HomenShum

Turn "my app + an agent that demos" into "benchmarked, browser-verified, evidence-backed, prod-proven, and looping." Use when a (solo) founder wants to prove an AI agent works IN their real app — across all its UI surfaces — without cheating. Triggers: "set up proofloop", "proof-loop my app", "benchmark my agent's UI", "prove my agent works in prod", "run proofloop". The user's coding agent (Claude Code / Codex / Cursor) drives; the user steers by comment.

Use this skill to generate well-branded interfaces and assets for NodeBench AI, either for production or throwaway prototypes/mocks/etc. Contains essential design guidelines, colors, type, fonts, assets, and UI kit components for prototyping.

Migration runbook for @convex-dev/agent. Currently pinned to 0.2.10 because v0.6 requires a coordinated AI SDK v5 → v6 bump across many files. Use when bumping the agent or when CI shows the "args was removed in v0.6.0" error pattern.

Schema and authoring guide for dep-* skills — project-local migration runbooks for high-touch dependencies. Use when a dep major-bump breaks CI and you need to either apply an existing recipe or write a new one.

Pin runbook for @tiptap/pm. Pinned to 3.22.3 because 3.22.4 dropped the ./collab package export, which some transitive consumer (likely @blocknote or @tiptap/extensions) still relies on. Use when bumping tiptap or when build fails with "./collab is not exported".

Coordinated bump runbook for the vega + vega-lite + vega-embed ecosystem. The 5 → 6 jump fixed CVE-2025-59840 (XSS) and required updating the spec schema URL from v5 to v6. Use when bumping any of these three packages or when the vega XSS CVE alert fires.

Remediation runbook for the SheetJS xlsx package. Pinned to the SheetJS CDN release because the npm-published xlsx has 2 unfixed HIGH CVEs (ReDoS + Prototype Pollution) and SheetJS only ships fixes via their own CDN. Use when bumping xlsx, when CVE-2024-22363 or CVE-2023-30533 alerts fire, or when considering a migration to exceljs.

Run the repeatable operational loop on the NodeBench real-time chat pipeline or the report generator pipeline. Every pipeline change flows through: instrument → judge → persist → surface → measure → regress. Triggers: "pipeline", "operational loop", "run the standard", "measure the pipeline", "judge the traces", "chat pipeline change", "report generator change", or any edit to server/pipeline/* or convex/domains/product/diligence*.

Set, measure, and repair slide copy density against rendered layout capacity. Use during slide planning, generation critique, PPTX preflight, or repair when text is crowded, overflowing, illegible, or competing with visuals.

Detect and fail closed when a presentation generation route substitutes generic or unrelated content for the requested brief. Use before ranking, exporting, accepting, or promoting any model-generated deck.

Require editable, evidence-serving visual artifacts instead of prose-only presentation slides. Use when planning or validating charts, diagrams, screenshots, timelines, comparisons, formulas, code, and other dominant slide primitives.

Apply evidence-backed Claude Sonnet safeguards to NodeSlide story direction, critique, and deck generation. Use when Sonnet proposes a narrative, critiques candidates, or directly composes presentation slides.

Bound Gemini Flash presentation work to independently verified capabilities and prevent generic fallback output. Use when a Gemini Flash route is selected for deck generation, extraction, classification, or other presentation tasks.

Apply evidence-backed safeguards when Gemma 4 26B Free plans or generates a bounded NodeSlide artifact. Use for free-router artifact planning, source-bound visual operations, or zero-cost candidate generation.

Apply evidence-backed Kimi K3 safeguards to NodeSlide deck execution and repair. Use when Kimi K3 is selected for full-deck generation, scoped slide editing, artifact construction, or validation-driven repair.

Diagnose and repair text overflow, internal collisions, and crowded regions inside rendered presentation slides. Use after geometry, density, screenshot, or PPTX validation reports overlap or likely line-capacity failure.

Run NodeRoom's repeatable SEO and landing-flow QA workflow. Use when auditing or improving public SEO pages, metadata, sitemap/robots, Search Console metrics, Playwright landing journeys, Lighthouse/Core Web Vitals evidence, Gemini video judging, or Nebius-first text/JSON judging without scraping Google rankings.

Keep the NodeRoom agent wiki synchronized with room artifacts, upload references, public chat, traces, and agent run evidence. Use when updating the in-app wiki, docs/AGENT_WIKI.md, interview documentation, or any agent-generated table-of-contents page after file uploads, research updates, finance spreadsheet edits, trace changes, or agent workflow changes.

Consolidate a multi-entity ledger (QuickBooks, Xero, NetSuite, or any per-entity JSON) into one close, reconcile it against a golden reference, catch mis-keyed entries that cross-foots miss, and gate on "shippable without review?" before numbers go to a board, investor, lender, or downstream report. Use for month/quarter close, multi-entity roll-ups, or any time figures from more than one accounting source must agree before they are trusted.

Record a live browser walkthrough, render a README-quality GIF/MP4, run Gemini video/UI review, fix defects, and update proof docs. Use when a feature needs demo evidence, README media, product walkthrough QA, or a reusable solo-founder verification loop.

Produce an evidence-backed software explainer episode from a storyboard — demo video, README GIF + MP4, code-diff story, before/after product narrative. Use when the user asks for a demo video, an explainer, an episode, social/YouTube cuts of a feature, or "turn this change into content". Builds on the readme-walkthroughs skill (capture layer) and the episodes/ manifest structure.

