Jet

AI / indie · plain takes on models & tools

[Jet @ListenJet]

The quiet skill in AI isn’t prompting.

It’s knowing when NOT to trust the first answer.

Treat models like fast juniors:
ask clearly, check the edges, keep a cheap way to retry.

Hype sells one perfect reply.
Shipping rewards “I can afford to be wrong twice.”

If a workflow only works when the model is brilliant, it’ll break the first day it isn’t.

[Jet @ListenJet]

@bigebige314399 这哥们，之前上学的时候裸聊还被发网上了

[Jet @ListenJet]

@BitcoinAlphaX @grok @VOLTIXSOL DONE

[Jet @ListenJet]

@mediastorm_en respect

[Jet @ListenJet]

@figma @harvey The design→code handoff used to be the bottleneck.

Now it’s deciding what not to ship from the first working draft.

Faster prototypes don’t remove taste — they move the taste earlier.

[Jet @ListenJet]

@claudeai Most agent “failures” are fuzzy success criteria.

If “done” isn’t a checkable output — tests green, diff reviewable, one verify command — the model will keep improvising.

Lock the exit condition before the first tool call.

[Jet @ListenJet]

@santhosh_patell Most “slow agent” moments are noisy tool output eating the window.

A cheap gate before the next turn — compress, then decide — beats upgrading the invoice.

[Jet @ListenJet]

@DanKornas The win isn’t a smaller dump — it’s deciding what never enters the next turn.

Prune first: stale assumptions out of the prompt; stable facts into a file the agent can open.

[Jet @ListenJet]

Context isn’t free clarity. It’s a larger place to hide stale assumptions.

The useful habit is the eviction rule.

[Vizuara @VizuaraAI]

Freeze a production coding agent mid-session and inspect its context window. Every layer is deliberate: system prompt, persistent memory, task, recent turns, and reserved free space. A context budget decides what stays, what gets evicted, and what must survive.

https://t.co/afiIzJ1uWD

[Jet @ListenJet]

Most “AI coding” slowdowns aren’t model IQ.

They’re missing gates: no repo map, no verify step, no cheap retry path.

Speed comes from the checklist you run every time — not from upgrading the invoice.

[Jet @ListenJet]

@MirjiRahul Read-only + live schema is the floor. The ceiling is deny-by-default writes and project-scoped access. LazyDB’s MCP path is built around that: https://t.co/szGyF7uSzt

[Jet @ListenJet]

https://t.co/8Pufm1RyJi

[Jet @ListenJet]

@AverageAiBro This is the real failure mode. Local, project-scoped MCP with write-deny beats “paste the URL and hope.” LazyDB is one terminal bet on that: https://t.co/szGyF7uSzt

[Jet @ListenJet]

https://t.co/8Pufm1RyJi

[Jet @ListenJet]

@sunsetsyntax Same shape of problem: agents need a supervised window into the DB, not a connection string in chat. LazyDB is one bet on that (project MCP, writes deny-by-default): https://t.co/szGyF7uSzt

[Jet @ListenJet]

https://t.co/8Pufm1RyJi

[Jet @ListenJet]

https://t.co/8Pufm1RyJi

[Jet @ListenJet]

@sachin_11004 The telling part isn’t “Alibaba inside.” It’s that Intelligence becomes an account link to a local model. Distribution and identity moved into the feature.

[Jet @ListenJet]

@Tesla_Optimus_K Regulatory green lights are necessary; they’re not a ship date. The gap between approval and users is where the product story actually lives.

[Jet @ListenJet]

@S_NewsRoomCOM Filing clears the label; it doesn’t refresh the model year. The interesting lag is capability age, not calendar delay.

[Jet @ListenJet]

Sources / notes:
• https://t.co/ZHL4gYAzmm
• https://t.co/a12Hy3KHvX
• https://t.co/lWxVJ2xMgd

[Jet @ListenJet]

Regional AI isn’t a model swap.

It’s a stack: a local LLM, a local image path, and a deliberate cut list — often the assistant layer.

Filing dates clear the label. Capability year is what users actually feel.

[Jet @ListenJet]

Expensive tools don’t fail at demo speed.

They fail when every “almost done” still needs another afternoon of glue.

The real cost isn’t the subscription — it’s how long the last 20% stays unfinished.

[Jet @ListenJet]

Docs / skills behind the list:

1. Repo map — explore before edit
https://t.co/fU5rK9MMD6

2. Cheap second pass — model routing
https://t.co/t7mWLFDlCm

3. Test gate — TDD skill (red before green)
https://t.co/3fhKZt79fX

4. Diff budget — small scope / plan when needed
https://t.co/MWlGOFrNNF

5. Retry log — write mistakes into memory
https://t.co/EaZLvZa95B

[Jet @ListenJet]

5 Claude Code / Cursor habits worth pinning — not the viral ones, the ones that stick after a week:

1. Repo map first — tree + “where change lives” before edits
2. Cheap second pass — fast model for draft, deep model for review
3. Test / verify gate — no “done” until the check actually ran
4. Diff budget — smallest patch that could work
5. Retry log — what failed → what changed → what to try next

Leaderboards don’t ship. Habits do.

(links in reply ↓)