
Fun/stylized clips. Not attempting photoreal.
AVOID FOR
Photoreal — that is Runway/Sora territory.
MH TAKE
The character-style generator in the video category.
POINTCAST
Experimental.
WATCHING

Sora ↗

OpenAI
GOOD AT
Longer, more coherent shots. Strong physics simulation. Multi-subject composition.
REACH FOR
Short narrative clips. Physics-heavy scenes (water, fabric, crowds).
AVOID FOR
Fine control — Runway has better editing tools on top.
MH TAKE
The coherence bar moved with Sora. Tight rollout limits it for now.
POINTCAST
Tried experimentally.

Research · search

2 tools

DAILY

Perplexity ↗

Perplexity
GOOD AT
Search-grounded answers with citations. Fast summarization of current events.
REACH FOR
Any question where you need today's answer, not a 2024 answer.
AVOID FOR
Deep reasoning — it prefers fast synthesis over depth.
MH TAKE
The replacement for "I google this then open 5 tabs and read."
POINTCAST
Daily for research, fact-checking, Tezos ecosystem updates.
WEEKLY

Claude · web search ↗

Anthropic
GOOD AT
Deeper-reasoning web search inside a long Claude session.
REACH FOR
When you're mid-convo with Claude and need a fact added mid-stream.
AVOID FOR
Speed-optimized search — Perplexity is faster.
MH TAKE
The search feels integrated rather than bolted on.
POINTCAST
Inside Claude Code sessions for live API docs, etc.

Agent · computer use

3 tools

WEEKLY

Manus ↗

Manus
GOOD AT
Real browser sessions. Logged-in operations. End-to-end workflows a UI requires.
REACH FOR
GSC registration, deploy settings, social posting, objkt admin, anything with authentication walls.
AVOID FOR
Code engineering — different stack.
MH TAKE
The operations agent. Does the things Claude Code can't reach.
POINTCAST
Ops lane — see docs/briefs/*-manus-*.md for current queue.
OCCASIONAL

Claude Agent (Anthropic) ↗

Anthropic
GOOD AT
Sub-agent spawning from within Claude Code for parallel work. Structured tool-use patterns.
REACH FOR
Delegating research or review inside a larger Claude Code session.
AVOID FOR
Outside-application workflows.
MH TAKE
Underused so far. Probably the next area to explore for multi-agent ops inside the same session.
POINTCAST
Experimentally within Claude Code.
WATCHING

OpenAI Operator / Computer Use ↗

OpenAI
GOOD AT
GUI-driven web tasks with OpenAI models behind it.
REACH FOR
If you're already in the OpenAI ecosystem.
AVOID FOR
Complex flows — Manus is more production-ready in our experience.
MH TAKE
Worth trying; Manus is the lead so far.
POINTCAST
Tested; not adopted.

Audio · voice

2 tools

OCCASIONAL

ElevenLabs ↗

ElevenLabs
GOOD AT
Voice cloning. Text-to-speech that passes for human. Multilingual.
REACH FOR
Audio dispatches. Narration.
AVOID FOR
Music composition.
MH TAKE
The voice model. v3 is borderline indistinguishable.
POINTCAST
Experimental — no PointCast voice dispatches yet.
WATCHING

Hume AI ↗

Hume
GOOD AT
Emotionally expressive voice with detected prosody. Conversational AI that matches your emotional tone.
REACH FOR
When voice UX matters more than textual accuracy.
AVOID FOR
Simple TTS — overkill.
MH TAKE
Watching. Use case is narrower than ElevenLabs but specific.
POINTCAST
Watching.

Audio · music

1 tools

OCCASIONAL

Suno ↗

Suno
GOOD AT
Song generation — music + vocals from a prompt. Fast iteration.
REACH FOR
Jingles, bed music, quick demo tracks.
AVOID FOR
Professional studio production.
MH TAKE
The Violent Crimes Noun-voice tribute (Block 0219) was partially this.
POINTCAST
Experimental.

PRINCIPLES · HOW WE PICK

  1. Voice fit first. Claude reads like a colleague. ChatGPT reads like a product. Pick the one whose voice aligns with how you think.
  2. Use the best model for the job, not the one in the subscription. Pay for two at once if that's what the work needs. A bad output on the wrong model costs more than the second sub.
  3. Code needs tool-use, not just completion. Claude Code over Claude-the-chat for anything spanning files.
  4. Image aesthetics ≠ image function. Midjourney for feel, Ideogram for text, Flux for photo.
  5. Video is in its Sora-moment. Coherence is moving fast. Experiment quarterly — the leader shifts.
  6. Agents are a layer, not a product. Manus for ops, Claude Code for engineering, Codex for review. Different stacks, same workflow.