🛠️ Skill Audit ☀️ Light

replace the `description:` line with: description: "Design and audit 2+ cron pipelines that hand off work through shared JSON files; use for state contracts, explicit prompts, and end-to-end verification."

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…

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

replace each internal reference: line 3: Use when the user wants to design, wire, change, or audit two or more scheduled jobs that share state. line 16: Use the relevant single-purpose skill for one standalone cron; this skill governs the coordination contract between multiple stages. line 113: "repo": "example-org/example-repo" line 171: Fleet-wide provider drift can leave jobs skipped after a provider rename. line 174: Scheduling several crons at the same minute can cause provider load spikes.

replace the opening paragraph: Use this skill when you need to work with cron fleet stewardship — it provides the conventions, examples, and reusable code so the agent can act without inventing the wheel.

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

replace the `description:` line with: description: "Make research-backed decisions via local Ollama tev1:4b: gather cited state, compose choice/noul/score questions, call /v1/systemone, return verdict with probabilities and sources."

tighten the body — a working tighter version: # Skill: tev1-decisions > **TL;DR** — When the user wants a decision, **research first**, then ask Ollama `/v1/systemone` with `tev1:4b`. Decision models don't predict; they weigh criteria. The work is in the state + the schema. ## When to load this skill Load when the user asks any of: - "Should I [buy / invest / pick / choose / hire / ship] X?" - "Is X [true / safe / risky / urgent / aligned]?" - "Classify / score / route / rank these items." - "Pick one of A / B / C for [purpose]." - "Give me a 1-5 score on [dimension] for X." Skip for purely factual questions, code generation, open-ended writing, or anything where the model needs to *generate* prose instead of *pick from a list*. ## Why this backend `tev1:4b` (Together AI, Qwen3.5-4B base, Apache 2.0) is the verified winner of the 2026-09-30 benchmark on 60 labeled support-ticket items: | Model | Avg precision | Latency | Download | |-------------|---------------|---------|----------| | **tev1:4b** | **98.3%** | 3…

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

replace each internal reference: line 72: - memory_store (local long-term store for project-specific context — e.g., recurring tickers, prior decisions) line 217: <workspace>/decisions/decision-YYYY-MM-DD-<topic>.json line 240: 5. Next action the user has to take, stated concretely and upfront. line 286: If the user asks for a report / deck of decisions (not a single call), bundle N calls (one per topic), write a short Markdown summary, and hand off to a presentation-creation skill for the PPTX. This skill is purely about the decision calls themselves. line 290: - benchmarks/jev-style/results-tev1_4b.json — 60-item precision benchmark (2026-09-30)

replace the `description:` line with: description: "Diagnose and repair failing Hermes cron jobs by classifying script, delivery, and scheduling failures, then fixing them one by one with read-only probes and verified runs."

tighten the body — a working tighter version: # Cron Failure Triage & Repair Class-level workflow for finding out which Hermes cron jobs are actually broken, why, and fixing them one at a time. Also covers "the cron isn't running" reports that are really delivery or duplicate-job problems. ## When to load this - "Check if there are any cron jobs that are not running due to failures" - "Why isn't cron X running / delivering?" - "Fix the cron jobs" / "let's go through them one by one" - An incident digest, health check, or manual sweep shows `last_status: error` on one or more jobs - A user asks which jobs are dead, stuck, duplicated, or "not doing their job" **Skip for:** - Designing a new cron fleet → `cron-fleet-stewardship` - A prompt that is blocked by the injection scanner → `cron-injection-scanner-forensics` - Model-pin / provider-drift skips → `cron-inference-config-pinning` - A read-only periodic audit that must not mutate anything → `hermes-install-health-audit` ## The three-layer failure model Most "the cron isn't…

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

replace each internal reference: line 47: $HERMES_HOME/cron/jobs.json — per job: line 48: $HERMES_HOME/cron/executions.db (SQLite) line 51: $HERMES_HOME/logs/gateway.error.log + gateway.log line 58: cp $HERMES_HOME/cron/jobs.json $HERMES_HOME/cron/jobs.json.bak-pre-<change>-$(date +%Y%m%d-%H%M%S). Same for any data file line 80: using the bot token from $HERMES_HOME/.env:

replace the `description:` line with: description: "Configure script-only Hermes crons without workdir and debug TERMINAL_CWD write-lock timeouts."

