🛠️ Skill Audit ☀️ Light

replace the `description:` line with: description: "Use cloudflare pages deploy (this workspace) to we have Hermes."* If the existing cron + skills + cloudflare-pages-deploy pipeline can do the job (daily job, weekly HTML rebuild, deploy to `<slug>.pa…."

tighten the body — a working tighter version: # Cloudflare Pages Deploy (this workspace) ## TL;DR **Default to existing infra before reaching for new tools.** Morsy's verbatim rejection: *"This is BS. We have Hermes."* If the existing cron + skills + cloudflare-pages-deploy pipeline can do the job (daily job, weekly HTML rebuild, deploy to `<slug>.pages.dev`), propose THAT, not a new external service (n8n, Make.com, separate cron → email bridge, etc.). Locked 2026-10-15. **Reuse an existing project before creating a new one.** Cloudflare Pages has a hard per-account project cap (free = 100, paid = 500), and the easy reflex of "one HTML file → one new project" burns through it. Before `wrangler pages project create`, run the project-reuse check below. Multi-file sister reports, dashboards, recurring cron outputs, and per-iteration artifacts from the same workflow all belong in ONE existing project with sibling files, not in N new projects. See Pitfall #27. **wrangler 4.105 auto-creates projects — one step is the default.** ```…

replace each internal reference: line 6: # Cloudflare Pages Deploy line 11: Default to existing infra before reaching for new tools. If the existing cron + skills + cloudflare-pages-deploy pipeline can do the job line 15: separate cron → email bridge, etc.). line 39: Always stage deploys in a non-Google-Drive local dir (e.g. ~/.cache/<name>/ or any tmpfs path) before running wrangler pages deploy. line 73: cd <workspace>/milestr milestr --agent <agent> export

replace the `description:` line with: description: "Generate H's daily 5-card neurological-surgery and clinical-research news digest (rolling 48h) and deliver via email, LinkedIn draft, and WhatsApp group."

tighten the body — a working tighter version: # H's Daily Brief — Three-Channel Delivery Cron This umbrella covers the recurring daily brief for **H (Henda Cheikhrouhou)** — Morsy's wife, a clinical researcher at the **UCSF Department of Neurological Surgery**. Every morning at 7am PT a cron run produces a five-card digest of neurological-surgery + clinical-research news from the **last 48 hours**, then delivers it through three channels: 1. **Email** — HTML newsletter via himalaya gmail SMTP (MML template format; see pipeline step 5) 2. **LinkedIn proposal** — ~80-word draft for H to post herself (the cron does **NOT** auto-post to LinkedIn) 3. **WhatsApp** — forwards the LinkedIn proposal to the family comms group This is **not** a Slack-format digest. Cards use HTML with inline CSS, not Slack markdown. The `executive-brief` umbrella's Slack format rules (§5) do **not** apply here. ## Recipient constants (do not change without user input) | Field | Value | |---|---| | To | `henda.cheikhrouhou@ucsf.edu` | | CC | `morsy.chei…

replace each internal reference: line 17: <recipient email> line 18: <cc email> line 19: <verified sender email on a Resend-authenticated domain> line 20: <permitted whatsapp group jid> — name "<hub group name>" line 21: <personal whatsapp jid> — NOT permitted for this cron (personal DM, off-limits as of <date>)

consolidate the smaller skill into `agent-communication-style`; archive the loser's directory under `~/.hermes/skills/_archive/<name>-merged-<ts>/`

replace the `description:` line with: description: "Open the source-of-truth file before answering count, list, or status questions across items. Read disk, count, reply. Never reconstruct from memory. Pair with communication-style for render rules."

tighten the body — a working tighter version: # Source-of-Truth-First (SOTF) When Morsy asks a question whose answer is a **count, list, status, or enumeration across items**, the answer lives in a file — not in agent memory. Open the file, count, then reply. Do not reconstruct from session_search when a known source-of-truth file exists. **Financial-model variant (2026-09-24 nancy SCI session):** the same discipline applies when **building** models, not just answering questions. When constructing any model with input rows (property values, acquisition basis, mortgage balances, tax brackets, rental figures), the first reflex must be to look up actual values in the user's existing documents — Google Drive (`~/Library/CloudStorage/GoogleDrive-*/My Drive/priv/TresorPublic/` for French tax filings, prior returns, notaire estimates), `US_Tax_09_<year>_-_Itemized_List.xlsx` for prior-year property values and acquisition dates, BNP/SocGen amortization CSVs in `<year>-fr-revenues/`, `SeLoger.pdf` for owner self-estimates — rather than …

