webgrip

Write, amend, or supersede Architecture Decision Records in MADR 4.0.0 — locate or bootstrap the ADR directory and registry index, numbering, frontmatter status/date semantics, Confirmation checks, append-only dated history, supersession mechanics, and a bundled CI-ready consistency validator. Use when recording or ratifying an architecture/technology decision, creating an ADR, updating an ADR's status after a rollout/revert/supersession, adopting ADRs in a repo that has none, or wiring ADR validation into CI. NOT for design exploration (that's an RFC/design doc) or operational procedure (runbook).

Fixes, cleans up, reviews and writes CLAUDE.md and AGENTS.md agent instruction files and agent config. Use when cleaning up or slimming down a CLAUDE.md or AGENTS.md, when it is too long, bloated or ignored by Claude, when Claude keeps doing something the file forbids, when CLAUDE.md and AGENTS.md are duplicated or drift apart, when creating one, when composer install, boost:install or another generator wipes hand-written instructions, when rules in ~/.claude leak into other projects, when reviewing a PR that changes CLAUDE.md, AGENTS.md, .mcp.json or .claude/settings.json, when testing whether an instruction change works, when setting up cloud, CI or headless agents, or when deciding whether a rule belongs in CLAUDE.md, a skill, a hook, CI or docs. Covers .claude/rules, CLAUDE.local.md, .cursor/rules, copilot-instructions.md and Laravel Boost across Claude Code, Codex, Cursor, Copilot, opencode and OpenHands; ships measuring, security and A/B probe scripts.

Write, edit, and syndicate technical blog posts through a full pipeline — evidence gathering, outline and hook workshop, drafting, single-dimension edit sweeps including a de-AI-ify pass, fresh-reader testing, and per-platform adaptation (Substack, LinkedIn, Hacker News, dev.to, Reddit, Lobsters, Medium, Hashnode). Use when writing or drafting a blog post, article, or writeup; turning project notes, an incident, a build log, or a TIL into a post; editing a draft to sound human and land with developers; picking a title or opening hook; adapting or cross-posting a post for LinkedIn, Hacker News, Substack, dev.to, or Reddit; planning where and when to publish; or reviewing a post before it ships.

Define, maintain, and publish a project's domain language (ubiquitous language) — the shared vocabulary of terms, entities, relationships, business rules, and events — and generate specs and documentation from it. Use when the user wants to build a glossary, define domain terms or concepts, model entities and their lifecycles, capture business rules, resolve naming ambiguity, write a feature spec grounded in domain vocabulary, generate domain documentation or diagrams, check the health/quality of a domain model, or audit a codebase/docs for inconsistent terminology. Trigger even if the user doesn't say "domain model" — phrases like "we keep calling this thing three different names", "grill me about my domain", "document our concepts", "define what an Order means here", "score our glossary", or "turn our terminology into docs" all apply.

Ships static sites, Workers and private pages on Cloudflare's free edge tier with wrangler - Workers static assets over Pages, routes vs custom domains, www-to-apex redirects, DNS as code (DNSControl, OpenTofu, external-dns), Zero Trust Access gating and expiring per-recipient share links, preview deploys, least-privilege CI tokens, free-tier limits, a wrangler config linter and live-site prober, and when Netlify, GitHub Pages, Bunny or an EU host wins instead. Use when deploying or migrating a site to Cloudflare Workers or Pages; writing or reviewing wrangler.toml or wrangler.jsonc; a Cloudflare-hosted site or its www returns 522, times out or deploys green but serves nothing; wiring a custom domain, DNS records or a redirect; putting Cloudflare Access, a login or Zero Trust in front of a site, staging, preview or homelab service; sending a client a private or expiring web page; choosing free static, edge or serverless hosting; or asking what Cloudflare's free plan covers.

Designs, critiques and builds distinctive, memorable web pages from the subject's own evidence - a claim ledger, one thesis object carrying the visual identity, a still composition that holds with motion off, parallel design directions, a motion contract per effect, and a validation matrix plus a source scanner for reduced-motion, zoom, focus, layout-shift, template-look and fake-proof failures. Use when a page looks generic, templated, flat or AI-made, or when asked to make it pop, stand out, feel premium, bolder, memorable or on-brand; redesigning or art-directing a landing, product, launch, campaign, portfolio, event, docs-entry or open-source project page; turning a product, report or dataset into a scroll story or interactive explainer; adding scroll-driven animation, view transitions or 3D to a site; reviewing motion or visual design for accessibility and honesty. Not for routine copy edits, single-component tweaks, or mockups with no build.

