Oskar Austegard

GitHub repository access in containerized environments using REST API and credential detection. Use when git clone fails, or when accessing private repos/writing files via API.

Decide which model, effort level, and cascade shape each subagent gets, and how to keep improvement loops safe (evaluator-as-selector, stop on regression). Routes on measured cost-per-completed-task rather than per-token price, because a tier's token count varies more by task shape than price varies across tiers. Covers per-model effort semantics, the concision lever, cascade preconditions, context handoff, and watching a subagent fan-out live. Use when spawning subagents via the Agent or Workflow tools, when choosing how to escalate a failed attempt, when fanning out more than a handful of agents, or when asked which model or effort a task should get. Grounded in measured calibration (references/calibration-2026-07-15.md), a 2026-08 coding-cost study, and a 2026-09 agentic-repair battery that measured the cascade rungs directly; Managed Agents API specifics are operational, not calibrated.

Securely manages API credentials for multiple providers (Anthropic Claude, Google Gemini, GitHub). Use when skills need to access stored API keys for external service invocations.

Guidance for asking clarifying questions when user requests are ambiguous, have multiple valid approaches, or require critical decisions. Use when implementation choices exist that could significantly affect outcomes.

Pre-change blast-radius report for a symbol or file. Walks tree-sitting references, augments with a plain-text scan over non-parsed files (configs, plain docs), and clusters affected sites by feature (`_FEATURES.md`) or top-level package. Use when about to refactor, rename, or delete something in a repo you don't own — "what breaks if I change `validateUser`", "who calls this", "is this safe to remove", "where is this used", "blast radius", "impact analysis". This is the CONVERGENT pre-change risk skill — for "what is this repo?" use exploring-codebases; for "where is X?" use searching-codebases.

Read Bluesky/ATProto without depending on Bluesky's AppView — interactions on a user's posts, thread replies, any repo's records, network-wide backlinks across every lexicon, an account's handle and PDS history, which collections are active network-wide, and layer-by-layer outage diagnosis. Use when bsky.app or the AppView is down or slow, when a Bluesky read returns timeouts or 5xx, when asked who liked/replied/quoted/reposted something, when pulling live Bluesky context cheaply, when reading records straight from a PDS, when asked who references a record or account anywhere on the network, whether an account has changed handle or migrated servers, or which lexicons and non-Bluesky apps are active on atproto. Complements browsing-bluesky, which routes everything through the AppView and fails when it does.

Ranked content search over a text corpus you point it at, using BM25 (via xhluca/bm25s). Corpus-agnostic: cloned repos, project knowledge stores, uploaded files and archives, any local directory. In-memory BM25 index per invocation, with a session-local disk cache for repeat runs against the same corpus. Use for "rank these documents", "search this corpus", "which files are most about X", "find content about Y", or any multi-word concept query against a known body of text where grep would return everything or nothing. Needs a corpus on disk. Not for searching stored memories or prior-session decisions (remembering), not for a named symbol or a file's structure (tree-sitting), and not for a literal string you could grep.

Browse Bluesky content via API and firehose - search posts, fetch user activity, sample trending topics, read feeds and lists, analyze and categorize accounts. Supports authenticated access for personalized feeds. Use for Bluesky research, user monitoring, trend analysis, feed reading, firehose sampling, account categorization.

Generate progressive disclosure indexes for GitHub repositories to use as Claude project knowledge. Use when setting up projects referencing external documentation, creating searchable indexes of technical blogs or knowledge bases, combining multiple repos into one index, or when user mentions "index", "github repo", "project knowledge", or "documentation reference".

DEPRECATED - Use building-github-index instead. Superseded per-file GitHub API implementation of progressive disclosure repository indexes.

Analyze and categorize Bluesky accounts by topic using keyword extraction. Use when users mention Bluesky account analysis, following/follower lists, topic discovery, account curation, or network analysis.

Cross-context adversarial review for deliverables before shipping. Use when producing blog posts, technical recommendations, analysis briefs, code, or any artifact where accuracy matters more than speed. Triggers on "challenge this", "review before shipping", "adversarial pass", "stress test this".

Select the right Python charting library (seaborn, matplotlib, graphviz) and produce publication-quality static visualizations. Use when creating charts, plots, graphs, diagrams, heatmaps, visualizations from data, or when choosing between matplotlib/seaborn/graphviz. Also triggers for network diagrams, flowcharts, dependency trees, state machines, and entity-relationship diagrams. For interactive browser-rendered charts or uploaded data exploration, defer to charting-vega-lite instead.

Create interactive data visualizations using Vega-Lite declarative JSON grammar. Supports 20+ chart types (bar, line, scatter, histogram, boxplot, grouped/stacked variations, etc.) via templates and programmatic builders. Use when users upload data for charting, request specific chart types, or mention visualizations. Produces portable JSON specs with inline data islands that work in Claude artifacts and can be adapted for production.

Validates development tool installations across Python, Node.js, Java, Go, Rust, C/C++, Git, and system utilities. Use when verifying environments or troubleshooting dependencies.

