digest-or-burn

digest-or-burn

A burn-pile CLI for your markdown hoard. Decide: absorb, act, or let it burn.

The problem

Your "Second Brain" is a graveyard with good UX.

Karpathy has 99k bookmarks. You have 5,000 notes in Obsidian. Neither of you knows what's in there. The bottleneck isn't retrieval — it's rejection. Every tool helps you save more. Nothing helps you decide what to let go.

"Most saved items go unread; using AI summaries instead" — HN user

"20k bookmarks, search became FIFO" — HN user

The fix

digest-or-burn scans your markdown folder and triages every file into three exits:

• 🔥 Burn — Not worth keeping. Delete it or archive it.
• 📚 Absorb — Worth integrating into your long-term knowledge.
• ⚡ Act — Contains a pending task or unfinished idea. Do it or delegate it.

Works in two modes:

• Heuristic (default, free) — Uses file age + word count to classify
• LLM (--llm flag) — Sends content to Claude Haiku for smarter verdicts

Install

npm install -g digest-or-burn

Or use via npx:

npx digest-or-burn scan ~/notes

Quick start

# Scan your vault — dry run, markdown report
dob scan ~/Documents/Obsidian/my-vault

# Only files untouched for 2+ weeks
dob scan ~/notes --older-than 14d

# AI-powered verdicts (bring your own API key)
ANTHROPIC_API_KEY=sk-... dob scan ~/notes --llm

# JSON output for piping
dob scan ~/vault --format json | jq '.burns[].relativePath'

Example output

# Digest or Burn — Triage Report

**Scanned:** `~/notes` | **Files:** 47 | **Mode:** heuristic

## Summary

| Verdict | Count |
|---------|-------|
| 🔥 burn | 12 |
| 📚 absorb | 23 |
| ⚡ act | 4 |
| 🔍 review | 8 |

## 🔥 Burn (12)

- **Meeting Notes Jan 15** (`meetings/jan-15.md`)
  92d old, 26 words — likely abandoned stub
- **Project Ideas 2024** (`ideas/old-ideas.md`)
  180d old, 45 words — abandoned stub
...

How it works

┌─────────────┐     ┌──────────────┐     ┌──────────────┐
│  Scan folder │ ──▶ │  Triage each │ ──▶ │  Report with │
│  for .md     │     │  file by age │     │  cleanup     │
│  files       │     │  + word count│     │  commands    │
└─────────────┘     │  (+ LLM opt) │     └──────────────┘
                    └──────────────┘

Heuristic rules:

Age	Words	Verdict
> 30 days	< 200	🔥 burn (high confidence)
> 14 days	any	🔥 burn (medium)
> 7 days	any	🔍 review
< 48 hours	> 500	📚 absorb
< 48 hours	any	📚 absorb

LLM mode (--llm) sends each file's content to Claude Haiku, which reads the actual text and classifies based on content quality, not just metadata.

Options

Flag	Description	Default
--llm	Use Claude Haiku for AI verdicts	off
--format <md|json>	Output format	md
--older-than <Nd>	Only files older than N days/hours/weeks	all
--ext <.md,.txt>	File extensions to scan	.md,.markdown,.mdx

Environment variables

Variable	Description
ANTHROPIC_API_KEY	Required for --llm mode. Get one at console.anthropic.com

Ecosystem

digest-or-burn is part of the Burn reading triage ecosystem:

Tool	What it does
digest-or-burn	This CLI — triage your local markdown
burn-mcp-server	MCP server for AI-powered reading queue
reading-routine	Daily reading backlog triage CLI
reading-feed-mcp	Vendor-neutral reading feed MCP
routine-templates	Claude Code Routine prompt templates
Burn	Full reading triage app (web + iOS)

Philosophy

The knowledge management world optimizes for capture and retrieval. But information has a half-life. Most of what you save today will be worthless in 6 months.

digest-or-burn is the opposite of a Second Brain. It's a burn pile — a place where your notes go to be honestly evaluated, and most of them get composted. The 20% that survive are the ones worth your attention.

The bottleneck is rejection, not retrieval.

Contributing

PRs welcome. Useful additions:

• More triage heuristics (link density, heading structure, code block ratio)
• Additional LLM providers (OpenAI, local models via Ollama)
• Export integrations (Obsidian, Notion, Logseq)

License

MIT

Built by @hawking520 | Try Burn for full reading triage