AI Agent Workflow Collection

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 (Haiku/Sonnet/Opus) and effort level each subagent gets, when to cascade cheap-first behind a verifier, and how to run improvement loops safely (evaluator-as-selector, stop on regression). Use when spawning subagents via the Agent or Workflow tools, when fanning out more than a handful of agents, or when the user asks which model or effort a task should route to. Grounded in measured calibration data (references/calibration-2026-07-15.md), not vibes.

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.

Ranked content search over any text corpus using BM25 (via xhluca/bm25s). Corpus-agnostic: works on cloned repos, project knowledge stores, uploaded files/archives, and any local directory. Stateless — builds an in-memory index each invocation, no cache, no persistence. Use when you need ranked multi-word content search beyond grep, or when picking the "most relevant files for these terms" across a corpus. Triggers on "rank these documents", "search this corpus", "find content about X", "which files are most about Y", or multi-word concept queries against a known body of text.

Generate progressive disclosure indexes for GitHub repositories to use as Claude project knowledge.

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.

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.

Build and cache a personalized container environment from a Dockerfile-like spec. Supports both single-layer (one Containerfile -> one cached tarball) and multi-layer composition (compose [base, scientific, mojo, ...] into one container with each layer cached independently). Use when the user mentions "container layer", "Containerfile", "custom container", "environment setup", "cache my installs", "uv shim", "composable layers", or wants to persist package installations, skills, or environment config across ephemeral sessions. Also triggers when the user asks to snapshot, restore, or rebuild their environment, or wants to capture ad-hoc package installs into a reproducible spec.

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.

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).

Create video from prompts by overseeing multi-clip AI generation end to end: write a shot list, generate each scene with Google Veo (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 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', '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.

Text-prompted image zone detection using TIPSv2 B/14 on CPU. Produces focustargets / focusedges 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.

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).

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.

DAG workflow runner that encodes control flow in code, not prose. Use when a procedure has 3+ steps with branching, retries, or validation that must be enforced — gates as when=, edge contracts as validate=, predicate loops as retryuntil=. The runner owns the graph; the LLM provides leaves. Also covers parallel execution, checkpoint resume, detached side-effects.

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 — checks that block 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 result is about to be published, acted on, or merged and a plausible-but-wrong value would survive review; when asking whether an existing test suite, linter rule, or check could actually fail; and when a check suite passes first try, passes suspiciously often, or was written by the same process that produced the thing it checks. Triggers on "verification loop", "calibration gate", "can this check fail", "known-bad", "negative control", "sanity check my results", "is this test actually testing anything".

Generate and maintain lat.md/ knowledge graphs with bidirectional source code anchoring. Docs link to code via [[src/filesymbol]] wiki links; code links back to docs via @lat: comments. lat check validates both directions, catching drift when either side changes. Use when user says "lat.md", "lattice", "knowledge graph", "document this codebase", "add back-links", or wants cross-referenced architecture docs anchored to source code symbols.

Break out of a locked problem frame by picking one disciplined move — reframe, provocation (Po), random stimulus, SCAMPER, inversion, perspective shift, or constraint play — 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". Complements challenging (which evaluates) and convening-experts (which synthesizes viewpoints); this skill generates distance, not judgment.