Why did each one die? Evidence, not vibes. The last commits touch Stripe files: it died at the payments wall. Another repo's first commit lands three days after this one's last: the killer gets named.

What's your pattern? Your projects die at day 19. Four of six were abandoned within 48 hours of starting something new.

Which one still has a pulse? Somewhere in that folder is a project that's 90% done: built, documented, never shipped. This finds it, checks what got easier since it died, and writes the short list of steps between it and a URL.

A representative scan (invented projects, real classifier verdicts). Your agent turns it into the funeral: an epitaph per corpse, your patterns named, and an offer to start resurrecting the strongest pulse right now. Yours will hurt more.

What it detects

Cause of death, read from git history:

Cause	How it knows
shiny object	Another repo you own had its first commit within days of this one's last. The killer is named.
deploy fear	README done, 20+ commits, real code, zero deploy config. It worked. It never shipped.
payments / auth wall	The final commits touch stripe/billing or oauth/login code.
boilerplate wall	60%+ of all file changes were config files. It died configuring.
rewrite spiral	Multiple rewrite/migrate commits; rebuilt instead of finished.
scope explosion	100+ files, no deploy config. It grew instead of shipping.
slow fade	Commit gaps stretched until they stopped. No wall, no killer; it drifted.

It also separates the finished (deployed, pushed, documented; done, not abandoned) from the unversioned (no git, so no autopsy). Then it ranks the dead by pulse, how close each is to shipping, and the agent takes over:

Autopsy interview: ambiguous deaths get a question; git evidence is marked (forensic), your answers (confirmed).
World-check: before prescribing a dig, it checks what changed since the death: the API that now has an SDK, the model that's 20x cheaper.
Resurrection: a ≤7-step plan ending at shipped, and an offer to start on step 1 right now.
Relapse watch: resurrections are recorded (--state + --mark-resurrected); every later scan reports whether the patient is holding.
Necromancer mode: ask your agent to build something new and it checks the graveyard first; you may have built 60% of it in 2024.

Install (10 seconds)

npx skills add https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/agent_skills/project-graveyard

The skills CLI installs it into whatever agents you have (Claude Code, Codex, Cursor, Copilot, Antigravity, and others); or copy this folder into your agent's skills dir. Then: "run the graveyard on ~/dev and ~/projects".

Standalone, no agent required:

python3 project-graveyard/scripts/graveyard.py ~/dev ~/projects

Scope and privacy

Everything runs locally: one plain-Python file, stdlib only, zero network calls, read-only. It reads git metadata (commit dates, messages, filenames), never your code's contents. Name folders and it scans only those; given none, it checks a fixed list of usual project spots (DEFAULT_ROOTS, line 30 of the script), never "everything on your machine." Want to post your report? --redact swaps project names for project-1..n.

Prove it works before installing, from a clone of this repo:

python3 agent_skills/evals/project-graveyard/test_graveyard.py   # 16 checks, ~10 seconds

Limits: no git means no autopsy (counted, not diagnosed). "Dead" is 45+ days silent, tunable with --days. Tested on macOS and Linux.

Files

project-graveyard/                  # ← this is all that gets copied
├── SKILL.md                        # agent instructions: report format, epitaph rules, resurrection protocol
├── README.md                       # this file
├── scripts/graveyard.py            # scanner + autopsy + pulse ranking (Python 3.8+, stdlib, offline)
└── references/causes-of-death.md   # the taxonomy: signals, confidence, resurrection strategy per cause

Part of awesome-llm-apps · Apache-2.0 · Last verified: July 2026