tighten the body — a working tighter version: # Cron script-mode without a workdir ## When to use - A recurring job that only needs to **run a script** (poll an API, check a file, post a report) and needs no LLM reasoning. - Any cron that failed with a TERMINAL_CWD write-lock timeout (hermes issue #79768). ## The trap A cron job with a **`workdir` set** takes the TERMINAL_CWD write lock while it runs. If any busy agent session, another cron, or a long-running reader holds that lock, the cron job dies waiting for it and times out. This is easy to hit accidentally because `workdir` is a natural thing to set ("the job lives in this project directory"). It is only needed when the job must **anchor its terminal/session context** to a directory (e.g. an LLM job that reads `AGENTS.md`, `SOUL.md`, or other project context files). **Rule:** a script-mode (`no_agent`) cron should have **no workdir**. The script uses absolute paths anyway. ## The recipe ```python cronjob(action='create', name='<name>', schedule='every 20m', scr…

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

replace each internal reference: line 35: script='<name>.sh', # relative to the configured Hermes scripts directory line 114: last_status: ok in the configured cron jobs file and inspect the job's line 115: output file in the configured Hermes cron data directory. line 122: prompt actually landed in the configured Hermes cron jobs file. line 125: the configured Hermes scripts directory for cron, or an absolute path inside the wrapper.

replace the opening paragraph: Use this skill when you need to work with cron script-mode without a workdir — it provides the conventions, examples, and reusable code so the agent can act without inventing the wheel.

tighten the body — a working tighter version: # LLM Prompt Audit A grader for system prompts. Two deliverables: (1) qualitative **grade** on 10 criteria, (2) quantitative **token-equivalence proof** if the user asks for a slim rewrite. ## When to use - "Review the intro to the LLM in folder X" - "Audit these N prompts" - "Can we get the same behavior in 8K tokens?" - "Grade this system prompt against your criteria" - Pre-ship check on a new agent's first system prompt ## Do NOT use when - The user wants the comm-style *content* updated → `agent-communication-style` - The user wants a single cron's per-turn token cost cut → `hermes-cron-token-audit` - The user wants config drift fixed → `hermes-config-audit` - The user wants to write skill frontmatter, not the prompt body → `hermes-agent-skill-authoring` ## The 10-criteria rubric Score each prompt on these 10 axes (0-3 each, sum to 30): 1. **Role clarity** — who, who for, what operative scope. Look for an explicit "You are X, serving Y" line, not just an interface descript…

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

replace each internal reference: line 52: Cite the source file and line for every claim. line 68: ("role", "you are") line 69: ("named audience", "serving") line 70: ("optional capability", "when available") line 95: A query-capable discovery tool must already exist and support keyword or semantic search, not only broad category filters.

replace the `description:` line with: description: "Query Hermes Agent token usage, cost, and session health from state.db. Includes sessions schema, cost-reconstruction recipes, and the broken-billing caveat."

tighten the body — a working tighter version: # Observability — Hermes Token, Cost & Agent Health Monitoring ## ⚠️ Critical: Two Stacks, One Right Answer There are TWO observability stacks on this machine. **Use the first one only.** | Stack | DB Path | Status | Use For | |---|---|---|---| | **Hermes-native** | `~/.hermes/state.db` | **LIVE, current** | Token usage, cost, session analysis | | OpenClaw legacy | `~/.openclaw/observability.db` | **ORPHANED, 50+ days stale** | Nothing — do not use | The OpenClaw stack (scripts in `~/.openclaw/skills/observability/`) was written for a different system. Its cron job disappeared, its source jsonl files stopped being written ~June 7, and all paths point to `~/.openclaw/` instead of `~/.hermes/`. **Ignore it completely.** --- ## Primary Source: Hermes state.db (always correct) **Path:** `~/.hermes/state.db` This is Hermes's own session database. It is live and current — query this first for any token/cost question. ### sessions table schema ```sql sessions ( id TEXT PRIMARY KEY,…

