wow-thats-tasty — AI Paper Meal Plan 🍽️

---

What it is

wow-thats-tasty is a concierge skill for the Hermes Agent that reads AI papers from dair-ai/AI-Papers-of-the-Week, ranks them with your personal tasty / bitter / sweet-but-heavy taxonomy, and writes the ranking straight into the Hermes memory graph.

You want to…	Run this
Rank the latest 3 weeks	wow-thats-tasty --weeks 3
Rank one paper	wow-thats-tasty --arxiv 2605.23904
Preview (no DB write)	wow-thats-tasty --dry-run
See all tasty papers	hindsight_recall("tasty papers tier 1")

---

Why it exists

Brandon’s KnowledgeCore vault used to live in Obsidian. Moving AI-papers there meant vault grep just to find Tier-1 papers:

grep "^tier: 1" 04-Reference/Papers/*.md

wow-thats-tasty kills that grep — the ranking now lives in a queryable sqlite table inside ~/.hermes/state.db, updated every weekday at 7:30 AM.

---

The Meal Plan

🍽️ Appetizer — Tasty (Tier 1, implement now)

Rank	Paper	Why
1	SkillOpt	Trainable external state for skills
2	Self-Harness	Weakness → proposal → validation
3	Harness Effect	41 % orchestration cost cut
…	…	…

🍰 Dessert — Sweet-but-heavy (Tier 2, later)

Paper	Decision
Generative Skill Comp	After SkillOpt is running
Bad Memory in Agents	After >500 bookbag entries

☕ Coffee — Bitter (Tier 3, skip)

Paper	Why skip
Red Queen / Gödel Machine	Too abstract, no recipe
Replicating ML Papers	Meta-framework, no action

The tasty ones become Hermes skills via book-to-skill: /learn <pdf> → ~/.hermes/skills/tier-1/<paper-slug>/SKILL.md

---

How to install

# clone into your Hermes skill dir (already auto-discovered on startup)
git clone https://github.com/branben/wow-thats-tasty.git \
  ~/.hermes/skills/wow-thats-tasty

# OPTIONAL: register a git tap so Hermes can update it
hermes skills tap add branben/wow-thats-tasty

# verify
skill_view(name='wow-thats-tasty')

Or use the Hermes one-liner:

hermes skills install https://github.com/branben/wow-thats-tasty

---

Usage

CLI

# rank last 3 weeks (writes to ~/.hermes/state.db)
wow-thats-tasty --weeks 3

# single paper
wow-thats-tasty --arxiv 2605.23904

# preview only
wow-thats-tasty --weeks 1 --dry-run

Flag	Default	Description
--weeks N	1	How many weekly editions to process
--arxiv ID	—	Rank a single paper by arXiv ID
--dry-run	off	Print ranks without writing the DB
--vault PATH	KC vault	Cross-reference existing tier: frontmatter

Cron (auto-runs daily M-F 7:30 AM)

Job name	Expression	What it does
Tasty-Papers Concierge	30 7 * * 1-5	Fetches latest papers → ranks → indexes

Find it with cronjob(action="list", name="Tasty-Papers Concierge").

Query the graph (no Obsidian!)

# tasty papers this week
sqlite3 ~/.hermes/state.db \
  "SELECT json_extract(value,'$.title') FROM state_meta
   WHERE key LIKE 'paper:%' AND json_extract(value,'$.rank')='tasty'"

# semantic search via the runtime tool
hindsight_recall("tasty papers tier 1", project="knowledge-core")

---

Architecture

  GitHub (dair-ai/AI-Papers)  →  concierge.py  →  state.db (state_meta)
                                                        ↓
                                              hindsight_recall("tasty …")
                                                        ↓
                                           /learn  →  book-to-skill  →  per-paper SKILL.md

---

Ranking logic (priority order)

1. KC vault tier (authoritative) — if the paper lives in 04-Reference/Papers/*.md with tier: N, use that tier.
2. Keyword heuristic — tasty hits agent-infra terms (memory, routing, orchestration, harness, verification); bitter flags abstract/theory-only papers; sweet-but-heavy catches improvement-flavored abstracts without a concrete action.

---

Stacking with book-to-skill

                ↓ tasty-rank from memory graph
  hindsight_recall("tasty papers tier 1")
                ↓ /learn <pdf>
  book-to-skill/scripts/extract.py
                ↓
  ~/.hermes/skills/tier-1/skillopt/SKILL.md
  ~/.hermes/skills/tier-1/self-harness/SKILL.md
  ~/.hermes/skills/tier-1/agent-as-router/SKILL.md

Each tasty paper → its own skill you can /load in any Hermes session.

---

Development

# test
python3 scripts/concierge.py --weeks 1 --dry-run   # 10 papers, no DB write
python3 scripts/index_paper_rank.py                  # vault → graph mirror

# run the CLI wrapper from anywhere
wow-thats-tasty --help

---

License

MIT — see LICENSE