AI Agent Skills

Create, edit and check Word (.docx) files with Python and python-docx: real styles, tables, headers, footers and page-number fields, edits that keep the original untouched, and a re-open check of the result.

A six-phase loop for hard bugs: build a reproduction that fails on exactly this bug, minimise it, test ranked hypotheses, then fix with a regression test.

Write Markdown documents with Mermaid diagrams for workflows, schemas, timelines and architecture: syntax guides for 21 diagram types, document templates, accessibility metadata and render checks.

Make an interface carry any language without redesign: message catalogs, locale-aware formatting, layouts that survive different text lengths, CJK and right-to-left support, and language switching with proper URLs.

Build complex, self-contained React web artifacts with TypeScript, Tailwind CSS, and shadcn/ui.

Write a structured spec before any code: surface assumptions, cover objective, commands, structure, style, testing and boundaries, reframe vague asks as testable success criteria, and stop for human approval before planning.

Build an animation from scratch in the right order: should it animate at all, what purpose, which tool, which properties, which curve and duration, how it interrupts and exits. It writes the code and refuses to animate what should not move.

Turn the papers and notes you provide into a structured review: themes, agreement and disagreement, methods, gaps and a source map, with no invented citations and the strength of evidence stated for each theme.

Builds a throwaway prototype to answer one design question: a clickable HTML model for logic, or several UI variations to compare.

Handle returns, refunds, damaged or lost parcels, chargebacks and angry messages fairly and with little loss: policy points, a case-handling process, reply drafts, escalation rules and a way to find why returns are high.

Decide whether self-serve, sales-led or a hybrid fits your product, then measure channel economics, activation and growth with simple rules.

Plan scalable, data-backed SEO pages while protecting against thin content and indexation risk.

Stress-test a component with the worst realistic data: long names, unbreakable emails, one-letter names, missing fields, 1,284 members, empty lists. It reports what broke, why, and the fix for each, and stops before changing anything.

Teach the agent to design like a premium agency: exact fonts, spacing, shadows, nested card structure and motion curves that make a site feel expensive, plus a ban list of the cheap AI defaults.

Create, edit and validate draw.io diagram files with correct mxGraph XML: flowcharts, architecture, sequence, ER and UML class diagrams, with templates and helper scripts.

Create, edit and check PowerPoint (.pptx) decks with Python and python-pptx: real layouts and placeholders, tables, charts, speaker notes, safe edits that keep the original and a re-open check of slides and notes.

Create, edit and check Excel (.xlsx) workbooks with Python and openpyxl: live formulas instead of pasted values, separate inputs, formatting, charts, safe edits that keep the original, and honest checks of what was and was not calculated.

Turn study material you provide into flashcards and a graded quiz for active recall, with sources, explanations, common mistakes and a spaced-review schedule, and nothing added from outside your material.

Handle common PDF jobs in Python: extract text and tables, merge, split and rotate pages, add watermarks, fill forms and generate reports. Never overwrites the original and re-opens the result to verify pages and text.

Upgrade an existing website or app to premium quality without rewriting it: scan the stack, audit typography, color, layout, states and content for generic AI patterns, then apply targeted fixes in priority order.

Choose how to produce marketing video: programmatic (Hyperframes, Remotion), AI generation (Veo, Runway, Kling), AI avatars (HeyGen, Synthesia) or editing and repurposing, with model comparisons, prompting tips and ready workflows.

Debug a failing GitHub Actions check on a pull request: find the failing checks, pull the logs, summarize the cause, draft a fix plan, and change code only after you approve.

Write B2B cold emails and short follow-up sequences that read like a thoughtful human: personal, brief, one low-friction ask, with subject lines and benchmark data.

Describe a motion effect in your own words and get its exact name: "the bouncy thing when a popover opens" is Pop in, "the iOS rubber-band scroll" is Rubber-banding. Handy for briefing a designer or prompting an AI.