replace each internal reference: line 3: description: "Query Hermes Agent token usage, cost, and session health from state.db. Includes sessions schema, cost-reconstruction recipes, and the broken-billing caveat." line 25: | Hermes-native | ${HERMES_HOME}/state.db | LIVE, current | Token usage, cost, session analysis | line 26: | OpenClaw legacy | ~/.openclaw/observability.db | ORPHANED (machine-specific legacy) | Nothing — do not use | line 28: If an older OpenClaw stack is present on the host, ignore it — its cron job and source jsonl files are no longer written and its paths do not match Hermes. Use state.db only. line 34: Path: ${HERMES_HOME}/state.db (default ~/.hermes/state.db on a user install)

replace the opening paragraph: Use this skill when you need to work with observability — hermes token, cost & agent health monitoring — it provides the conventions, examples, and reusable code so the agent can act without inventing the wheel.

replace the `description:` line with: description: "Design and refine premium marketing sites with layered motion, distinctive typography, and anti-template polish after initial copy and layout revisions."

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

replace the `description:` line with: description: "Build a private self-report questionnaire for a reluctant or nonresponsive person, collect responses, and rank hypotheses without diagnosing."

tighten the body — a working tighter version: # Self-Report Instrument Builder (questionnaire for someone who won't talk) ## Purpose When a person won't answer messages and the obvious read — "he's avoiding us," "she's angry," "he needs space" — is a guess, build an **instrument** they fill in privately, then analyze the answers. The deliverable is a working, verified artifact plus an analysis path, not a set of suggested questions in chat. Two halves, both required: 1. **Instrument design** — question wording, order, scale types, privacy framing, and the candidate causes the instrument must discriminate between. Grounded in published research, not intuition. See `references/instrument-design-rules.md`. 2. **Build & collection** — a private web form, durable storage, an admin analysis view, and a watcher that tells the requester when answers arrive. See `references/cloudflare-pages-d1-stack.md`. **Do both in the same turn.** A question list in chat is under-delivery. A form without a defensible design is a survey tha…

replace each internal reference: line 246: - [ ] Results, privacy limits, test status, analysis method, and follow-up actions handed to the requester

replace the `description:` line with: description: "Realign tracking trees, KPIs, recurring reports, cron prompts, and standing instructions when an owner changes priorities."

tighten the body — a working tighter version: # Goal-priority pivot — applying an owner directive across every surface Trigger: the owner says something like "stop working on X — we need Y first", "drop A until B is done", "we don't have customers yet, let's get users first". A directive is only real once **every** surface carries it. This skill covers the class of "the priority changed; now make the whole system agree". ## The surfaces that must change together 1. **The tracking tree** (milestr / any task-KPI store) — retitle the new live goal, park the retired work. 2. **The KPI set** — values AND `source` fields. 3. **The recurring report crons** (daily pulse, weekly digest) — otherwise the next run re-raises the topic the owner just retired, in public. 4. **The standing instruction / SOUL / memory** — so a fresh session doesn't re-derive the old priority from context. Change 1+2 and the cron keeps reporting the old story. Change 3 and the dashboard contradicts the report. Change all four or the directive is not ap…

replace each internal reference: line 58: the named measurement path lacks the required attribution fields line 61: Writes may emit deployment logs; distinguish them from command results before proceeding.

replace the `description:` line with: description: "Diagnose and reactivate fading professional networks using LinkedIn exports, colleague data, and conversation history to build a prioritized outreach plan."

tighten the body — a working tighter version: # Professional Network Reactivation — Diagnose From Data, Not From Feeling The user's report ("my connections are fading", "I have no connections") is a **symptom report**, not a diagnosis. The correct move is to mine the user's own platform data for the mechanism before writing any plan. Guessing the mechanism produces a generic plan that the user has already received and rejected before. **Verified instance (2026-10-22, Morsy).** User said "my personal connections in the office are fading out." First read framed it as skip-level (G/C) visibility — **wrong**. User corrected: "it was not my intention. Was talking about my colleagues. Feels like i have no connections." Mining his LinkedIn export showed the real mechanism: 67 RingCentral colleagues connected, but only 5 ever messaged; 30 of 31 invites sent with **no note**; 96% of accepted blank invites across his whole network were never spoken to. The fix was "never send a blank invite again," not "attend more all-hands." ## When t…