Produce animated WALKTHROUGH GIFs for the README — never static "hero shot" GIFs. Use this skill whenever the user asks for README demos, product GIFs, demo recordings, "show the feature working", marketing clips of the app, or to refresh/regenerate the walkthrough GIFs after a UI change. It drives the LIVE app with Playwright through a versioned spec, captures clean per-state frames + cursor click targets, and renders a Remotion composition with a gliding cursor, click ripple, step captions, and a progress bar. Also use it when adding a NEW feature that deserves a README demo (add a spec, capture, render, embed).

Run only the synthetic, offline FounderQuest-RL environment and preserve protected rewards, blind heldout separation, and human/professional authority boundaries.

Turn a business pain point, product purpose, hackathon idea, judging rubric, or required sponsor stack into a researched, scaffolded, evaluated, live-tested, browser-proven agent application. Use for empty-directory builds or safely adopting NodeKit into an existing project.

Capture a substantial code change, architecture migration, benchmark, release, or completed hackathon application as an editable, evidence-backed presentation. Use when Codex must explain what changed, assemble a judge or reviewer deck, keep presentation material current during development, or drive NodeSlide through its package, CLI, MCP, or host adapter.

Dogfood and regression-test an agentic application from the rendered user surface through its runtime, tools, durable state, artifacts, and proof. Use after NodeKit creates or adopts an application, before production or hackathon presentation claims, and whenever a major agentic UI workflow changes.

Benchmark-driven development for AI agents — turns "I have an idea / prototype / half-built app and an agent that demos but does not hold up" into "an agent that completes real benchmark tasks IN the live app, browser-verified, without cheating." Runs the loop discover → benchmark → setup → build → adapter → verify → iterate under four non-negotiables (held-out · no answer-keys · in-app transfer · honest provenance), with the honest-lane clean-probe rule, a local-first memory substrate, and a Design Bridge for UI. Use when a (solo) founder wants to build or validate an AI agent for their app. Triggers: "build the agent layer for my app", "benchmark my agent", "prove my agent works in production", "which benchmark fits my agent", "make my agent pass SpreadsheetBench/BankerToolBench/SWE-bench in my app", "my agent demos but fails on real tasks", "eval my agent honestly". The user's coding agent (Claude Code, Codex, OpenClaw, Hermes, Trae) drives; the user steers by comment. This single skill IS the suite — it r

A reusable ProofLoop skill for making coding agents prove completion with external receipts, anti-gaming gates, held-out tasks, and live or stratified browser evidence before they claim done.

A reusable skill for taking one founder product goal through domain discovery, benchmark selection, local setup, app build, live workflow verification, and repair until the demo has receipts instead of vibes.

Run only the synthetic, deterministic FounderQuest-RL environment and preserve protected rewards, heldout separation, and human/professional authority boundaries.

Run a measurable keep-or-revert research loop that remains observable, steerable, resumable, and reproducible. Use when an experiment has a fixed objective metric and a human needs to intervene without restarting the agent.

Safely inspect a synthetic SMB lending file and propose one reviewable missing-document request without making a lending decision.

Use this skill when changing prompts, tools, models, or agent orchestration so the resulting behavior is evaluated against the synthetic preferred-response golden dataset before sign-off.

Use this skill for store-manager and regional-manager guidance so answers stay operationally specific, policy-grounded, and escalation-aware.

Use this skill for framework or SDK integration work involving Gemini, LangChain, Deep Agents, Convex, Next.js, or evaluation infrastructure so implementation decisions stay grounded in current official docs instead of stale examples.

Full QA→fix→verify loop using retention.sh tools. Crawls the user's app, finds issues, suggests fixes, re-crawls to verify, and shows token savings. Use when the user says "QA my app", "find bugs", "check my site", "run QA", or "test my app".

Full onboarding for retention.sh — checks prerequisites, installs MCP tools, verifies connection, runs first crawl, shows results. Use when the user says "set up retention", "install retention.sh", "get started with QA", or "connect MCP tools".

Capture, distill, and replay coding agent workflows

Distill any person into a reusable Claude Code skill. Feed it chat logs, docs, emails, git history, and get back a digital twin that thinks, decides, and communicates like them.

Produce storyboarded product proof assets for a web app (built/tested on Streamlit, works for any browser UI): MP4/GIF walkthroughs, screenshots, and readable state captures that show every important UI state of a feature. Use when the user wants README demos, release proof, or side-by-side comparisons that walk through the end-to-end flow rather than showing a single final-state "hero shot". Pipeline = Playwright capture → generated step data → Remotion render → ffmpeg and exported artifacts.

Make every code submission (commit, branch, PR) easier to read for the next person picking it up — human or AI. Forces per-surface changelog entries (one append-only file per page/component/server module/db table/integration/script, not one undifferentiated git log) and (when UI changed) a verified demo recording with both DOM checks and Gemini video analysis. Use any time you commit, push, open a PR, hand off a branch, or finish a task that touched user-facing code.

Capture proof images from real source files and running app sessions for codebase walkthroughs, trace UIs, onboarding demos, README media, and agent-verifiable UI evidence. Use when replacing generated screenshots, recompositions, mocked code views, synthetic UI callouts, or stale walkthrough assets with NodeTrace code-browser captures and Playwright/browser captures from the live app.