# Close Claude Code. Don't Open It Again Until These Five Files Are In It.
Every Claude Code setup has the same five blind spots by default: no second reader checking
whether a "done" is real, no way to know if a stranger could find their way around your own
setup, no clean path from a vague goal to dispatched work, no place approved-but-forgotten fixes
land, and nothing watching your saved shortcuts for rot. None of that is a Claude limitation — it's
missing scaffolding. Five small files close all five gaps, and none of them cost anything to run.
This page hands you The Agent Stack — the five agent files, the install guide, and the map of
how they work together.
## The five files, one line each
1. verifier — the maker/checker gate. Checks any "done" claim or outbound draft before you
trust it, with a mechanical 4-point checklist (schema, confidence, truncation, source-grounding)
instead of vibes. Catches the #1 documented failure mode of autonomous agents: claiming success
on a login-wall or a limit-kill.
2. cold-start-verifier — simulates a fresh agent with zero memory landing on your setup. Given
only your entry file, it has to find a named skill, your project's truth file, a script's usage
line — purely by following committed pointers. Reports the exact broken hop, never fixes it
itself.
3. hermes-decomposer — turns one vague goal ("I want X") into 1-5 scored, dispatchable
missions instead of a stalled wish. Interrogates first ("what does done look like?"), proposes,
never dispatches on its own.
4. the debt scanner — a read-only aggregator that finds every approved-but-never-run
suggestion across your projects by convention, writes one deduped ledger, and lets you run any
of them from one place with an evidence receipt.
5. skill-doctor — audits your saved skills and long docs for rot on a schedule: dead
references, stale paths, bloated files, unmerged learnings. Read-and-propose only — it never
silently rewrites anything.
Why install all five together, not one at a time
Each file alone catches its own class of failure. Installed together they close a loop:hermes-decomposer starts work cleanly, verifier catches it lying about being done,cold-start-verifier and skill-doctor catch the setup itself decaying between uses, and
anything any of the four surfaces but doesn't fix lands in the debt scanner instead of vanishing.
That loop — not any single file — is the compounding part. Pay the five-file install once, and
every session after runs with a checker, a newcomer-test, a goal-splitter, a debt scanner, and a
skill doctor built in.
The full step-by-step (about 5 minutes)
1. Make the folders (skip any that already exist): ~/.claude/agents,
~/.claude/skills/debt, ~/.claude/skills/skill-doctor.
2. Drop in the two agent files + the decomposer: verifier.md, cold-start-verifier.md, and
hermes-decomposer.md go straight into ~/.claude/agents/.
3. Install the debt scanner: the SKILL.md and its script go into ~/.claude/skills/debt/.
Open the script and fill in the small config block at the top (about 6 lines) with your own
backlog file and flags file.
4. Install skill-doctor: the SKILL.md goes into ~/.claude/skills/skill-doctor/.
5. Restart your session so it picks up the new agents/skills, and try each one once — full
worked examples are in README.md inside the kit.
What's in the kit
- The 5 agent files — sanitized: no personal paths, no names, no keys, generic ~/.claude/
locations throughout.
- The install guide (README.md) — install-in-5-minutes steps a stranger can follow with zero
context, plus a worked example for each file.
- The stack map (STACK-MAP.md) — the one-page "which file fires when," so you install the
five as a system instead of five unrelated tools.
Comment STACK on today's video and the kit lands in your DMs — or grab it right here with the
button above. The comment word rotates out at 4pm today when tomorrow's drop lands.