moonlight-lupin

Autonomous multi-step deep research engine implementing an iterative Think → Search → Extract → Synthesize → Stop loop. The LLM drives every decision: what to search, what's relevant, what's missing, and when to stop. Produces a cited, magazine-quality report with inline citations, category- specific formatting, and research stats. Trigger when the user asks for "deep research", "research report on", "comprehensive analysis of", "look into X in depth", "write a report on X", or any question needing multi-source synthesis beyond a single search. For entity vetting/dossiers use entity-research; for news digests use news-monitoring; for source-grounded Q&A use notebooklm-mode.

Track upstream drift and sync adapted skill libraries

Use when the user wants to organize, tidy, or restructure a messy directory (Downloads, Desktop, documents folder). LLM-powered file organizer that scans content, proposes a structure, and executes moves in chunks with a mandatory user confirmation gate.

Compile a task brief BEFORE starting substantial work — pin down the goal (what "done" looks like), the context (everything known that bears on the task), the constraints (guardrails, with the organisation's standing rules included automatically) and the tooling (which installed skill or tool will do the work) — confirm it with the user, then execute against it. Use whenever the user says "brief this task", "brief this first", "make sure you understand before you start", "scope this properly", "compile a brief", "what do you need from me", "improve my prompt", "help me get better output", or hands over a SUBSTANTIAL task — multi-step, multi-document, or producing a deliverable someone else will read — as a thin one-line request. Asks at most 2-3 clarifying questions, and only where the answer would change the output; otherwise it states its assumptions in the brief and proceeds. Not for quick one-step asks — just answer those directly.

Convert Claude Code plugins into self-contained Hermes plugins — discovery analysis then full conversion

Parse agent log files to identify error patterns, rate limit hits, timeout clusters, tool failures, and component-level error counts. Produces a structured anomaly report. Cron-compatible — silent if no issues, alert digest if anomalies found.

Generate short marketing/social VIDEO clips via fal.ai through a staged studio workflow — brainstorm the motion, draft cheaply, then produce the final at quality. Three modes: text-to-video (a clip from a prompt, no source image); animate (bring a still image to life — a product, people, b-roll, a space); and camera-move (a gentle push-in, pan, tilt or shallow orbit over a still, including an interior "3D view"). Use when the user wants to "make a video/clip", "generate a video from text", "animate this image/photo", "create a marketing reel / social video / teaser", "do a 3D / parallax move", "pan/zoom/orbit a shot", or when they mention fal.ai, Kling, Veo or Seedance for video. Do not use for still images (use image-studio), Canva-style template designs, slide decks, data charts/dashboards, or flowcharts/diagrams. Carries an honesty discipline for real subjects: never use text-to-video to depict a real, identifiable place, product or person. Do not override platform-native video tools where the host environ

Generate, edit and upscale AI images via fal.ai through a three-stage studio workflow — brainstorm a strong prompt with the user, prototype cheaply and iterate on feedback, then produce a finalised image. Use when the user wants image generation or editing through the local fal.ai helper workflow, including requests to "generate an image", "make an image/picture/illustration/graphic of…", "create an AI image", "edit/change this image", "make a variation", "upscale this", create a "production-ready image", "clean up / enhance a photo", "make this phone shot look professional", create imagery "for the deck/post/website/newsletter", or when they explicitly mention fal.ai or nano-banana. Do not use for Canva template designs, branded PowerPoint decks, data charts/dashboards, or flowcharts/diagrams — those are layout, data, or structure tasks, not generative imagery. Video generation is out of scope. Do not override platform-native image generation tools where the host environment requires them.

Reconcile any two record sets (A vs B) and triage the discrepancies — match line-by-line on a shared key, or heuristically on amount + date, then classify every unreconciled item (category, materiality, probable cause, suggested action) into a reconciliation working paper (.xlsx). Use whenever the user wants to "reconcile A to B", "reconcile the bank to the ledger", "reconcile the invoice tracker to the accounts", "tie out", "match these two lists/exports", "find the differences between", "what doesn't match", "reconciliation working paper", or "discrepancy triage". Ships presets for common recurring reconciliations (invoice tracker vs accounting records, bank vs cashbook, fund administrator vs internal records, payments/PRF vs bank); generic for anything else. Deterministic and computed locally — it never force-fits a match and never posts an adjustment; it produces a working paper for finance to review and sign off. NOT budget/variance analysis, and NOT deal/asset analysis (that is out of scope).

