Author
Hektor Jacynycz
CTO at Naktor
Technical architecture and project execution. Hektor leads Naktor's engineering, designing and building scalable and robust AI systems. With over 10 years of experience in distributed systems, data governance, and autonomous and decision-making systems based on generative AI, he ensures every project is executed with technical excellence.
Elsewhere
Pieces he signs
- Prompting for OpenAI: GPT-6 and the GPT-6.1 family
How to write prompts for GPT-6 Astra and the GPT-6.1 family: the developer channel, conflicting rules, quoted text and effort as a parameter.
- Prompting for Claude: Opus 5.5, Sonnet 5.5 and Haiku 5.5
How to write prompts for Opus 5.5, Sonnet 5.5 and Haiku 5.5: effort as the primary control, an append-only history, and verification per model.
- Prompting for Gemini: Gemini 4 Argon and the 3.x family
How to write prompts for Gemini 4 Argon and the 3.x family: the four input shapes, thinking_level, the parameters the API ignores, and context-first ordering.
- Prompting that works on any model
How to write a prompt that does not depend on the vendor: role and task, checkable constraints, the exact shape of input and output, and delimiters.
- Task prompts: one input, one output
How to write a prompt that does one thing: input and output described field by field, a declared value when the input cannot answer, and cases to test.
- Agent prompts: assistants with memory and tools
How to write an agent prompt: what persists between turns, where a rule belongs so it is still alive at turn twenty, and what route every tool failure has.
- Goal prompts: agents that chase a success criterion
How to write a prompt that hands over an objective and a success criterion and leaves the steps to the agent, with three endings: met, exhausted, blocked.
- Loop prompts: agents that repeat until a condition holds
How to write a prompt that repeats: the boundary between runs, the exit condition as a check against stored state, and the report after repeated failures.
- Tool design for agents: the description decides the call
How to write a tool definition an agent uses well: the description that decides selection, the verb_object name and parameters with a type and a unit.
- The end of temperature: what the three vendors are removing
All three vendors now remove or ignore temperature, top_p and top_k, each on a different date. What is left for the prompts you have in production.
- Gemini 4 Argon: thinking_level and the 400s from prefill
Gemini 4 Argon swaps a reasoning budget for a level and ignores temperature, top_p and top_k. The 400 comes from a prefilled model turn.
- Claude 5.5: effort rules, and the history is append-only
Claude 5.5 moves control into effort and makes the conversation an append-only history: editing an earlier turn invalidates every thinking block after it.
- GPT-6 Astra: the authority ladder over your system prompt
GPT-6 Astra ranks instructions into five authority levels and instructions expires when the reply completes, so the system block is superseded.
- Assistant prefill is gone: what replaces it
A trailing assistant or model turn now returns HTTP 400 on Anthropic and Google. The three ways to get what prefill used to buy you.
- prompts/v1 dies on 30 November: put your prompts in code
OpenAI is retiring prompt objects from its API: from editing an object in the account to keeping a file with review, typed inputs and tests.
- Gemini 3.5 Flash keeps its thinking: delete the recap
Gemini 3.5 Flash carries reasoning across turns and drops the default thinking effort to medium. Restating what was decided costs tokens for nothing.
- GPT-5.5 changed the effort default under your prompt
OpenAI moved reasoning_effort to medium by default in GPT-5.5 and left caching to the extended path only. Your code did not change.
- Your prompt did not change and costs 30% more
Opus 4.7 ships a new tokenizer: the same text is roughly 30% more tokens. How to measure the new cost, what to re-check and what not to do.
- Tool search: your tools no longer live in the prompt
GPT-5.4 defers the tool catalogue to runtime: what stops going into the prompt, what breaks, and what to check in your own tool list.