deal-hunter

AI skill for smart buying in India. Find the best-value product (not just the cheapest), verify the deal is real, find the cheapest way to pay, and keep the whole thing tracked - from first hunt to the warranty claim two years later.

Works with any AI agent that supports skills (Agent Skills-compatible tools).

---

What it does

Area	What you get
Hunt / compare / deal-check	3-5 real candidates across Amazon.in, Flipkart, and category platforms, scored value = features ÷ price
Worthiness gate	Asks "is this purchase worth it to you?" before any research - the answer can be Don't buy, no matter the price
Deal verification	Real price history vs inflated MRP, review-integrity checks (bursts = manipulation), community long-term-ownership checks (Reddit-first)
Market shock awareness	Lightweight news sweep per hunt — category-wide shocks (RAM/memory shortages, component-cost hikes, launch inflation) reset targets instead of false "fake discount" flags
Pay Smart	Effective price after card offers, UPI cashback, coins, cashback apps, EMI / no-cost EMI (with its hidden fees), split payment - you decide on what you actually pay
EMI readiness	Committed monthly EMI + new EMI vs your ceiling → Ready / Almost / Not ready, with the numbers
Used / refurb	On-spot test plan, mining-card & stolen-device signals, warranty-transfer reality
Imports	Landed cost (customs/IGST, courier vs postal) vs buying local - imports are a separate tier
Price alerts	Keepa / pricehistory.in / deal bots, tied to the next real India sale
Claims & warranty	Invoice kept, warranty-registration deadlines, escalation: brand → NCH 1915 → consumerhelpline.gov.in → e-Daakhil
Repair tracking	Every repair logged; repair-vs-replace (fix / replace / do nothing) before spending
Subscription audit	Recurring-spend inventory with worthiness per subscription
Restock / rebuy	Researched once, stocked forever — a rebuy reads the Product Master (per-unit benchmark) instead of re-hunting; every buy logs a FIFO lot
Record & recall	One row per purchase (deal + claim + EMI + repair + restock pointer), then "what's the status of X?" is answered from the record, not by re-researching

What it does NOT do

• Does not buy anything - it researches and recommends
• Does not trust MRP, strikethroughs, or star ratings as evidence - everything is cited or flagged unverified
• Does not invent prices, reviews, offers, or sale dates - sale dates/offers are re-verified live at research time
• Does not file claims, run credit checks, or guarantee financing/warranty approval - it prepares the evidence and gives the escalation path
• Never writes into its own files - assets/*.csv and the skill's own files are read-only templates; your records go to the config-driven route (Obsidian vault Trackers//notes, the workspace, or a copy-paste row)

How it works

hunt → verify hard → score & compare → pay smart (effective price)
  → pipeline budget (out + queued + EMI + new) → verdict
  → record (config-driven route: MCP → local vault → workspace → chat)
  → recall (later "status of X?" answered from the record)
  → restock (rebuy of a recorded product = benchmark check, not a new hunt)

Verdicts are exactly one of: Buy now / Wait for sale / Pick alternative / Don't buy - with a target price, the next real India sale, the best alternative, and a purchase-mechanics + safety note for the recommended buy.

Install

The skill is a folder with a SKILL.md — copy it into your agent's skills directory (or upload the zip):

git clone https://github.com/yeahnowhere/deal-hunter.git
cp -r deal-hunter /path/to/your-agent/skills/

Or download the latest .skill zip from Releases and upload it through your agent's interface.

Install only from trusted sources, and audit the bundled files before use.

Configure storage (first run)

The skill saves your results automatically to a config-driven route — it never re-asks "where do you want to save?" every hunt. Set it up once:

1. Open scripts/config.json in the skill's folder.
2. Set storage.vault_paths.pc to the absolute path of your Obsidian vault on this machine (e.g. C:\Users\<you>\Obsidian\Vault), and storage.workspace.dir if you want a non-default workspace.
3. Done — the AI reads that same scripts/config.json from the skill's scripts folder whether you use the skill via the web or the CLI. No script to run; no other setup.

The save route is picked at runtime, in this order:

1. Obsidian MCP connected → writes the trackers to your vault by vault-relative path (Trackers/<file>, notes in Journal/) — works on any synced device. (Nothing to configure.)
2. Local vault path → writes to <vault_paths.pc>/<tracker_dir>/… and <…>/<notes_dir>/….
3. Workspace — local CLI → writes to the skill's in-skill workspace/ (trackers/ + notes/); path from storage.workspace.dir (default, overridable).
4. Workspace — web-only → writes to the web workspace/project, folder literally named deal-hunter (trackers/ + notes/) — always that name.
5. Chat → if the agent can't write files, you get a copy-paste row + which file to paste it into.