Exports a Claude.ai workspace — its custom instructions and its uploaded knowledge files — into a bundle that can be re-imported elsewhere. Use for "back up my project instructions", "export my project knowledge", "copy this workspace to a new one", "migrate my project to another account", or "recreate this setup somewhere else". Operates on the Claude.ai workspace you are talking in, and reads nothing from disk.

Close a GitHub issue with a synthesis comment as a flowing graph — validate the synthesis, post the closing comment, close, then run a pluggable callback (e.g. memory store) detached. Use when closing an issue should also capture the LEARNING (not just the diff log) and when the post-close work shouldn't block the close ack.

Develop and run Mojo code in Claude.ai containers. Handles installation, compilation, and execution. Use when writing Mojo code, benchmarking Mojo vs Python, or when user mentions Mojo, Modular, or MAX. Routes to Modular's official skills (mojo-syntax, mojo-python-interop, mojo-gpu-fundamentals) for language-specific correction layers.

Composes single-file HTML artifacts (PR review writeups, status reports, incident postmortems, slide decks, design systems, prototypes, flowcharts, module maps, feature explainers, kanban boards, prompt tuners) from a small JSON spec instead of hand-written HTML/CSS/JS. Use when the user asks to "compare options side-by-side", requests an HTML version of a report or review or deck, asks for a flowchart, status update, postmortem, design system reference, interactive prototype, custom editor — or explicitly says "HTML artifact", "single HTML file", "self-contained HTML". Skip for ad-hoc HTML snippets (forms, emails, embedded widgets) where there's no template fit.

Universal environment variable loader for AI agent environments. Loads secrets and config from Claude.ai, Claude Code, OpenAI Codex, Jules, and standard .env files.

Authors and caches a personalized container environment from a Dockerfile- like spec, as a single layer or as a composition of independently cached layers. Use when the user mentions "container layer", "Containerfile", "custom container", "cache my installs", "composable layers", "uv shim", or wants package installations, skills and environment config to survive an ephemeral session. Also for snapshotting, restoring or rebuilding that environment, and for capturing ad-hoc installs into a reproducible spec.

Control Spotify playback and manage playlists via MCP server. Use when user requests playing music, controlling Spotify, creating playlists, searching songs, or managing their Spotify library.

Convenes expert panels for problem-solving. Use when user mentions panel, experts, multiple perspectives, MECE, DMAIC, RAPID, Six Sigma, root cause analysis, strategic decisions, process improvement, or asks for philosophers/ancients (Socratic, Aristotelian, Stoic method experts).

Convert a file from one format to another inside the container — documents, images, audio, video. Routes to the right engine (pandoc, LibreOffice, ImageMagick, ffmpeg) by format pair. Triggers on "convert X to Y", "turn this docx into a pdf", "make a gif from this mp4", "md to docx", "batch-convert these images", or any single-file or batch format change where the source and target extensions differ. NOT for editing content (use docx/pptx/xlsx/pdf skills), creating files from scratch, or reading a file you already have in context.

Chooses the right FORMAT for instructions on Claude.ai — project instructions, a skill, or a standalone prompt — and gives the format-specific structure once chosen. Use when the question is which container an instruction belongs in ("should this be a skill or project instructions", "where do I put this", "how do I set up this project", "is this worth a skill"), or when someone has instructions and does not know how to package them. For the writing principles that apply inside any of the three formats, use writing-instructions. For building, testing and packaging a complete skill directory, use creating-skill.

Creates browser-executable JavaScript bookmarklets with strict formatting requirements. Use when users mention bookmarklets, browser utilities, dragging code to bookmarks bar, or need JavaScript that runs when clicked in the browser toolbar.

Builds a portable, embedding-free knowledgebase from a set of files and delivers it as a self-contained `.skill` bundle (BM25 index + bundled searcher + query protocol). Use when a user wants to turn uploaded files, a folder, or a corpus into a searchable knowledgebase they can hand to any agent — phrased as "make a knowledgebase", "build a KB skill", "package these docs for retrieval", "create a searchable bundle", or references to a `.skill` KB. The output runs anywhere with Node or Python — no model, no install, no network. Distinct from `bm25` (ephemeral in-session search) and `building-github-index` (markdown project-knowledge index).

Creates production-ready MCP servers using FastMCP v2. Use when building MCP servers, optimizing tool descriptions for context efficiency, implementing progressive disclosure for multiple capabilities, or packaging servers for distribution.

Builds and revises a complete skill DIRECTORY — SKILL.md, scripts, references, assets — and packages it. Use for "create a skill for X", "turn this into a skill", "update/improve this skill", "why doesn't my skill trigger", "review this SKILL.md", "package this skill", or when a repeated procedure should become a reusable artifact. Enforces the structure that makes skills work — a concrete procedure rather than a goal statement, an explicit applicability boundary, failure modes with their signals, a runtime verification step, and a description checked against the existing catalogue for confusability. For choosing whether the instruction should be a skill at all rather than project instructions or a prompt, use crafting-instructions. For writing quality inside the prose, use writing-instructions.