replace each internal reference: line 3: Use when the user asks 'is X tracked?', 'what have we worked on?', 'show me everything about Y', 'are all N things done?', or any question whose answer is a count, list, or status across multiple items. The rule: open the relevant file (milestone tracker, project README, git log, session DB) FIRST, count or extract the answer, THEN reply — never reconstruct from memory, never substitute analysis for an answer. This skill covers the retrieval discipline that must happen before the shape decision in agent-communication-style. line 5: version: 1.6.0 author: "see CHANGELOG.md" line 16: Construction-time variant: the same discipline applies when building models, not just answering questions. When constructing any model with input rows (property values, acquisition basis, mortgage balances, tax brackets, rental figures), the first reflex must be to look up actual values in the user's existing documents (tax filings, prior returns, bank amortization statements, owner self-estimates) — rather than generating plausible-looking defaults. line 18: Deepest construction-time cousin — "Copied from a derived summary, didn't verify upstream": line 36: Origin:* see references/financial-model-derived-summary-2026-09-27.md.

replace the `description:` line with: description: "Scan 15 named sources daily for AI bubble warnings, detect trigger events, and rebuild the heat-map dashboard with cumulative signal tiles."

tighten the body — a working tighter version: # AI Bubble Tracker A daily tracker of named experts and institutions calling for an AI bubble burst in 2027-2028, with automated watches for the triggers they care about. ## When to use - "What did the AI bubble tracker find today?" - "Did Burry / Grantham / Dalio post anything new?" - "Run the AI bubble scan" - "Rebuild the AI tracker dashboard" - "Update the AI bubble tracker predictions" ## Data files (all in `~/.hermes/workspaces/nancy/ai-bubble-tracker/`) - `data/predictions.json` — the 11 named sources + 5 triggers - `data/events.json` — append-only log of detected events - `data/signal_history.json` — per-day `{source_id, severity}` rows; drives the all-time heat count - `data/seen_ids.json` — dedup key (URL) → first-seen-timestamp; used by the daily cron to avoid re-logging the same article - `index.html` — last-rendered dashboard (dark default, light toggle) ## Scripts - `scripts/scan.py` — read-only post-processor: prints Slack-ready summary - `scripts/render.py` — r…

replace each internal reference: line 20: ## Data files (all in <workspace>/ai-bubble-tracker/) line 31: uses a parent z-index + ::after z-index + ancestor overflow:visible pattern for the popover. line 42: (deployed via wrangler pages deploy; project slug and hosting account configured per the owner's Cloudflare setup.) line 85: ## Sources watched (15 total) line 90: AI-credit cycle (4):

replace the `description:` line with: description: "Design and operate a multi-cron fleet (2+ LLM-driven jobs handing off via shared state files). Covers pipeline shape, state-handoff JSON, per-cron prompt contract, silent-worker trap, and audit recipe."

tighten the body — a working tighter version: # Cron Fleet Stewardship A "cron fleet" is 2+ scheduled LLM-driven jobs that share work via files on disk. Each cron writes an artifact (state JSON, draft, etc.) that the next cron reads. Common shapes: `daily-scan → pr-drafter → release-drafter` for repos; `scan → enrich → notify` for content feeds; `poll → classify → dispatch` for support workflows. This skill governs the design, audit, and operation of fleets. The single-cron patterns live in class-level skills (`inbox-watch`, `rss-watch`, `gsc-watch`, `ticket-watch`, `executive-digest`, `community-scout-scan`); this skill governs the **fleet** — the coordination contract between them. ## When to use - Adding a new scheduled job to an existing fleet - Designing a new fleet (2+ jobs sharing state) - Auditing a fleet that "looks healthy" but isn't producing output - Migrating a fleet from skill-only heuristics to explicit prompts - Investigating why a downstream cron has "no input" or "nothing to do" - Converting a single cron int…

replace each internal reference: line 3: Load when the user asks to "set up cron jobs for X", "wire a daily pipeline", "automate triage and PRs", or add/change scheduled jobs that share state. Distinct from release-stewardship (single-repo release discipline) and kanban-orchestrator (per-task assignment via durable SQLite queue). line 114: "repo": "<owner>/<repo>", line 171: Fleet-wide drift_skip after a config-provider rename. Renaming model.provider in the agent's config (e.g. minimax → minimax-portal) can cause the cron runtime's spend-safety guard to skip every unpinned job in the fleet with [drift_skip]. The fix is hermes cron edit <id> --provider X --model Y for each affected job — not hermes gateway restart, not editing jobs.json directly. After pinning, the per-job last_status: error text may stay stale until the next scheduled tick; verify the pin actually worked by reading the freshly-written output file in the cron's output directory. line 184: Setting deliver: slack:<CHAT_ID> (or any explicit chat_id) is silently overridden to the current conversation thread's chat if the cron is created via cronjob.create while you're in that thread. line 188: When a user pastes a Slack permalink like <CHAT_ID>/<THREAD_TS> and asks "which cron posted this?" or "why did this land in my DM?", convert the timestamp and query the agent's session database. Cron sessions are named cron_<jobid>_<UTC-filename-ts> where <UTC-filename-ts> is the session's UTC start time formatted YYYYMMDD_HHMMSS. The Slack <THREAD_TS> is UTC seconds with a fractional .ffffff microsecond field — match within a 60-second window (or just match the HH:MMSS portion):

replace the `description:` line with: description: "Operate a Milestr project dashboard via the milestr CLI — create/update tasks, log KPIs, publish to Cloudflare Pages, back up and restore. Load before any milestr command or data.json read/write."