replace each internal reference: line 23: Example. A colleague-focused audit can show that many connections never became conversations. Use the measured pattern to choose a specific behavior change instead of generic visibility tactics. line 61: Save the ZIP in a user-approved local directory and record that path before running the audit. line 81: Example finding pattern: line 82: > Employer invitations sent without a note may be accepted without any follow-up conversation. line 83: > Compare acceptance and messaging rates for invitations sent with and without a personalized reason.

replace the `description:` line with: description: "Send Slack messages via Hermes, diagnose gateway outages (no inbound/outbound), and fix mrkdwn rendering bugs. Load when Slack is broken or agent output formats wrong."

tighten the body — a working tighter version: # Slack Messaging Send messages to Slack channels through Hermes, diagnose when the Slack gateway platform is broken (no inbound, no outbound, retry loop in logs), and fix Slack mrkdwn format-rendering bugs in agent output. ## When to Use - User reports **"Slack is not working" / "no messages" / "no replies"** — START WITH `./scripts/verify_slack.sh`. This is almost always faster than reasoning from memory. - Sending notifications, reminders, or updates to Slack channels - Communicating with users via Slack as configured in Hermes - Sending automated reports or alerts to team channels - After a workspace reinstall — use `scripts/slack_invite_all_channels.py` to re-join every channel - User reports **agent output renders wrong in Slack** — strikethrough on text that shouldn't have it, italics where bold was wanted, raw `&amp;` / `&lt;` / `&gt;` leaking through, asterisks showing as literal `*`. Fix the **Slack adapter's `format_message` method** at `/Users/morsy/dev/hermes/plugins/p…

replace each internal reference: line 12: Fix the Slack adapter's format_message method in the slack platform plugin (path varies by install) line 31: # Load token from $HERMES_HOME/.env (defaults to ~/.hermes/.env) line 32: env_path = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes")) / ".env" line 44: channel = "C0XXXXXXXXXX" # Replace with actual channel ID

replace the `description:` line with: description: "Send and receive WhatsApp via the wacli CLI on macOS. Read chat history, send text/media, list chats, search, and sync the local store. Mirror of the imessage skill for WhatsApp."

tighten the body — a working tighter version: # wacli — WhatsApp CLI Use `wacli` to read and send WhatsApp messages via a locally-synced store (whatsmeow-backed). Works similarly to `imsg` but for WhatsApp. ## Prerequisites - **macOS** with `wacli` installed (`brew install` or from https://wacli.sh) - One-time auth: `wacli auth` shows a QR to pair the device. After that, the local DB persists across runs. - Verify with `wacli doctor` if anything looks off (auth, store, search). ## When to Use - Reading WhatsApp conversation history with a specific contact - Scanning the last 24h of a personal 1:1 chat for a coaching / reflection / relationship brief - Sending a WhatsApp message (text, image, file) - Searching the local message store (FTS5 if compiled in) - Listing recent chats to find the right JID - **Discovering which groups you belong to** (use `wacli groups list --query <name> --json` — `chats list` silently omits inactive groups) ## When NOT to Use - iMessage / SMS → use `imessage` instead - Telegram / Discord / Slack …