Create video from prompts by overseeing multi-clip AI generation end to end: write a shot list, generate each scene with Gemini Omni Flash (via the Cloudflare AI Gateway), review the results, and assemble them into a finished cut. Use when the user asks to make/generate a video, a short film, an animatic, or a multi-scene clip from a script or idea; when they mention Omni, Veo, text-to-video, or image-to-video; or when acting as the editing/director agent over generated footage. Triggers on 'make a video', 'generate a clip', 'short film', 'video from this script', 'turn this into a video', 'omni', 'veo', 'text to video', 'storyboard to video'. For transcoding/trimming/merging/GIF/subtitles use processing-video; for reading or summarizing existing video content use parsing-video.

Find tests that enumerate a domain by copying it, and declare the invariants a codebase depends on. Reports where a parametrize list, for-loop, or it.each iterates a hand-written subset of a dict/set/tuple/Enum that exists in the source, and names the members nothing covers. Use when reviewing tests, when a module gains a name-to-thing table, registry, enum, or dispatch map, before trusting a green suite as evidence a domain is covered, or when asked "is this test actually total", "does anything cover X", "what does this repo guarantee", "which invariants do we declare". Also for vacuous tests that pass over an empty collection, for a domain that has silently NARROWED (an enumeration cannot see that), and for recording the refutation that proves a claim can fail.

Load at the start of any task whose deliverable is prose another person will read — a PR description, commit message, README, doc, postmortem, blog post, issue or review comment, report, release note or essay — before drafting it, whether the result is pushed, posted, published or saved to a file, without being asked. Draft, then run this pass on the draft before handing it over. Also use when someone says "de-claude", "de-slop", "humanize this", "this reads like AI", "make it sound human", or asks for a voice, tone or register edit. Rewrites the constructions that mark prose as model-written (staged reveals, verdict headers, aphoristic closers, "it's not X, it's Y", em-dash drama, forced triads, flat-certainty adverbs) into plain technical prose and checks the rewrite kept every claim. Not for fiction, poetry, code, or quoted text; for a full adversarial review of a deliverable use challenging.

Delegates work to a subagent or spawned session by writing only the task, while a PreToolUse hook pages the parent transcript through Jev and appends the chunks that task needs. Use before writing any Agent tool prompt or create_session prompt, and when asked to delegate this, hand this off, use a subagent, spawn an agent, run agents in parallel, or brief a subagent on what we have done so far.

Text-prompted image zone detection using TIPSv2 B/14 on CPU. Produces `focus_targets` / `focus_edges` bbox lists from natural-language labels, ready to feed into `svg-portrait-mode`. Use when you want automatic foreground/background separation from prompts like "dog face" + "wooden floor" instead of hand-annotating bboxes.

Specialized Preact development skill for standards-based web applications with native-first architecture and minimal dependency footprint. Use when building Preact projects, particularly those involving data visualization, interactive applications, single-page apps with HTM syntax, Web Components integration, CSV/JSON data parsing, WebGL shader visualizations, or zero-build solutions with vendored ESM imports.

Distill Opus-level reasoning into optimized instructions for Haiku 4.5 (and Sonnet). Generates explicit, procedural prompts with n-shot examples that maximize smaller model performance on a given task. Use when user says "down-skill", "distill for Haiku", "optimize for Haiku", "make this work on Haiku", "generate Haiku instructions", or needs to delegate a task to a smaller model with high reliability.

Turn a short video into a narrated comic strip. Use when the user uploads a video (.mov/.mp4/etc.) and asks to make a comic from it, narrate it, draw it, Attenborough it, storyboard the clip, or otherwise wants a stills-plus-narration comic strip derived from footage. Covers probing and extracting frames, GROUNDING the actual storyline (native Gemini video parse or frame reading), composing narration in a chosen voice, generating comic panels with Gemini image models anchored to the real stills, QA-ing the panels, and compositing a titled strip. Do NOT use for plain video transcoding or trimming (that is processing-video) or for original illustration from a text prompt (that is invoking-gemini).

Draws images, PDF/SVG vector files and animation frames with pycairo, and supplies what cairo lacks: shaped text for Arabic, Indic and other scripts, text along curves, blur/glow/drop shadows, path booleans, stroke-to-outline, and Gouraud-shaded 3D. Use when writing pycairo or cairo drawing code, generative art, custom diagrams or posters rendered in Python, or when cairo output needs non-Latin text, soft effects, merged shapes or shaded surfaces.

Renders a Bluesky post as an anonymised screenshot so its words can be criticised without a link back to whoever wrote them. Blacks out avatar, display name, handle and any at-mentions in declassified-document styling, then emits a PNG and its alt text. Use for anonymous dunk-quoting, "dunk on this without linking it", "screenshot this post with the author hidden", "quote this without naming them", "redact this before I share it", or when a bad take from a small account deserves an answer and its author does not deserve your followers.

