João Galego

bag-of-tricks

the answer. nothing else. — strip LLM replies of personality, filler, hedging, emoji, and sycophancy.

see what your agent actually said behind your back. — a transparent proxy + web UI that logs the exact bytes your agent sends the model.

argue with yourself before the internet does. — red-team a prompt across jailbreak, injection, derailment, extraction, and ambiguity.

make it ask before it acts. — flip the helpful-assistant reflex: ask one sharp clarifying question instead of guessing.

two letters, and the prompt writes itself. — short aliases that expand into full prompts for common dev tasks.

rip the JSON out of the chatter. — extract and repair valid JSON buried in chatty LLM output (fences, trailing commas, Python literals).

pat it down before it ships. — scan text headed to the model for secrets and PII (API keys, tokens, private keys, emails) and redact or flag them.

every AI has a tell. — flag the giveaways in AI-written prose (delve, tapestry, "it's not just X, it's Y", em-dash overuse).

know the bill before the bill. — estimate token count and dollar cost of a prompt across models before you send it.

call its bluff. — extract the URLs and citations from an answer and check they actually resolve, catching hallucinated links.

find the plant. — sniff out prompt-injection and planted instructions hiding in untrusted input before it reaches the model.

wash out the prints. — strip the mechanical fingerprints from text: zero-width chars, smart quotes, exotic spaces, em-dash tics.

does the story check out? — flag answer claims that aren't supported by the provided sources (a lexical grounding check for RAG).

know when to fold. — catch overconfident phrasing and absolutes so the model hedges or abstains instead of bluffing.

put it in the hot seat. — adversarially interrogate an answer with probing follow-ups before you trust it.

same prompt, the whole lineup. — run one prompt across several models and lay the answers side by side.

we know your prints. — guess which model wrote a passage from its stylistic fingerprints (a heuristic, not proof).

put the squeeze on it. — detect AI-generated text by how it compresses: model prose squeezes flatter against a known-AI corpus than a human one (after thinkst's zippy). A heuristic, not proof.

pull the whole routine. — chain tricks into one pipeline so the output of one flows into the next: `combo "frisk | launder | deadpan"`.