The Senior Engineer's AI Vault.

Formerly the AI Prompt Vault. The 50 prompts are still here. Version 2 adds the setup around them.

How it works

1. Unzip it. Every setup step is a plain file, so you can read what your agent will run first.
2. Open Claude Code or Codex in the folder and say: "read README.md and set this up."
3. The agent interviews you, shows you a plan, and waits for your yes. Then it builds the setup, runs the verify script, and walks you through what it built.

> read README.md and set this up

… interview, plan, your yes …

PASS vault:personal
PASS start-hook
PASS guard-blocks-env
PASS end-hook
PASS claude-hooks
PASS dotfiles-links

What your agent builds

• A memory vault. A private git repo of markdown notes the agent reads at the start of every session. One vault by default, or separate ones for personal and work.
• Agent wiring. Claude Code loads the vault through its global CLAUDE.md, Codex through its global AGENTS.md.
• Three hooks. Session start loads your notes and your latest handoff. Session end writes a short session note. A secrets guard blocks reads and edits of .env files and common key files.
• Handoff and resume. Stop in one session, agent or machine, and pick up in another from the handoff note.
• A dotfiles repo that holds the hooks and skills, so your setup lives in git.
• A multi-agent starter. An empty model routing table and a one-page orchestrator pattern.
• A verify script that checks the vaults, hooks, agent wiring and links, and prints PASS or FAIL for each.

Everything it builds is structure and wiring. The notes are yours to fill in.

The five levels

The guide in the zip walks from one level to the next, one short page each.

1. Chat prompts. You write a good prompt, paste in the code or the error, and read the answer.
2. Project memory. The agent reads a set of notes at the start of every session: who you are, how you work, what each project is and what was decided.
3. Skills and hooks. A skill is a procedure written down once, which the agent runs when you ask for it by name or when the task matches its description.
4. Subagents and parallel sessions. One session plans and judges. Other sessions, or subagents it starts, each do one settled step and report back.
5. Sessions, agents and machines. Work moves. You stop in one session and pick up in another, start in Claude Code and finish in Codex, or leave the desk and continue on a laptop.

Safe to run

• It backs up every existing file before it changes it.
• It keeps what you already have in CLAUDE.md, AGENTS.md and your settings, and adds its own block next to it.
• It shows you the full plan and waits for your yes before it writes anything outside the vault.
• It asks before creating any repo, and only creates private ones.
• It is safe to run again. A second run replaces its own block and skips hook entries that are already there.
• It writes what it did, and every place it adapted a step to your machine, to setup-log.md in your vault.

Platforms. Linux is tested end to end with Claude Code. Codex, macOS and Windows are beta. Windows needs Git for Windows.

The secrets guard is a guardrail, not a security boundary. When it blocks something, your agent stops and asks you. Keep real secrets in a secrets manager.

Also in the zip

• A five-level guide, from chat prompts to working across sessions, agents and machines.
• The 50 prompts, 10 per category, as Markdown, CSV and PDF:
  • Code Review. Blast radius, security, concurrency, error handling, API contracts, test gaps, naming, dependency boundaries, migration safety, PR descriptions.
  • Debugging. Hypothesis-driven, distributed traces, memory leaks, race conditions, slow queries, flaky tests, incident triage, stack traces, config drift, time zones.
  • Refactoring. Extract domain, reduce complexity, untangle god classes, value objects, illegal states, flag arguments, strangler fig, long functions, async modernization, feature flag cleanup.
  • System Design. Capacity planning, queues, multi-region, idempotency, rate limiting, distributed transactions, cache invalidation, schema migrations, webhook delivery, risky rollouts.
  • Docs & PRs. PR descriptions, ADRs, runbooks, API references, changelogs, postmortems, READMEs, onboarding docs, tech specs, migration notices.
• Bonuses: 5 .cursorrules files (one per category) and 3 CLAUDE.md starters (Go services, Elixir umbrellas, AI agent projects).
• Every setup step as a plain file, so you can read exactly what your agent will run.

Level 1 preview: two of the 50 prompts

Level 1 is chat prompts. These two are free to read and copy. The other 48 are in the kit.

1. Blast Radius Analysis Code review

Act as a Staff engineer reviewing this diff for blast radius before merge.

For each changed file, answer:
1. What systems, services, or callers depend on this code path?
2. What's the worst realistic failure mode if this ships broken?
3. Is the change reversible without a backfill, replay, or schema rollback?
4. What monitoring would I check in the first 30 minutes after deploy?

End with a single sentence verdict: SAFE TO MERGE, MERGE WITH MITIGATION,
or HOLD. If MERGE WITH MITIGATION, list the mitigations in priority order.

Diff:
<paste diff or attach file>

Why it works. The single-sentence verdict forces commitment instead of a wishy-washy "it depends." The four-question structure stops the model from nitpicking syntax when there's a real systems-level problem to flag.

2. Hypothesis-Driven Debugging Debugging

I'm debugging a problem. Reset me. Walk me through this rigorously.

Symptom (what I observe):
<describe>

What should happen:
<describe>

What I've already ruled out:
<list>

Now:
1. Generate 5 hypotheses ordered by likelihood given typical failure modes
   in this stack.
2. For each hypothesis, write the cheapest test that would falsify it.
3. Identify which test gives me the most information per minute spent.
4. If two hypotheses are entangled, propose the ordering that disentangles
   them fastest.

Don't suggest fixes yet. We're isolating, not fixing.

Why it works. Most debugging time is wasted on the first plausible explanation. Forcing five hypotheses + cheapest-falsification rebuilds the scientific method when fatigue has eroded it. The "don't suggest fixes yet" line is the hardest one for the model to obey and the most important.

Built by Jared Smith

Founder and engineering leader with 13+ years shipping production systems. CISO at Lavender. Backend at BlockFi. Elixir at AAMP Global.

Today I'm a founding engineer at an early-stage startup, where most of the code is typed by coding agents and every line is mine to answer for. Read the full positioning page →

FAQ