Analyse a dataset and deliver the insights and key metrics that matter for it — an insight brief with headline findings, trends, breakdowns, concentration, outliers and ageing, tailored to the type of data (transactions, receivables, pipeline, survey, task list, any table). Use when the user says "analyse this data", "what are the key metrics", "any insights from this spreadsheet", "summarise this export", "what's driving the numbers", "who are the top customers", or hands over a table and asks what it says. Works INTENT-FIRST: asks what question the data should answer, profiles it, proposes a plan → you confirm → computes every metric deterministically (local engine, exact Decimal, currency-aware) → writes a calibrated brief separating observation from interpretation. Metrics computed locally by the engine. NOT data cleaning (data-tidy), NOT matching two datasets (data-reconcile), NOT a dashboard (data-visualise — natural next step); descriptive analysis only, never financial or investment advice.

Re-express structured data to hit a DIFFERENT target — map a source (system A's export) onto another system's import CONTRACT, and/or RESHAPE its structure (wide↔long, nested JSON↔flat table, split one file into many, union many into one). Use when the user wants to "convert this to X's format", "map these columns to the import template", "get this into the format System B needs", "reshape / pivot / unpivot this", "flatten this JSON", "split this file by column", "combine these files", or "prepare this for upload/import". Works INTENT-FIRST: proposes the target shape from the contract + purpose, you confirm, it applies deterministically and writes a reusable **conversion card** (Markdown + an embedded JSON spec) a future agent re-runs after sense-checking the source. Delegates cleaning to data-tidy; never invents values; live inputs (FX) are pinned/user-approved, not fetched. NOT data cleaning (data-tidy), NOT matching two record sets (data-reconcile), NOT raw file-format-only conversion.

Get STRUCTURED data OUT of documents — PDFs (incl. multi-table/scanned via local OCR), Word, PowerPoint (.pptx), and Outlook .msg — into a clean .xlsx plus an audit report. Use when the user wants to "extract data from this PDF/document", "pull the table out of this report", "get the figures from these statements/certificates", "turn these confirmations into a table", "read the fields off this form", or "extract these line items". Two modes: key-value/FORM extraction (label → value, one record per document — certificates, confirmations, cover sheets) and TABLE extraction (list a document's tables, pick one, pull it). Intent-first; normalises via the shared engine (dates → DD MMM YYYY, currency → amount + code) and flags anything unfound or uncertain — never invents values. Extraction is computed locally (local OCR only). NOT for already-tabular data (use data-tidy) or deal-document intelligence like lease abstraction/model review (out of scope).

Clean up messy data into a structured, validated table — from any source (a junk-filled .xlsx/CSV, a pasted email/markdown table, a Word table, or an Outlook .msg) to a clean .xlsx plus an audit/change report. Use when the user wants to "clean up this data", "tidy this spreadsheet", "normalise this list", "structure this messy export", "dedupe this", "standardise these dates / currencies", or "turn this into a clean table". (For getting data OUT of PDFs — tables, forms, scanned documents — use data-extract, not this skill.) Works INTENT-FIRST: it asks what the data is for and the expected output up front, then profiles → proposes a transform recipe → you confirm → applies it deterministically → reports every change and flags cells for review. Never mangles silently; transforms are computed locally by a deterministic engine. Standalone — no other toolkit required. NOT deal-document intelligence (lease abstraction, model review, comps) — that's out of scope.

Orchestrate visual output from tabular data or a data-analyse analysis.json into either (1) a brandable self-contained HTML dashboard (print/PDF / artifact) or (2) an Excel workbook of native charts for analysts. Use when the user says "build a dashboard", "visualise this", "make a chart / KPI cards / scorecard", "Excel charts", "chart this in a spreadsheet", "a one-pager of these numbers", "RAG status board", or wants a shareable visual summary. HTML path: inline SVG, no CDN. Excel path: openpyxl charts with OfficeCLI-aligned chartType names (column/bar/line/pie/doughnut/waterfall). Draft for review, not advice. NOT PowerPoint or letters; clean/extract first via data-tidy / data-extract; compute metrics via data-analyse when numbers must be exact.