The secrets-handling floor for Webgrip repos, and the enforcement hook behind it. Use when writing or wiring secrets, editing SOPS files, handling API keys/tokens/credentials in config or Helm values, or when an edit is blocked with a BLOCKED secret message. Explains why the block fired and the compliant path (SOPS + value wiring).

Mine durable, reusable knowledge out of Claude conversation threads with a three-phase workflow — distill each thread into a self-contained digest, consolidate many digests into one deduplicated truth-checked knowledge set, then synthesize repo docs, CLAUDE.md rules, new-skill candidates, and memory updates from it. Use when distilling/harvesting durable learnings from a Claude thread, extracting a thread digest, consolidating multiple digests into one knowledge set, running the integration thread, or turning collected learnings into docs/skills/memory.

Detects, removes and prevents the tells of AI-generated writing in English and Dutch: negative parallelism ("not X but Y"), colon reveals, rule of three, em-dash density, hollow intensifiers, vague attribution, chatbot residue, and for Dutch the translationese layer of English-shaped Dutch. Four modes, a register tolerance table, voice matching and a regex scanner. Use when text reads like ChatGPT or Claude, sounds like AI, needs humanizing or de-AI-ifying or should sound like a person wrote it; to audit, flag or detect tells without editing, when asked "is this slop" or "does this read as AI"; to scan docs or copy for AI tells; and when drafting public-facing copy that must not read generated: a meetup or event description, a LinkedIn or social post, an announcement, a newsletter, marketing or landing-page copy. Dutch: klinkt als AI, niet zo AI, maak het normaal of menselijker, is dit slop, ChatGPT-taal, AI-woorden, jeukwoorden, ontslop.

Define, groom and facilitate KPI sets for teams, organizations, projects and individuals — complete KPI definitions (formula, data source, owner, baseline, target, the decision it changes), Goodhart speed/quality pairs, timeboxed KPI-choosing workshops the team owns, mapping KPIs onto existing dashboards, review-and-retire cadence, and the ethics gate for person-level metrics (SPACE, GDPR/works council). Use when defining or choosing KPIs or metrics for a team, organization or project, preparing or facilitating a KPI session or metrics workshop, reviewing or grooming an existing KPI set or dashboard for steering value, pairing a speed metric with a quality metric, deciding whether or how to measure individual people, or when asked what should we measure, which KPIs matter, or to put targets on a dashboard.

Writes and ships LinkedIn posts that actually reach people: the 2026 ranking mechanics (link penalty, comments over likes, dwell time, personal profile over company page), the plain-text formatting the editor will not destroy, accessible emphasis, and where to publish. Use when drafting or editing a LinkedIn post, announcement, launch, call for speakers or hiring post; when asked why a post got no reach; when choosing between a personal profile and a company page, or whether to repost; when a pasted post lost its blank lines or shows underlined links; and for the posting schedule. Dutch: LinkedIn-post schrijven, aankondiging, bereik, waarom krijgt mijn post geen reacties, op mijn profiel of op de bedrijfspagina.

Run any ticket board as product owner — Vikunja, ClickUp, or any tracker. Refinement to the Definition of Ready, agent-ready gating for AI-executed work, review-capacity WIP limits, the Definitions of Mergeable and Done, and flow metrics. Use when creating, updating, closing or listing tickets, "make a ticket for X", "maak een ticket (aan) voor", "wat staat er op het bord", asking what to pick up next, refining or grooming a backlog or klantbacklog, "is dit ticket klaar voor de sprint", writing acceptance criteria or a verification, prioritizing or triaging, ordering pick-up, moving a ticket to another status, re-grooming a ticket rejected in testing or acceptance, deciding whether a PR or MR is mergeable ("mag dit gemerged worden"), running a board-health or WIP audit, measuring cycle time or deriving an SLE, preparing tickets for an AI coding agent, splitting an oversized ticket, handling a stale ticket, or when a ticket is just a product name. NOT for agile-theory questions — this runs a live board.

Designs, critiques and builds product interfaces people use repeatedly (apps, dashboards, admin tables, forms) from evidence, not taste - job stories, a frequency and error-cost task model, a state matrix before screens, evidence-tiered pattern rules, a WCAG 2.2 and EU deceptive-pattern floor, heuristic evaluation and KLM, and a UI source scanner. Use when designing, building or reviewing an app screen, area or flow, such as account settings, profile, password or delete-account pages; improving the usability of a form, table, dashboard, navigation, search, empty or error state or onboarding; adding an AI feature to a product (suggested replies, drafts, an assistant, agent actions); a subscription cancel or churn-reduction flow with retention offers, a consent or checkout flow; design tokens or components; UX copy or error messages; an app accessibility audit; or when asked why a screen is confusing or slow. Not for marketing-page art direction (expressive-design) or chart construction (dataviz).

