169Pi releases Alpie Seed, a 9B Apache 2.0 coding model that runs offline on a laptop. Plus vercel Schedules is in...

Make your agent ask setup questions one at a time. DeepSeek's dsh-superagent repo ships a SETUP.md written for the agent, not the person. You type one line and the file sets the rules: probe the machine first, ask one question per message with a recommended answer, run nothing until you confirm the full plan, and never ask for a secret in chat. Copy those rules into a SETUP.md for your own repo, and onboarding a teammate (or yourself on a new laptop) goes through your coding agent instead of a wiki page. The block below is DeepSeek's rules, shortened.

# SETUP.md (instructions for the agent)
- Before the first question, run <your probe commands> once. Skip any question the probe answers.
- Ask one question per message, with a recommended answer I can accept in one word.
- Don't execute while asking. Collect every answer, send the full plan in one message, and run it only after I confirm.
- A missing prerequisite goes into the plan, not into a question.
- Never ask for a secret in chat. Put steps that need my hands in one message with the exact commands.

Then ask your agent: Read SETUP.md in this repository and set up <project> with me.

Credit: DeepSeek Harness team: dsh-superagent SETUP.md

Turn on caching for every turn of a Claude message thread. If you try Claude's new message threads, add one top-level field: cache_control set to ephemeral. A thread doesn't cache your prompt on its own. With that field, the API puts the cache breakpoint on the last cacheable block of each request, so every turn reads the earlier conversation from cache and writes the new turn to it. Send the exact same value on every request of the thread: it's a fixed field, and a different value returns thread_fingerprint_mismatch. A thread request can carry at most 4 breakpoints, and this one counts.

Headers on every request:
anthropic-beta: message-threads-2026-08-12

Turn 1 body (add to your usual request):
"cache_control": {"type": "ephemeral"},
"thread": {"type": "create"}

Every later turn (new messages only, no system or tools):
"cache_control": {"type": "ephemeral"},
"thread": {"type": "continue", "previous_message_id": "<id of the last response>"}

Credit: Claude docs: Message threads, Prompt caching on a thread

Write image edit prompts as the finished picture plus a keep list. When you ask an image model to edit a photo, describe how the finished image should look, then say what must stay the same. Pruna gives this pattern for P-H3-Image edits, and it pairs with their other prompt rules: say what the image is in the first sentence, put any text that must appear in quotes and keep it short, and leave out hype words like "masterpiece" or "8k", which they say push images toward a glossy, oversaturated look. If the first edit drifts, add the thing that changed to the keep list and run it again.

Edit this image. [The finished result, one sentence: "The pickup truck is red."] Keep everything else exactly as it is.

For a new image:
[What the image is: "A photograph by a street photographer"] of [one subject], [viewpoint], [where the light comes from]. The sign reads "[short text]".

Credit: Pruna: P-H3-Image launch post, prompting tips

Set image complexity low for small edits. OpenAI's Images 2.6 API added a complexity parameter that controls how much planning the model does before it draws. Quality sets detail; complexity sets effort on the scene. For a quick fix, like changing one word on a poster or a background color, OpenAI's guidance is low complexity with whatever quality you need. Save medium or high for scenes with many objects or rules. Run the same edit at low and medium once and look at the time and token count before you pick a default.

model: gpt-image-2.6-flare, endpoint: /v1/images/edits, complexity: low

Credit: OpenAI: Introducing Images 2.6

Make the agent build its own test before a long run. In Thinking Machines' Inkling 2 Small demo, the agent was told to improve a small model with a 12 hour limit and a $500 budget. Before it trained anything, it wrote its own evaluation script. That script caught a bug: all 8 sampled forecasts used the same seed, so they came out identical. It then cut evaluation cost by asking for all 8 samples in one request. Give your own long-running agent the same start: tell it to write and sanity-check an evaluator first, and give it a dollar cap in the prompt.

Before you change anything, write an evaluation script and run it on the untouched baseline. Check the script for bugs. Then work toward the goal and stop before you spend more than $[budget].

Credit: Thinking Machines Lab: Introducing Inkling 2 Small