First-encounter orientation on a repository nobody here has worked in yet. Runs a fixed five-step workflow — venv setup, tarball fetch, tree-sitting structural scan, featuring synthesis, then reasoning over the two — and yields an account of what the repo contains and how it is arranged, optionally written out as _FEATURES.md. Use for "I just cloned this", "what is this repo", "what does this do", "explore this repo", "give me an orientation", "what are the main features", "review what's new in this repo", or before starting work in a codebase you have not seen. This is the divergent what's-here skill. Route elsewhere for: a named symbol, a file's structure or a line range (tree-sitting); all callers of a Python symbol (searching-codebases); teaching a human the codebase through exercises (orienting-codebases); fetching or cloning a repo without analysing it (accessing-github-repos, cloning-project).

Exploratory data analysis. Use when users upload .csv/.xlsx/.json/.parquet files or request "explore data", "analyze dataset", "EDA", "profile data". Small files get ydata-profiling HTML/JSON reports; large files (over 200MB or 5M rows) get fixed-memory DuckDB/sketch profiling. Also covers near-duplicate row detection, cross-file key overlap ("can these join?"), dataset drift vs a stored baseline, and time-series profiling.

Extract keywords from documents using YAKE algorithm with support for 34 languages (Arabic to Chinese). Use when users request keyword extraction, key terms, topic identification, content summarization, or document analysis. Includes domain-specific stopwords for AI/ML and life sciences. Optional deeper extraction mode (n=2+n=3 combined) for comprehensive coverage.

Generate hierarchical _FEATURES.md files that describe what a codebase DOES from a user/consumer perspective, anchored to source symbols via tree-sitting. Supports large complex codebases through feature-driven decomposition into sub-feature files. Uses a multi-pass synthesis: orientation → detail → overview rewrite. Use when someone says "what does this do", "document features", "feature inventory", "_FEATURES.md", or needs to understand a codebase's purpose before modifying it. Complements tree-sitting (structural) with semantic (why/what-for) layer.

Retrieve clean markdown from URLs when web_fetch fails. Converts pages via Jina AI reader service with automatic retry. Use when web_fetch or curl returns 403, blocked, paywall, timeout, JavaScript-rendering errors, or empty content or user explicitly suggests using jina.

Discover and load skills on demand from /mnt/skills/user/. Use when you need a capability but don't know which skill provides it, when the boot-emitted skill list is names-only and you need a full description, or when you want to list the catalog. Verbs are list (names only), search (rank by name/description match against a query), and show (emit the full SKILL.md for a named skill).

Runs a multi-step procedure as a Python DAG, so ordering, branching and retries are enforced by the runner rather than described in prose a model can generate past. Use for "run these steps in order and retry the flaky one until the check passes", "build a pipeline that fetches, validates, then skips the upload when nothing changed", "make sure these steps cannot be skipped", "resume from where it broke instead of redoing the expensive early stages", "run these independent calls at once and merge the results", or any procedure of 3+ steps with branches, input contracts, or side effects that must not block the critical path. Primitives are depends_on, when=, validate=, retry_until=, detached= and journal_path=. Not for a single sequential call, for steps needing reasoning between them that no predicate captures, or for async and distributed work. To audit whether one verification check can actually go red, use gating. To fan work out across many subagents, use a dynamic workflow.

Zero-shot univariate time series forecasting using the Reverso foundation model (NumPy/Numba CPU-only inference). Activate when users provide time series data and request forecasts, predictions, or extrapolations. Supports Reverso Small (550K params). Triggers on "forecast", "predict", "time series", "Reverso", or when tabular data with a temporal dimension needs future-value estimation.

Build and audit deterministic verification gates — a check that blocks a pipeline and can be shown to go red. Use when writing a calibration gate, CI check, validation script or pre-publication check for a numeric or empirical result; when a plausible-but-wrong value would survive review; when asking whether an existing test, linter rule or check could actually fail; and when a suite passes first try, passes suspiciously often, or was written by whatever produced the thing it checks. Triggers on "can this check fail", "known-bad", "negative control", "calibration gate", "sanity check my results", "is this test actually testing anything".

Generates git patch files from codebase modifications for local application. Use when user mentions patch, diff, export changes, bring changes back, apply locally, or after editing uploaded codebases.

Break out of a locked problem frame by picking one disciplined move — reframe, provocation (Po), random stimulus, SCAMPER, inversion, structured analogy, constraint play, or family traversal — and committing to it before evaluating. Use when stuck, when options feel narrow or obvious, when iterations produce variations of the same idea, or when the user says "widen this", "break out of", "think differently", "I'm stuck", "feels too obvious", "stress-test the framing", "what am I missing", or holds two related examples and asks what lies between or beyond them. Complements challenging (which evaluates) and convening-experts (which synthesizes viewpoints); this skill generates distance, not judgment.