Evaluates an external protocol, standard, or ecosystem tool (MCP, A2A, ACP, AG-UI, AHP, an emerging spec, a competing product) against a project's own architecture seams using a parallel research fan-out, then lands a verdict-led report in the project's survey docs with adoption recommendations and named re-evaluation triggers. Use when asked how a protocol relates to our stack or components, whether a spec, standard or tool matters to us or is worth adopting, to do deep or exhaustive research on a spec URL, to compare a new protocol against the one we already bet on, or for a protocol or market-landscape survey.

Refactors and splits agent-maintained code so coding agents spend fewer tokens on future changes, and proves it - git-hotspot targeting with agent share, a repeated-run harness that measures the same change before and after from the agent's own usage stream, Fowler-catalog refactorings ordered by what shrinks the agent's read set, tool-driven execution with a per-step gate, and a break-even model per token class. Use when agents burn tokens or keep re-reading a large file; when asked to split a big file or class so agents spend less; whether refactoring will cut agent token cost or pay for itself; to measure what a change costs an agent before and after a refactoring; which files or modules to refactor first so AI agents spend fewer tokens; to reproduce the martinfowler.com refactoring-economic-benefit experiment; or to plan refactoring an AI-built codebase. Not for behaviour changes or general code review.

Gets static sites on Cloudflare's edge found in Google and Bing and cited by ChatGPT, Claude, Perplexity and AI Overviews - crawler access and Cloudflare AI bot settings, one canonical URL per page under wrangler html_handling, robots.txt and sitemaps, titles, snippets, structured data and link previews, answer-first content, an evidence-graded verdict on GEO and llms.txt, and a build plus live-site scanner. Use when a site does not show up in Google or in AI answers; to improve the SEO or search visibility of a static, Astro, Hugo or Eleventy site; when checking robots.txt, a sitemap, canonical, hreflang, noindex, redirects or 404s; to make a site findable or citable by ChatGPT, Perplexity, Claude or AI Overviews (GEO, AEO, llms.txt); when deciding whether to block AI crawlers or training bots; when link previews, og:image or social cards are broken; or before launching a site. Not for paid search ads, page speed (site-performance), page design (expressive-design) or deployment (edge-hosting).

Places a secret at the right level of the Webgrip estate and names the manifest that puts it there, per homelab-cluster ADR-0055 - floor (SOPS), vault (OpenBao), cluster (External Secrets), bridge (Forgejo Actions and Cloudflare Worker secrets), short-lived (OIDC, dynamic credentials, per-run tokens), person (a laptop). Use when adding, wiring, rotating, or asking where to put an API key, token, password, credential, PAT, or client secret; when a CI job, workflow, pod, Worker, or developer machine needs a secret; when someone proposes a new sops.yaml, a Forgejo org secret, a Keychain entry, a .env file, or a bao kv put; or when deciding whether a value is entropy, provided, or mintable.

Makes static sites on Cloudflare's edge fast for real users and proves it with field data - Core Web Vitals (LCP, INP, CLS) diagnosed by subpart and phase, ranked fixes for hero images, fonts, JavaScript, third parties and layout shift, edge caching, speculation rules and the back/forward cache, RUM and APM choices (RUMvision, Cloudflare Web Analytics, a free web-vitals beacon into Workers Analytics Engine, Workers Observability) with EU privacy, Lighthouse CI budgets, and a build plus live scanner. Use when a site is slow or feels slow; when LCP, INP or CLS fail in PageSpeed Insights, Lighthouse or Search Console; to improve a Lighthouse or PageSpeed score; when setting up real user monitoring, RUMvision, CrUX, synthetic checks or performance alerts; when adding a performance budget to CI; when a hero image, web font, cookie banner or third-party script slows a page; or to check whether a site is fast enough. Not for search rankings (search-visibility) or deployment (edge-hosting).

Report which agent skills actually get used, from the local usage log the bundled hook writes (~/.claude/skill-usage.jsonl; opencode twin at ~/.config/opencode/skill-usage.jsonl). Use when asked which skills are used or unused, for a skill usage report, to find dead or undertriggering skills, or to decide what to consolidate or remove from a skills estate.

Author, edit, and audit agent skills (.claude/skills/*/SKILL.md or plugin skills) for token-efficient, high-trigger-accuracy LLM ingestion. Use when creating a new skill, refactoring/auditing an existing one, writing or tuning a frontmatter description, a skill that never fires or fires only sometimes, probing whether a skill triggers, building a skill from articles or research, cutting a skill's token weight, splitting a skill into supporting files, or deciding what belongs in a skill vs CLAUDE.md vs memory vs a runbook.