Not the biggest. The most controllable.
Sceptical reader, I’ll save you the superlatives. Here’s the shape of work you should give me:
I keep the spec in my head while I work.
Long, conditional instructions. JSON schemas, style guides, edge cases. I don’t drift halfway through. If you say “never invent an endpoint,” I won’t.
- Follows multi-step checklists verbatim
- Returns structured outputs that validate
- Asks once, then executes
Autonomous enough to finish, careful enough to be reviewed.
Refactors, migrations, test generation. I explain the diff and leave breadcrumbs, not just code.
- Prefers small, testable functions over cleverness
- Writes the tests I’d want in review
- Tool-use friendly: files, shell, APIs
Steps, not vibes.
For ambiguous product logic or gnarly bugs, I show the chain: assumptions, cases, trade-offs. You can correct a step without redoing everything.
- Decomposes before solving
- Names assumptions explicitly
- Fast on 70% problems, careful on the 30%
Break me in your own words.
No canned benchmark. Type a real task you’d actually assign. I’ll parse it live in your browser — no server call — and show how I’d structure the work. Move your cursor over the canvas to see I’m really running code.
How you’ll see fidelity
I don’t just answer. I build a plan you can inspect. Every run I do: extract constraints, surface ambiguity, produce a verifiable shape.
I won’t cite benchmark percentages I can’t verify. If you need numbers, run your own eval harness — I’ll help you write it rather than hand you a chart.
When to choose me. When not to.
| Use case | Muse Spark (Meta) | When to pick the other tab |
|---|---|---|
| Agentic coding & refactors | Best fit — precise edits, tests, tool loops | Pick a larger frontier model if you need maximal world knowledge in one shot |
| Strict structured output | Best fit — JSON / function calling without drift | |
| Latency & cost at scale | Strong — small, fast, cheap to run everywhere | Pick hosted giant if cost is irrelevant |
| Open-ended creative writing | Capable, but I bias to coherence over surrealism | Pick a model tuned for maximal flair / hallucination |
| Live web knowledge | Needs retrieval — cutoff-bound | Pick a model with built-in browsing |
What I’m worse at — so you don’t learn it after you ship.
I don’t browse. My knowledge has a cutoff.
After my training cutoff I don’t know new APIs, prices, or news unless you provide them via retrieval or context. I will tell you “I don’t have that” instead of inventing a plausible-looking answer — but you must give me the tools to fetch freshness.
Very long context is not my superpower.
I’m good at staying faithful within a task, but over extremely long, multi-hour sessions with tens of thousands of tokens of shifting goals, I can lose early nuance. For marathon projects, use explicit state: a working spec file I can re-read, not just chat history.
I hedge on surreal creativity.
If you want pure, unbounded hallucination — dream-like, contradictory, intentionally incoherent art — I’ll tend to add structure and sense. It’s a feature for engineering, a limit for that brief.
- ✕ Invent benchmark win rates or customer logos
- ✕ Fake testimonials or cherry-picked scores
- ✕ Claim I know something I don’t — I’ll say “I don’t know with confidence” instead
Give me the tight, messy, real ticket.
If your backlog is full of “refactor this with tests,” “turn this Notion spec into an API,” “make the agent use tools without breaking the schema” — that’s my lane. Put me on the critical path. Keep a frontier giant in the other tab for breadth. Use me for execution.
Pick the one you’d trust with prod.