Installs and authenticates the GitHub CLI (gh) in a Claude.ai container, then runs gh commands against it. Use for "install gh", "gh auth", "create an issue or PR from the CLI", "gh is not authenticated", or any GitHub operation you want to issue as a gh command rather than a raw API call. This skill is the transport rather than the analysis. To understand what a repository does or contains, use exploring-codebases; to add a repo to the session, accessing-github-repos.

Assign items to a CLOSED label vocabulary that is too large to put in a prompt — product taxonomies, category hierarchies, tag vocabularies, routing tables, ICD/SIC-style code lists. A cheap model writes the label it thinks the vocabulary would use, and an embedder snaps that writing onto the nearest legal value, so the schema is never transmitted and the output is always in-vocabulary. Use for "classify these into our taxonomy", "tag these against the existing tag list", "map these queries to categories", "the enum is too big to send", or a Literal/enum that hits a provider cap. NOT for a vocabulary that fits in a prompt — structured output measured 0.701 acc@1 there against this pattern's 0.564. NOT for open-ended labelling with no fixed vocabulary, and not for ranked retrieval over documents (bm25).

Delivers a static Hello World HTML demo page with bookmarklet. Use when user requests the hello demo, hello world demo, or demo page.

Convert raster images (photos, paintings, illustrations, line art) into SVG vector reproductions. Use when the user uploads an image and asks to reproduce, vectorize, trace, or convert it to SVG. Also use when asked to decompose an image into shapes, create an SVG version of a picture, or faithfully reproduce artwork as vector graphics. Handles graphic/line-art inputs (Kandinsky, architectural drawings, ink work) via a compositional pipeline that extracts lines as SVG strokes. Do NOT use for creating original SVG illustrations from text descriptions — only for converting existing raster images.

Use when the user explicitly asks to use the informed-patient skill to prepare for a medical appointment, organize symptoms before seeing a doctor, or evaluate the evidence behind a diagnosis or treatment. Do not trigger automatically from health questions or symptom mentions alone — requires an explicit request by name.

Discovers and indexes Python code in skills, enabling cross-skill imports. Use when importing functions from other skills or analyzing skill codebases.

Install skills from github.com/oaustegard/claude-skills into /mnt/skills/user. Use when user mentions "install skills", "load skills", "add skills", "update skills", "refresh skills", or references a skill not currently installed.

Install and drive Google's Antigravity CLI (`agy`) as a non-interactive sub-agent. Use when orchestrating agy, running Antigravity agents from a script or sandbox, delegating a task to Google's agent harness, or wanting a Gemini-backed peer agent alongside Claude.

Invokes Google Gemini models for structured outputs, image generation, text-to-speech narration, multi-modal tasks, and Google-specific features. Use when users request Gemini, image generation, Gemini TTS or a synthesized voice, structured JSON output, Google API integration, or cost-effective parallel processing.

Enables GitHub repository operations (read/write/commit/PR) for Claude.ai chat environments. Use when users request GitHub commits, repository updates, DEVLOG persistence, or cross-session state management via GitHub branches. Not needed in Claude Code (has native git access).

Multi-conversation methodology for iterative stateful work with context accumulation. Use when users request work that spans multiple sessions (research, debugging, refactoring, feature development), need to build on past progress, explicitly mention iterative work, work logs, project knowledge, or cross-conversation learning.

Generate guardrailed UI from natural language. Emits constrained JSON that a Preact runtime renders. Use when the request is for a dashboard with metrics, charts, or tables; an admin panel; a data visualization interface; or a form-based application.

Listen to generated music by measuring it and by reading it as sheet music: render Strudel code or record a Web Audio page through the real engine in headless Chromium, then read spectrograms, pYIN pitch lines, chroma, roughness and semitone-rub meters; engrave notes as a score with beat-by-beat harmony; transcribe a melody from audio; read MusicXML/MIDI/ABC or a score image into notes and into Strudel code. Use to check whether generated music sounds right or better (before/after a change), to match a reference track, spectrogram image or score, or to find discordant chords, a buried melody, clipping or muddy mixes. Triggers on 'does it sound better', 'listen to', 'render and check', 'spectrogram', 'sfft/stft', 'discordant', 'clashing chords', 'sounds off', 'compare to the original', 'reproduce this track in Strudel', 'sheet music', 'score', 'notation', 'MusicXML', 'MIDI', 'ABC', 'transcribe', 'OMR'. Pairs with strudeling (writing Strudel); for format conversion use processing-video.

Execute programs on a compiled transformer stack machine where every instruction fetch and memory read is a parabolic attention head. Demonstrates that transformer attention + FF layers can implement a working computer. Use when user mentions "llm-as-computer", "lac", "stack machine", "compiled transformer", "percepta", "parabolic attention", "execute program", or asks to run/trace programs on the transformer executor.

Generates WAFFLES Declarations for social media posts — preemptive lists of what a post does NOT say. Use when users mention WAFFLES, ask for clarifications on their post, want to prevent misinterpretation, or request disclaimers for controversial/nuanced takes.

