The Prompt Library

01Hermes Starter Tips

You got Hermes running yesterday. Here is the fast path to feeling at home. Hermes is an agent that runs on your machine, gives a language model real tools (terminal, files, web, memory, skills), and remembers you between sessions. Full docs: hermes-agent.nousresearch.com/docs.

Commands worth knowing

Security habits from day one

02The Prompts

Copy, adapt, run. The first four are the build prompts from the field guide. The rest are day-to-day life with an agent.

Read the docs at https://hindsight.vectorize.io/developer/installation first.
Then:
1. Run the Hindsight memory server in Docker (embedded database, local
   embeddings and reranker, ports 8888 API and 9999 UI).
2. Configure an LLM provider. Use my API key from the environment
   variable (never print it). Default to the cheapest fast model.
3. Create a bank named "personal" with a reflect mission.
4. Prove it works end to end with real commands:
   - create the bank
   - retain one memory
   - recall it back
   Show me the actual outputs and the token usage from the retain call.
Ground rules:
- Never write, print, or store a secret anywhere.
- No em dashes in any user-visible string.
- Keep scripts ASCII-only.
- Verify with real probes, not "it should work".

Install RTK (Rust Token Killer) for my coding agent.
1. Download the binary for this OS from the GitHub releases page and put
   it on PATH.
2. Check which agent integrations are supported (claude, cursor, codex,
   hermes, and others). Install for the agent I actually use.
3. CRITICAL: before installing, resolve the real agent home directory
   (on Windows it is usually AppData\Local\hermes, not ~/.hermes). Set
   the HOME env var the installer expects and verify the plugin files
   landed in the right place.
4. Prove the rewrite works: run one real command through the proxy and
   show the compacted output vs the raw output.
Rules: fail open means never break the original command. No secrets.
No em dashes. Verify with real output.

Build a "packet" for my setup so another machine can install it.
1. Create a folder with:
   - packet.yaml manifest (id, version, kind, privacy, requires,
     installs, touches, data_policy with sends: [] and verify list)
   - SKILL.md documenting the setup end to end
   - an idempotent install script (PowerShell for Windows, bash for
     POSIX) that checks prerequisites, backs up config before touching
     it, and skips anything already installed
   - a verify script with real probes that exits non-zero on failure
2. The install script must fail loud if required values are still
   placeholders (e.g. LLM model = "your-model"), listing the options.
3. Run the verify script and show the pass/fail output.
4. Write a registry entry with SHA-256 checksums of every file.
Ground rules: no secrets anywhere, ever. Idempotent (run twice =
same state). ASCII-only scripts. No em dashes. Real probes only.

Take my internal packet and produce a public flavor for distribution.
1. Copy it to a clean output directory.
2. Replace every owner identifier (paths, project names, machine names,
   personal handles) with generic placeholders (e.g. $HERMES_HOME,
   your-bank-id, your-project).
3. Rewrite the manifest: privacy: public, generic author, data_policy
   sends: [] preserved.
4. Run a secrets sweep over the output; if any credential-shaped string
   survives, fail with file:line.
5. Re-scan the output for the original identifier list; if ANY owner
   identifier survives, fail. A public packet that still contains
   internals must not ship.
6. Show me the scrub report (what was replaced) and confirm exit 0.
The fail-closed guard is the deliverable. No telemetry. No em dashes.

This is my first real session. Orient me:
1. List what you can actually do: tools available, skills installed,
   MCP servers connected, memory settings.
2. Show me my config location and any defaults I should know about.
3. Suggest three useful things I can do this week with what I have.
4. Tell me how to install a skill and how to turn a workflow into one.
Rules: no secrets in the output. Be concrete, not generic.

I keep doing the same workflow and I want it as a reusable skill.
1. Look at the procedure I describe and write it as a SKILL.md with
   front matter (name, description, trigger, version).
2. Include: numbered steps, exact commands, expected outputs, a
   pitfalls section with anything that bit me before.
3. The description must fit the trigger budget: one sentence, trigger
   first, ends with a period.
4. Save it under the skills directory with the correct category.
5. Show me the final file and how to invoke it next time.
No secrets. No em dashes.

Review my persistent memory for hygiene.
1. List the entries and flag anything stale, redundant, or expired.
2. Flag anything that looks like a temporary task detail that should
   not be a permanent fact (task progress, file counts, one-off
   outcomes).
3. Suggest consolidations: merge overlapping entries, drop anything
   that will not matter in a week.
4. Show me the proposed final set before changing anything.
Rules: never suggest removing a user preference or correction. No
secrets in the output.

Connect me to this MCP server: [name, url or command].
1. Verify the endpoint is reachable before configuring (one probe).
2. Add it with the correct transport (HTTP url, or stdio command).
3. If it needs auth, use env vars or the MCP login flow; never print
   tokens or ask me to paste secrets into chat.
4. Run the real status check and list the tools discovered.
5. Tell me when the tools will appear (next session) and name the
   mcp__ prefix so I can spot them.
Rules: no secrets in output. If it fails, show me the actual error.