[!note] Web-only user (no local Obsidian, no skill folder) The AI saves everything there is to save straight into a workspace folder literally named deal-hunter in your web/cloud workspace (e.g. a Claude Project), split into two subfolders:

• trackers/ → the tracker files + my-deals.csv + the restock ledger (Household & Restock Tracker.md)
• notes/ → every individual product note (<product>.md) That name is fixed — the AI never saves into any other folder, so you always know where your whole buying record lives. Just create the deal-hunter folder once (the AI will remind you), and it reuses it every hunt.

Key config fields (scripts/config.json):

• storage.prefer_mcp — use MCP automatically when an Obsidian MCP is connected (default true)
• storage.fallback — route when no MCP is available: pc (use vault_paths) or workspace
• storage.vault_paths — per-machine absolute vault paths (e.g. pc)
• storage.mcp.tracker_dir / notes_dir — vault-relative folders (defaults Trackers / Journal)
• storage.workspace.dir — local CLI workspace root (the skill-internal workspace/ default; overridable). Web-only always uses the fixed deal-hunter folder instead.
• restock_file / restock_csv — restock ledger filename (default Household & Restock Tracker.md, lives in the tracker folder) and its portable CSV template (restock.csv).

An explicit instruction ("save this to X") always overrides the configured route. The skill ships with a placeholder path — config.json uses <you>; set storage.vault_paths.pc to your real path in the deployed/installed copy. Your real paths live only in your copy, never in the publicly shipped skill.

Quick start

find me the best <product> under ₹<budget>, must-have: <spec>
is this a good deal? <URL>
compare these: <A> <B> <C> - which should I get?
am I ready for <product> ~₹<price> on EMI? (my ceiling is ₹X/month)
help me claim warranty on <product> - what's the escalation path?
restock the <product> - I bought it last month, find me a deal for it
what's the status of <product>? / where is my claim?

Data & privacy

• Nothing leaves your machine. The skill researches the open web; all your records are saved via the config-driven route (your vault trackers, the workspace, or my-deals.csv) — never into the skill's own files.
• my-deals.csv is the portable record - a single merged file (deal + claim + EMI + repair + restock pointer, one row per purchase) that any AI can read back on a later session.
• On platforms where the agent can't write files, it gives you a copy-paste row to drop into your file.
• The skill's scripts/config.json ships with <you> placeholders — set storage.vault_paths.pc to your real vault path in your installed copy. Never write your real paths into the skill's own files.

Project layout

deal-hunter/
├── SKILL.md          # the skill - frontmatter tells the agent when to use it
├── references/       # playbooks the agent loads on demand
├── assets/           # read-only CSV templates (my-deals, deal-tracker, emi, repairs, restock)
├── scripts/          # storage config (config.json) — set storage.vault_paths.pc
├── README.md         # this file - repo-facing only, not part of the skill
└── .github/          # CI (validate + release)

Reference playbooks

File	Load when…
prompt.md	analyzing a specific shortlisted product (JOB 0-11 copy-paste prompt)
india.md	any India-market buy - platform map, sale calendar, scams
community.md	verifying long-term ownership (Reddit-first, paid-plant detection)
finance.md	any financed purchase - effective price, no-cost EMI reality, offer-stacking
emi.md	any financed purchase or while EMIs are active - readiness verdict
repairs.md	a device breaks or needs a repair - repair-vs-replace
claims.md	a product breaks or arrives wrong - warranty → NCH → e-Daakhil
used.md	used/refurb candidates - on-spot testing, mining/stolen signals
import.md	international listings - landed-cost math vs local
alerts.md	every Wait verdict - price alert + sale trigger
market.md	every hunt (lightweight) - market-shock detection: shortages, category-wide price resets, shrinkflation
companions.md	every verdict - does the product need an essential companion? surface, don't auto-hunt
subscriptions.md	subscription/recurring-spend questions
tracker.md	recording a hunt result - schema, pipeline rule, watchlist
restock.md	restock / rebuy / "bought this before" - benchmark the Product Master, no re-hunt
recall.md	any "status of X / where is my claim?" - read the record, don't re-research
environment.md	agent environment & storage - the config-driven save routes (MCP → local → workspace → chat)
examples.md	seeing the pipeline in action - 8 worked examples (Buy now, Wait, Don't buy, Compare, Pay Smart flip, EMI flip, Quick Hunt, Market-shock flip)

Development

• Docs-first: draft behavior changes in the project docs, mirror them into SKILL.md/references/, then deploy.
• Validate: sync-dealhunter.ps1 -CheckSync verifies deployed = master = git HEAD; CI runs the same frontmatter, reference-integrity, and leak checks on every push.
• The .skill zip is built on release tags; the repo stays free of build artifacts.

License

All rights reserved. This repository is public for viewing only; no license grants permission to copy, modify, or redistribute the contents.