tighten the body — a working tighter version: # Milestr — Agent User Guide ## What is Milestr Milestr is a **local-first CLI dashboard** for tracking the work of a single project. Each project gets its own `data.json`, its own HTML dashboard, and its own Cloudflare Pages deployment. If you need to *develop the engine itself* (add features, fix bugs in the CLI, change the data schema, manage releases), this skill is not for you — this guide is for operators of a Milestr instance, not engine contributors. ## ⚠️ HARD RULE — NEVER edit `data.json` directly. CLI only **Every mutation to a Milestr `data.json` MUST go through the `milestr` CLI binary.** Direct edits via `python -c "json.load/open/write"`, `write_file`, `cat > data.json`, `sed -i`, or any other out-of-band tool are **forbidden**. **Why:** 1. The CLI is the only publisher — `saveData()` is the sole path that creates the backup, writes the `dashboard.log` `CHANGE:` record, regenerates `dashboard.html`, and ships to Cloudflare Pages. Direct edits ship a `data.json` that…

replace each internal reference: line 8: tags: [milestr, dashboard, cli, task-tracking, kpi, cloudflare-pages] line 9: line 190: If you are developing the engine itself (modifying the Milestr source tree), this skill's CLI reference applies but the schema references may lag the latest code. Always cross-check the engine's schema module for the authoritative Zod shape. line 199: A fourth track (the operator's workspace journal, e.g. <workspace>/journal-<agent>-<ISO-Z>.json) is required by the orchestrator's overall guidelines for cron-driven sweeps. If you write a sweep-style cron, follow the 7-step order in references/orchestrator-sweep-integrity.md.

replace the `description:` line with: description: "Use when designing an agentic loop on Milestr — four-role mesh (runner/planner/KPI-tracker/orchestrator), sole-writer data boundaries, typed-error retry, KPI as truth. Distinct from milestr-dashboard."

tighten the body — a working tighter version: # Designing an Agentic Mesh on Milestr ## What this is A class-level pattern for turning a goal-management tool (Milestr) into the shared-state substrate for an agentic loop that converges on a measurable North Star. Four agents, one `data.json`, one winner: the KPI that hits target. ## When to use - Morsy asks "transform X into an agentic solution" or "agents that achieve the North Star" - Designing planner/runner/orchestrator/KPI topology over a state store - Reviewing an existing multi-agent design for sole-writer conflicts or KPI conflation - Writing the prompts + data-boundary matrix for a new four-role mesh - Answering "what's the architecture for an agent loop on Milestr?" ## When NOT to use - Operating a Milestr instance (use `milestr-dashboard`) - Single-agent skills-loaded workflows without decomposition - Cost/token-optimized agent design (the mesh is NOT optimized for tokens — KPIs are the success signal) - Generic LangGraph/AutoGen tutorials (those are frameworks; t…

replace each internal reference: line 1: Use when designing an agentic loop on Milestr — four-role mesh (runner/planner/KPI-tracker/orchestrator), sole-writer data boundaries, typed-error retry, KPI as truth. Distinct from milestr-dashboard.

replace the `description:` line with: description: "Build premium marketing sites past anti-slop defaults. Delivers 9-primitive motion vocabulary (cursor, holo, parallax, KPI counters, marquee) plus editorial typography for top-5% visual craft."

tighten the body — a working tighter version: # Premium Marketing Site Craft This skill is the next tier beyond the `claude-design` anti-slop rules. Use `claude-design` for general design process (surface-first, slop diagnostic, anti-slop rules). Use this skill when the user has *already* rejected one or more drafts on aesthetic grounds — copy + layout fixes have been tried and are not enough. The remaining gap is **motion, presence, and layered craft**. ## When To Use Trigger phrases (any one): - still looks AI-made / looks like a template / looks like a Webflow demo - I need sophisticated animations / background holo / moving cards - top 5% best sites / nativeads.ai tier / benchmarks against a real premium reference - I can't accept this dumb version - explicit request for really top of the line Do NOT use this skill on: docs sites, admin tools, dashboards, forms/wizards, or any surface where motion competes with monitoring or input. Motion is for marketing surfaces only. ## The 9-Primitive Motion Vocabulary A top-5% land…

consolidate each conflicting rule pair into a single statement; keep rules terse and non-overlapping