The reliability layer for Claude Code

Finish multi-session work — without losing context or shipping silent failures.

Claude Code is brilliant for an afternoon. Across a week of sessions it forgets where it was, re-introduces fixed bugs, and quietly marks work "done" that was never verified. This kit gives it a durable harness so every session resumes cleanly and every "done" is a proven done.

# session 1 — scaffold the harness from your plan > long-running-task-init 12 features translated, each with a real verification step harness/ committed — resume anytime # session 7, days later — pick up exactly where it left off > long-running-task-resume baseline green · read prior progress + lessons built add-rate-limiter → verified (429 returned under load) → committed stopped after one feature. your turn to review.

Why long tasks fall apart

The expensive failures aren't dramatic crashes. They're silent — work that looks finished but isn't.

Lost context

A new session starts cold. It re-derives decisions you already made and undoes work it doesn't remember doing.

Silent "done"

A feature gets marked complete because the code looks right — not because anything was run and observed.

Repeated bugs

The same trap gets hit session after session because last week's fix was never written down anywhere durable.

The harness fixes this

Four small files at your project root turn ad-hoc sessions into a disciplined, resumable loop. Two skills drive it.

  1. Scaffold once. Point long-running-task-init at your plan. It turns each checklist item into a tracked feature with a concrete verification step, and seeds a progress log + a lessons fix-log from your project docs.
  2. Resume any session. long-running-task-resume reads prior progress, runs a baseline-green check, and refuses to build on a broken tree.
  3. One verified feature at a time. It implements exactly one feature, runs its verification, and stops — bounded scope, reviewable diffs, no runaway sessions.
  4. Nothing slips. Anti-cheat rules baked in: don't edit tests to pass, don't mark done without verifying, don't fabricate lessons. Failures get logged, not flailed through.

What's included

The harness is the hero. Around it, the skills that make Claude work like a careful engineer — your whole loop, in one install.

★ The hero — the multi-session harness

long-running-task-init  +  long-running-task-resume

Durable, resumable, verification-gated execution across as many sessions as a project takes. Ships with the full convention spec and copy-ready templates (feature list, progress log, lessons, baseline check). This is the part you can't get free.

persistent-memorycross-session facts: one file each + a curated index
research-first-debuggingfind the root cause before changing anything
verify-before-shippingprove it works before calling it done
token-efficient-workflowcheapest-tool reads, model routing, cache hygiene
git-commit-and-prguard-railed commit → push → PR
session-to-docscapture learnings into your docs at session end
skill-authorwrite your own skills with progressive disclosure
subagent-orchestrationfan-out, find→verify, pipelines for big tasks

Plus two bonus hook utilities (desktop notifications + permission tracking) and a one-command installer.

Why this, and not a free pack

It's a system, not a pile

Free marketplaces give you dozens of loose skills. This is one opinionated, integrated loop — execution, memory, and verification designed to work together.

The harness is genuinely rare

Free "resume" skills recover old context. None give you the structured feature-list + verification + anti-cheat discipline that stops silent failures. That's the core here.

Built from real use

These are battle-tested conventions from running multi-session work, sanitized and packaged — not theory, not AI-generated filler.

Yours forever

One-time purchase. Plain Markdown you can read and edit. Updates ship as new versioned releases.

Get the kit

$34 $29
one-time · intro price for the first buyers
  • The multi-session harness (2 skills + spec + templates)
  • 8 more skills for your full working loop
  • 2 bonus hook utilities + one-command installer
  • Plain Markdown — read it, edit it, own it
  • Free versioned updates
  • 14-day refund, no questions
Buy now — $29

Instant download · works with Claude Code

Honest note: about half of these skills have capable free alternatives — git, docs, code review. You're buying the harness and the curated, integrated whole, not a skill count. If that's not worth $29 to you, the free marketplaces are genuinely good; start there.

FAQ

What exactly do I get?

A downloadable folder: 10 skills as SKILL.md files, a reference spec for the harness convention, copy-ready templates, two bonus hook scripts, and an install.sh that drops everything into ~/.claude/skills/.

How do I install it?

Run bash install.sh, or copy the skill folders into ~/.claude/skills/ yourself. Then ask Claude to run the long-running-task-init skill in a git project that has a plan file. Takes about a minute.

What are the requirements?

Claude Code, a project under git, and some documented context (a CLAUDE.md or README) plus a plan/checklist to work from. The hook utilities also need jq.

Isn't a lot of this available free?

Some of it, yes — and we say so above. The differentiated piece is the harness: structured, verification-gated, anti-silent-failure multi-session execution, which the free options don't provide. The rest rounds it into one integrated loop you don't have to assemble yourself.

Do you offer support?

This is a self-serve kit — the bundled docs are thorough and are the support. No 1:1 setup help. Updates ship as new versioned releases to existing buyers.

Refunds?

14 days, no questions asked.