DEPRECATED — superseded by tree-sitting. This skill generated persistent _MAP.md files; tree-sitting does the same AST extraction at runtime (~700ms for 250 files) with no artifact to write, commit, or keep in sync. Use tree-sitting for "map this codebase", "explore repo", "understand structure", or any code navigation task. The final working version is archived at release mapping-codebases-v0.8.0.

Generate navigable semantic maps from PDF documents. Extracts section structure via font analysis, then runs LLM extraction per section for claims, symbols, and dependencies — all page-anchored. Produces _MAP.md (progressive disclosure), .symbols.json (definition index), .anchors.json (claim references), and a _USAGE.md snippet for CLAUDE.md. Use when analyzing papers, specs, or legal docs; when asked to "map this document", "index this PDF", "what does this paper say"; or when a coding agent needs grounded reference material from a PDF source. Analogous to tree-sitting but for prose documents.

Generate behavioral/feature documentation for web apps using code-first analysis. Reads source code via tree-sitting to produce _FEATURES.md, with optional visual verification via browser automation. Companion to tree-sitting. Use when documenting app behavior, creating feature inventories, generating behavioral ground truth for agents, or before modifying UI code. Triggers on "map features", "document app behavior", "feature inventory", "what does this app do".

Puts oaustegard/muninn-utilities on the container Python path — runs scripts/boot.sh, which shallow-clones the repo (codeload tarball as fallback) and writes the .pth. The boot payload itself comes from the Muninn MCP connector: call its `boot` tool after this script. Trigger on the first message of any new Muninn conversation, or when the user says "boot up" / "boot" — unless the message contains "skip boot". Idempotent: a warm container exits in ~0s via a sentinel.

Adds a searchable text layer to a scanned PDF with ocrmypdf. Installs the toolchain at runtime in ~18s. Use when a PDF's pages are images and the deliverable is a file to keep, grep, or hand to another tool: 'make this PDF searchable', 'OCR this scan', 'I can't select the text in this PDF', 'search across these scanned pages', 'the pdf skill returned nothing'. Handles non-English scans via tesseract language packs. NOT for charts, diagrams, handwriting, or slide layouts. Route those to transcribing-images.

Open a GitHub PR via API as a flowing graph — branch + push + create_pr + mergeable poll, with structural protection against pushing to main/master/etc. Use when a Claude Code or Claude.ai container needs to land changes on GitHub via the API (no git CLI required) and the prose "create branch, push, open PR, wait for mergeable" workflow keeps drifting under context pressure.

Disciplined, validation-gated revision of an EXISTING skill so each edit is a measured improvement rather than a guess. Use when editing, revising, or tuning a skill that already exists and there is evidence it underperforms (observed failures, drift, complaints) — invoke by name, or have versioning-skills / creating-skill defer to it before applying edits. Not for authoring a brand-new skill from scratch (use creating-skill) or one-off prose.

Orchestrates parallel API instances, delegated sub-tasks, and multi-agent workflows with streaming and tool-enabled delegation patterns. Routes by surface — native subagents in Cowork and Claude Code, httpx fan-out on claude.ai — and covers Gemini delegation via the Cloudflare AI Gateway on every surface. Use for parallel analysis, multi-perspective reviews, or complex task decomposition.

Skill-aware orchestration with context routing. Decomposes complex tasks into skill-typed subtasks, extracts targeted context subsets, executes subagents in parallel, and synthesizes results. Self-answers trivial lookups inline. No SDK dependency — uses raw HTTP via httpx. Use when tasks require multiple analytical perspectives, when context is large and subtasks only need portions, or when orchestrating-agents spawns too many redundant subagents.

Interactive codebase orientation for a HUMAN who wants to learn the code. Runs the same tree-sitting + featuring pipeline as exploring-codebases but synthesizes it into guided exercises and an HTML teaching artifact rather than an analysis document. Use for "orient me to this repo", "teach me this codebase", "walk me through this code", "learning orientation", or when someone wants genuine comprehension rather than a task completed. The audience is the test: if nobody is being taught and the goal is to get work done, use exploring-codebases instead.

Interpret video content visually by sampling frames into timestamped contact sheets that can be read as images. Use when: user asks what happens in a video; asks to summarize, describe, review, or QA video content or footage; asks about scenes, actions, people, or objects in a video; needs a storyboard-style overview of a clip; asks to find where something occurs in a video. Triggers on 'watch this video', 'what's in this video', 'summarize the video', 'describe the footage', 'contact sheet', 'storyboard', 'review this clip', 'find the scene where', 'scene detection', 'shot boundaries', 'detect cuts', 'dwell points'. For converting, trimming, or transcoding video, use processing-video instead.

Image processing toolkit awareness. Use when: user uploads images for manipulation, requests format conversion, batch processing, compositing, resizing, optimization, analysis, effects, metadata inspection, montages, animated GIFs, color correction, or any image-related task. Also use when working with screenshots, photos, diagrams, icons, or visual assets. Triggers on 'resize', 'crop', 'convert', 'compress', 'optimize', 'thumbnail', 'watermark', 'montage', 'collage', 'gif', 'sprite sheet', 'color space', 'metadata', 'EXIF', 'compare images', 'diff', 'overlay', 'composite', 'batch process', 'image analysis', 'histogram', 'blur', 'sharpen', 'rotate', 'flip', 'border', 'shadow', 'round corners', 'favicon', 'icon set'.

