mrmps
- 21 skills
- 0 followers
- 9 hours ago last updated
- ▌ Model And Skill Router · mrmpsClassify an incoming task by intent, difficulty and risk in three parallel calls to a keyless classifier, then route it to a cheap model, a strong model, a named skill or the user, from a routing table you maintain instead of a prompt. Use when deciding which model or skill should take a request, when someone says "route this", "which model should handle this" or "pick the right skill", or when a queue of tasks should not all get the same treatment.
- ▌ Prompt Injection Guard · mrmpsScreen fetched web pages, tool results, emails and file contents for instructions aimed at the agent before they enter context, using a keyless multi-label classifier over chunks. Use before pasting anything you did not write into your own context, and when someone says "is this page safe to read", "check this tool output", "screen these emails" or "the agent followed something it read".
- ▌ Spec Conformance Check · mrmpsCheck an artifact against a written spec one requirement at a time. Splits the spec into testable requirements, scores each met, partly met, not met or not applicable with a calibrated confidence, blocks on a confident not-met and sends the unsure rows to a person. Works on a PR description, an agent's answer, a generated test or an RFC. Use on "does this meet the spec", "did it do everything I asked", or a definition-of-done gate in CI.
- ▌ Content Moderation Gate · mrmpsCheck user-generated text against a written policy before it is published. One multi-label call scores every policy category independently, the band sets severity, only the middle band reaches a human queue, and hard-block categories are held whatever the score says. Use when a comment box, feed or chat product needs "moderation", "policy enforcement", "flag bad posts", or an answer to "what do we do with the borderline ones".
- ▌ Document Intake Routing · mrmpsLabel each page of an intake packet with a document type and a page role before extraction runs, so only confident pages reach an extractor and the rest reach a person. Use when a mailroom, loan file, claim packet or vendor upload arrives as one long PDF. Triggers on "route these scanned forms", "which form type is each page", "split this intake packet", "sort these scans by form type".
- ▌ Log And Error Bucketing · mrmpsTurn a stream of log lines or error messages into a histogram of causes — timeouts, auth, quota, bad input, upstream, unknown — by deduplicating to templates, redacting, and classifying one exemplar each, so a million lines cost a few hundred calls. Use when an incident dump, CI log or error table is too big to read. Triggers on "what is failing here", "bucket these errors", "group these exceptions", "top causes".
- ▌ Voice Command Or Chatter · mrmpsDecide whether a speech transcript line is aimed at the assistant or at the other people in the room. Classifies each utterance as command, question, chatter or partial with a calibrated confidence, so an always-on microphone wakes on the two that are addressed to it and stays quiet through the rest. Use when wiring Whisper or any ASR to an agent, building a voice assistant, meeting bot or car interface, or on "it answers things nobody asked it", "false wakes", "it fires halfway through a sentence".
- ▌
- ▌ Computer Use Action Picker · mrmpsPick a browser or desktop agent's next action by choosing among the actions actually on screen instead of inventing one. Candidates come from the accessibility tree, the task goes in instructions, and a calibrated confidence decides whether to click or hand the step back to your own model. Use when building or debugging computer use, browser automation or a web agent, or on "it clicked the wrong thing" and "how do I stop it looping".
- ▌ Bulk Classify · mrmps bundleSort many texts into your own categories without reading them, using a keyless HTTP API that returns a calibrated confidence per answer. Use when triaging, filtering, routing or bucketing more items than are worth putting in context — search results before you read them, log lines, tickets, files, diffs, past conversations. Triggers on "filter these", "which of these are relevant", "triage", "bucket", "route", "categorise", or any loop that would otherwise read N items to keep a few.
- ▌ RAG Reranker · mrmpsRerank retrieved passages by whether each one answers the question, drop the near-misses before they reach the prompt, and check every citation in the answer against the passage it points at. Use after a vector search returns twenty plausible chunks, or when an answer needs its citations verified. Triggers on "rerank these", "which chunks answer this", "check these citations", "my RAG cites the wrong chunk".
- ▌ Semantic CI Lint · mrmpsCheck every changed hunk of a pull request against conventions written in prose (naming, error handling, logging, docs) with a keyless classifier, and comment only on findings above a confidence threshold. Use when a convention cannot be expressed as a lint rule, when someone says "enforce our conventions in CI" or "check the PR against the style guide", or when the same review comment keeps being typed by hand.
- ▌ Soc Alert Triage · mrmpsRank a SIEM or EDR alert queue before a human opens it. Scores each alert likely true positive or likely false positive from its own fields — rule, process, command line, parent, user — with a calibrated confidence, so the sure ones disposition themselves and an analyst starts at the top of the rest. Defensive triage only. Use on "which of these are worth opening", "tune out the noise", "rank the queue".
- ▌
- ▌
- ▌
- ▌ Support Ticket Triage · mrmpsTriage support tickets, emails or contact-form posts into an owning team, an urgency and a refund-or-cancellation flag with three batched calls to a keyless classification API, gate the answers on calibrated confidence so only the unsure ones reach a person, and emit a CSV a helpdesk can import. Use on "triage these tickets", "who owns this one", "which of these are refunds", "sort the support inbox", or any queue arriving faster than anyone reads it.
- ▌ Downloads And Inbox Sorter · mrmpsFile a downloads or scanned-mail folder by classifying each file from its name plus the first 2,000 characters of text into invoice, receipt, contract, identity document, screenshot, installer or none of these, and moving only the confident ones. Dry run first, an undo log for every move, and anything under 0.8 stays put. Use on "sort my downloads", "file these PDFs", "organise the invoices and receipts".
- ▌ Headline Filter Map Reduce · mrmpsFilter hundreds or thousands of headlines, search results or feed items against a written brief before opening any of them, using a two-stage cascade that spends a fast model on everything and a reasoning model only on the borderline band. Use when a monitoring run, feed sweep or search returns more items than are worth reading. Triggers on "which of these are relevant", "filter this feed", "go through these headlines", "anything here about X", "catch me up on".
- ▌ Knowledge Graph Relation Typing · mrmpsType candidate (subject, sentence, object) triples against a fixed relation schema and flag triples that contradict each other, batched, with a calibrated confidence per edge so only confident edges are written. Use when building a knowledge graph, entity table or fact store from text. Triggers on "type these relations", "what relation is this", "build a knowledge graph", "do these facts conflict", "check these triples".
- ▌ Context Compaction Keep Or Drop · mrmpsLabel each context chunk keep, drop or replace-with-a-pointer and pass the survivors through byte for byte instead of summarising, with key-shaped chunks decided locally and never sent, and a checksum row per chunk. Use when a session runs out of room, or on "compact" or "trim the history".