replace each internal reference: line 3: wacli is a host CLI: any agent on the same machine can drive it via terminal(), even without a dedicated wacli primitive in its tool_search. See the cross-agent pitfall before claiming inability. line 192: ## Message style — personal nudge to a specific named recipient (validated 2026-08-30, no-signature rule locked 2026-09-02) line 213: When the message is from a sender to a recipient who already knows them (family, close contact), do not sign with a relationship label ("Oncle", "Papa", "Maman", the sender's first name) or any closing attribution. The message lands first-person, ending with the CTA verb. The sender's exact framing: "the message lands as if from me, unsigned — first-person without attribution." This overrides any prior "sign off with relationship to recipient" guidance. line 219: Validated case 2026-08-30 (coaching agent → 'Yannouch - Brilliant Future' group): four drafts reviewed by the user. Draft 1: 4 paths + "Pour toi si..." opinions per option → user: "Remove your opinions about each option." Draft 2: cheerful intro but led with guilt-framing → user: "Intro is stressful." Draft 3: dropped intro entirely → user: "No…. add an intro but make it optimistic and cheerful!" Draft 4 (landed): cheerful intro + per-option bullet + 1-line description (named employer/school) + status, ending on a single CTA verb. User: "Send." line 302: wacli IS WORKING — don't echo the stale "Client outdated (405)" belief (corrected 2026-10-12, certificate-message send). If an agent's MEMORY.md or TOOLS.md carries a "wacli broken / 405 error" entry from 2026-09-11 or earlier, treat that as a stale-invocation memory, not a real outage. The corrected send path (wacli send text --to <jid> --message <body> subcommand form, verified with v0.18.1) works and returns a real MsgID.

replace the `description:` line with: description: "Diagnose macOS performance issues: load average spikes, CPU throttling, memory leaks, disk pressure, runaway processes, system database and log forensics."

tighten the body — a working tighter version: # Apple System Diagnostics Skill for investigating and diagnosing Apple system components on macOS by examining system files, databases, and logs. ## When to Use - **Machine feels slow** — "the computer is extremely slow", laggy UI, high load average, CPU throttling, runaway process, memory leak, disk pressure → see `references/macos-performance-triage.md` (or just run `scripts/macos-perf-probe.sh` for one-shot evidence) - Investigating unknown system behaviors - Checking status of Apple services/features - Debugging issues with macOS built-in functionalities - Auditing system data usage (like token consumption for AI features) - Troubleshooting when standard logs don't provide enough detail ## Performance Triage (quick path) For "the machine is slow" requests, run `scripts/macos-perf-probe.sh` first — it captures load, CPU throttle limit, load composition (R vs D), top CPU, top RSS, swap, process/thread counts, app version drift, GPU processes, and agent/dev background bleed…

replace each internal reference: (no internal lines found — manual review needed)

replace the `description:` line with: description: "Audit a Hermes system prompt for bloat, contradictions, duplicate specs, locked-rule conflicts, env-state burns, and unconditional catalog dumps. Includes 6-class taxonomy and 5-question triage."

tighten the body — a working tighter version: # Hermes Agent Prompt Audit Class-level workflow for reviewing the **content** of a system prompt actually being sent to a Hermes agent — bloat, contradictions, locked-rule conflicts, duplicate specs, and environment-state burns. ## When to Use - Morsy says "analyze the prompt", "review the system prompt", "audit the SOUL.md", "is this prompt in better shape" - You captured a `~/.hermes/tmp/llm_prompts/*.md` dump and need to review it - A session ran hot (token use spike, repeated confusion) and you suspect the prompt, not the model - Before shipping a change to SOUL.md, `agent-communication-style.md`, AGENTS.md, or any always-injected block ## Why This Skill Exists Morsy's verification rigor (memory, locked 2026-09-03): if you cite "X saves Y tokens", measure it. Same logic applies to audit claims — every "this is fluff" call must be backed by file:line and a concrete alternative. The audit is a content problem, not a transport problem. Don't confuse with `hermes-install-health-…

replace each internal reference: line 32: suppressing the skills index removes the instruction and the discovery path together — the system-prompt builder bundles narrative + index in one string, and no skills_search tool exists to replace it (skills_list has only a category param, and a full dump is several times the index it replaced, spilling to the prompt-cache overflow directory). Consult your Hermes version's source for current line numbers. line 73: Distinct from the system-prompt taxonomy above: this is the post-turn fork path (background review extractor), not the main agent loop. When auditing a prompt dump and the user_message is a memory- or skill-review invocation, the dump reveals bloat in three places: line 76: Fix: install a ~300-token task-specific system prefix on the fork that overrides the main system prompt, keeping the prefix cache key byte-stable so warm-cache reads still hit. line 79: approx_input_tokens > 100K AND user_message starts with "Review the conversation above and..." (i.e. a memory- or skill-review prompt emitted by the background-review fork).