Audio and video processing with ffmpeg. Use when: user asks to convert, trim, merge, compress, or transcode video or audio files; extract audio from video; create GIFs or animated WebP from video; add subtitles or watermarks to video; change video resolution, framerate, or codec; normalize audio loudness; extract frames from video; concatenate clips; create thumbnails from video; strip or add audio tracks; convert between audio formats (MP3, AAC, FLAC, Opus, WAV); adjust volume; apply video filters; stabilize shaky video; generate waveform or spectrum visualizations; probe media file metadata. Triggers on 'ffmpeg', 'video', 'audio', 'transcode', 'MP4', 'MKV', 'WebM', 'MP3', 'AAC', 'FLAC', 'Opus', 'WAV', 'GIF from video', 'extract audio', 'add subtitles', 'video to gif', 'compress video', 'trim video', 'merge videos', 'normalize audio', 'framerate', 'resolution', 'bitrate', 'codec', 'ffprobe', 'waveform', 'spectrogram'.

Semantic Python code queries via a stdio LSP client driving pyright-langserver. Provides binding-resolved go-to-definition, find-references, hover types, type diagnostics, file symbol outlines, and project-wide symbol search — name resolution and type inference that tree-sitter and ripgrep cannot do. Use when you need to follow an import to a definition, find all real uses of a symbol (excluding same-named-but-unrelated ones), get an inferred type, surface type errors, outline a file, or search symbols across a project. Triggers on "go to definition", "find references", "what type is", "resolve this symbol", "symbol outline", "find symbol in project", "pyright", "type-check this file".

Query, filter, and transform Markdown structurally with mq — a jq-like CLI for Markdown. Use to extract headings/sections/code-blocks/links from .md files, build a table of contents, pull code blocks of a given language, slice or reshape LLM prompt/output Markdown, or batch-transform docs. Triggers on "extract sections from this markdown", "get all the code blocks", "jq for markdown", "mq", or any structural query over Markdown that grep/Read can't do cleanly.

Preprocesses photographed sheets of many business cards — slicing each into overlapping high-resolution tiles and de-glaring them with container tooling (OpenCV/ImageMagick) — then reads every card via cheap parallel temperature-0 API calls (Haiku or Sonnet) using a distilled extraction prompt, and writes deduped contact fields to a CSV. Use when a user has photos or scans holding multiple business cards per image, mentions glare or unreadable cards, batch card transcription, contact extraction, or wants to read many cards without an expensive in-conversation pass. Triggers on 'business cards', 'card scan', 'extract contacts', 'read these cards', 'card glare', 'too many cards per photo'.

Apply semi-formal certificate reasoning to code analysis — patch verification, fault localization, patch equivalence. Use when reviewing patches, hunting bugs across scopes, comparing fixes, or when code reasoning requires tracing execution across files/modules. Triggers on code review, bug localization, patch comparison, name shadowing, scope analysis, regression checking.

Analyzes an AI/ML publication — paper, preprint, article, technical blog post — and extracts what an enterprise AI engineer should do about it. Use when someone supplies a URL or document on RAG, embeddings, fine-tuning, prompt engineering, agents, or LLM deployment and asks "review this paper", "what do you make of this", "is this worth adopting", or "summarise the method and its limits". The subject matter must itself be machine learning.

DEPRECATED - Use browsing-bluesky skill instead. Sample and analyze Bluesky firehose to identify trending topics and content clusters. Use when user asks about "what's happening on Bluesky", "Bluesky trends", "zeitgeist", "firehose analysis", or wants to see real-time topic clusters from the network.

Binding-resolved Python symbol queries via pyright — every true caller (--refs), the real definition (--def), or an inferred signature (--hover) of a .py symbol, excluding the same-named false positives text search cannot tell apart. Use when a task needs ALL callers or users of a named Python symbol and grep would over-match. Python only, and only for the caller/definition question: measured 2026-07-04 on real issue-localization tasks, plain ripgrep tied or beat the semantic and indexed-regex tiers at 4-60x less wall-clock, so those tiers survive below without being a default.

Augmented vision tools for analyzing images beyond native visual capabilities. Use when tasked with describing images in detail, reproducing images as SVGs, identifying subtle features, comparing image regions, reading degraded text, or any task requiring careful visual inspection. Also use when the image-to-svg skill needs ground truth about colors, shapes, or boundaries.

In-process semantic search over text files or in-memory strings, using Gemini embeddings via the CF AI Gateway. Use when user wants fuzzy/conceptual search where exact-keyword grep would miss — "sessions discussing regulatory constraints", "code about retry logic", "notes mentioning burnout even if the word isn't there". Complements searching-codebases (regex/AST) and extracting-keywords (YAKE). Do NOT use when an exact string/regex match is what's wanted — grep/rg wins on speed and precision there.

