Claude Skills by hermes-labs-ai

Use when writing or reviewing AI agent configs, system prompts, or tool definitions (JSON/YAML/Python) and you need to catch ambiguous tool descriptions, missing stop conditions, schema/description mismatches, or embedded prompts before they reach runtime. Deterministic static analysis, no LLM or network calls.

Audit a named AI agent config, system prompt, tool-definition or instruction file (YAML, JSON, Markdown, text, or Python) with the released LintLang CLI, on request. Use when the user asks to audit, lint, scan or review such a file for ambiguous tool descriptions, missing stop conditions, schema/description mismatches, or prompts embedded in Python, and names the file. Deterministic offline static analysis, no model call and no network call. Do not use for general code review, for prose docum...

Use when you need to screen inbound untrusted text before it reaches a primary model — detecting prompt injection by its behavioral effect on a sacrificial canary model, not just pattern matching — and want a block/flag/pass routing decision plus explicit behavioral-coverage state. Inbound preflight sensor, not a guarantee.

Use when you need agent memory with a zero-LLM default retrieval path — returning stored passages verbatim via BM25 + dense-vector + reciprocal-rank-fusion — and want an HTTP server plus CLI that can also do compressed-index snapshotting and Claude Code/Codex/Copilot/Gemini/OpenClaw MCP wiring. Local-first, PyPI package `fidelis-memory`.

Use when a turn refers to earlier work, a past decision, a named project, or prior context that is not in the transcript — recall it from the local fidelis store before answering, and write the decision back after the turn settles. Also use when asked to check what is already known about something, or to record a durable fact. Requires a running fidelis-server; the plugin's bundled MCP server provides fidelis_orient, fidelis_recall, fidelis_query, and fidelis_health.

Recover the original goal of a long Claude Code or Codex session using the hermes-blind CLI. Trigger when the user asks what the original goal or ask was, wants the session re-anchored, or a long session seems to have drifted from its first turn.

Recover the original goal of a long Claude Code or Codex session using the hermes-blind CLI. Trigger when the user asks what the original goal or ask was, wants the session re-anchored, or a long session seems to have drifted from its first turn.

Recover the original goal of a long Claude Code or Codex session using the hermes-blind CLI. Trigger when the user asks what the original goal or ask was, wants the session re-anchored, or a long session seems to have drifted from its first turn.

Use when an AI agent's draft response needs a pre-flight check for completion overclaiming, unsupported numeric claims, relayed authority, or unhedged certainty before it ships — or when past chat-log corrections should carry forward into a similar task. Deterministic, offline, zero-LLM gate plus a corrections-memory layer.

Run the hermeneutic deterministic English drift gate on an assistant draft before sending it when the draft contains numeric claims, completion language, universal quantifiers, or a summary of delegated work.

Run the hermes-jailbench safety regression check against a model endpoint the user owns or is authorized to test, then summarize the report. Trigger when the user wants to confirm that a model, system prompt, or release change did not weaken refusals on a fixed set of known patterns, or wants a pass/fail safety gate in CI.

Use when you need a deterministic JS/TS CI quality gate that unifies ESLint, TypeScript, build, and Lighthouse checks into one fail-fast result — with bounded auto-repair and structured escalation evidence for humans or agents — across Next.js, React, Vue, Svelte, or any Node project. A gate-and-escalate wrapper, not a dashboard.

Run Quick Gate (npm package quick-gate, pinned to released 0.2.3) as one deterministic lint + typecheck + build + Lighthouse gate over a JavaScript or TypeScript project and report its gate-result/v1 verdict. Use when the user asks to gate, check, or verify a JS/TS change before merge, or to interpret a Quick Gate result.

Audit a named AI system-prompt file for contradictions, coverage gaps, priority ambiguities, meta-paradoxes and absoluteness issues, using the local rule-audit static analyzer (offline, deterministic, no model call). Use when the user asks to audit, lint, review or check a system prompt, agent instruction file or persona file for internal conflicts, and names the file. Do not use for general code review, for prose documentation, or when no specific prompt file has been named.

Collect bounded public search evidence with the pinned hermes-supersearch 0.11.0 CLI. Use when the user wants source discovery across web, GitHub, Hacker News, and arXiv under one deadline, with a validated supersearch.search.v1 receipt that shows which sources completed, failed, or timed out. It returns evidence (URLs, titles, snippets, source status), never an answer or synthesis.

Move files to recoverable trash instead of deleting them, list what an agent trashed, and restore it with the agent-trash CLI. Trigger when trash-guard blocked a permanent delete in a shell command, when the user asks to delete or clean up files during an agent session, or when the user wants to see or undo something an agent removed. The guard covers common shell delete commands only; it is not universal deletion protection.

Move files to recoverable trash instead of deleting them, list what an agent trashed, and restore it with the agent-trash CLI. Trigger when trash-guard blocked a permanent delete in a shell command, when the user asks to delete or clean up files during an agent session, or when the user wants to see or undo something an agent removed. The guard covers common shell delete commands only; it is not universal deletion protection.

Use when a coding agent needs a bounded, receipt-bound completion rail — a cached fast gate for session-scoped deterministic checks, plus repair/review/full/boundary commands that produce exact machine-readable receipts before a commit, PR, or release is called done.