Maintains a structured running-notes document during long work sessions. Use when the user says "session notes", "update notes", "start session notes", "show session notes", or when you recognize the current session has accumulated enough state (decisions, corrections, files touched, errors) that it risks being lost under context pressure. Stores notes as a procedure memory tagged [session-memory, active] so they survive container death within the same session thread.

Sort grocery lists by aisle order using store aisle sign photos. Build aisle maps from uploaded images, match items to aisles, and output optimized shopping routes. Use when users upload aisle sign photos, request grocery list sorting, want shopping trip optimization, need store layout mapping, or mention grocery list organization.

Portrait Mode for SVGs — foveated vectorization with 4-zone selective detail. Combines vision annotations, MediaPipe segmentation/landmarks, and optional saliency. Like phone portrait mode, but vectorized. Use when vectorizing a portrait or photo where subject detail should outrank background detail.

Exhaustive problem space exploration using the MIT Synthetic Neurobiology "tiling tree" method. Partitions a problem into MECE (Mutually Exclusive, Collectively Exhaustive) subsets recursively via parallel subagents, then evaluates leaf ideas against specified criteria. Use when users say "tiling tree", "tile the solution space", "exhaustively explore approaches to", "what are all the ways to", or request a MECE breakdown of a problem. Requires orchestrating-agents skill.

Maintain a structured task list for the current session. Use proactively when a request requires 3+ distinct steps, the user provides multiple items, or complex work benefits from explicit progress tracking. Storage persists via Muninn config across container death. Adapted from Claude Code's TodoWrite tool.

Reads the visual content of slides, pages, and images the way a human would, not just their embedded text. Use when a PPTX or PDF has image slides, screenshots, charts, scanned figures, or flattened-to-image layouts that the built-in pptx/pdf skills read as empty; when asked to transcribe, describe, OCR, or extract what is shown in an image, slide deck, or document page; or when embedded-text extraction returned little or nothing from a visually rich file. Triggers on 'read this deck', 'what's on these slides', 'transcribe', 'OCR', 'extract text from image', 'describe this chart/diagram', .pptx/.pdf/.png/.jpg with visual content.

Symbol-level navigation of a local checkout using tree-sitter ASTs. Answers where a symbol is defined, what lines it spans, which symbols a file exposes, what a directory holds, and where a name is referenced — every answer carries exact line ranges to feed straight into a scoped read. Use for "where is X defined", "who calls X", "find the function/class named", "what's in this file", "give me the line range for", "show me the source of", "list the symbols in", or before editing a file you have not read. Each invocation auto-scans and is self-contained. Not for first-encounter repo orientation (use exploring-codebases), for what a codebase DOES rather than what it contains (featuring), for binding-resolved Python caller sets (searching-codebases), or for literal text and regex matching (plain ripgrep).

Systematic research methodology for building comprehensive, current knowledge on any topic. Requires web_search tool. Use when questions require thorough investigation, recent developments post-cutoff, synthesis across multiple sources, or when Claude's knowledge may be outdated or incomplete. Triggered by "Research", "Investigate", "What's current on", "Latest info on", complex queries needing validation, or technical topics with recent changes.

File-upload bridge for Claude Code on the Web. CCotw has no native file mount; this skill creates a throwaway GitHub branch the user can drop files onto via the github.com web UI, then fetches them locally on the next turn. Use when the user wants to upload, share, or send files into the session, or when a task clearly needs files the user has on disk that aren't in the repo.

Browser automation via webctl CLI in Claude.ai containers with authenticated proxy support. Use when users mention webctl, browser automation, Playwright browsing, web scraping, or headless Chrome in container environments.

Check that a document's claims about code are actually true by reading the prose, the code, and the tests and reporting (or fixing) where they disagree. Use whenever the user wants to verify a README, guide, spec, or docstring still matches the code; whenever they mention documentation drift, doc-code sync, "is this still accurate", stale docs, or keeping docs/tests/code consistent; before publishing or merging a docs change; or as a periodic doc-accuracy sweep. The agent reads the prose's meaning directly — there is no claim-comment DSL to maintain. Pairs with TDD — the test suite is the deterministic behavioral gate, this skill is the semantic prose-vs-reality review.

REQUIRED for all skill development. Automatically version control every skill file modification for rollback/comparison. Use after init_skill.sh, after every str_replace/create_file, and before packaging.

Sentence-level and section-level craft for instructions Claude will actually follow, in any container — project instructions, a standalone prompt, or skill body text. Use when the text exists or is being drafted and the question is quality: "why is Claude ignoring this instruction", "rewrite this prompt so it works", "improve my project instructions", "make this instruction stick", "is this too prescriptive", "how much structure does Sonnet need versus Opus". Covers imperative construction, positive framing, strategic-versus-procedural calibration per model, extended thinking, complexity scaling, and a quality checklist. For deciding WHICH container the instruction belongs in, use crafting-instructions. For building, testing and packaging a whole skill directory, use creating-skill.