agentsclimarketplace

Subscription audit

Skill trac3r00/agent-guards/skills/subscription-audit

Agent Guards for long-lived AI agents: context-budget (audit per-turn token weight, fail CI on bloat) + claim-audit (flag unverified factual claims). Standalone tools + CI-ready, Claude Code / Codex / OpenCode compatible.

Install
npx -y skills add trac3r00/agent-guards --skill subscription-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 29 days oldThe repository was created 29 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Finds the recurring charges (subscriptions) hiding in a bank or credit-card CSV, tells you the true monthly and yearly cost, and flags the ones that look forgotten — a free trial that started charging, or a service that quietly kept billing after you stopped using it. Reads a statement you exported yourself; no bank login, no third-party service, fully offline and deterministic.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

6.4 KB, as published. Nobody here has run it

Subscription Audit

Re-read a year of statements the way no human bothers to.

Overview

Subscriptions are engineered to be invisible. Each charge is small, they land on different days, the merchant name is ADOBE *CREATIVE CLOUD 800-833 instead of "Adobe", and nobody sits down to re-read twelve months of a bank statement by hand. So the free trial that started charging, the gym you stopped going to, the app you used once — they just keep billing.

subscription_audit.py is the re-reader. Feed it a CSV you exported from your bank or card and it:

  • finds the recurring charges by clustering repeat payments to the same merchant at a regular cadence (weekly / monthly / quarterly / yearly), while ignoring one-off spending like groceries, gas, or a flight;
  • normalizes messy merchant names (AMZN Mktp US*RT4G9amazon, store numbers and auth codes stripped) so the same service groups together;
  • reports the true cost — monthly and annualized — across everything it found;
  • flags review candidates: a subscription whose expected next charge is well overdue (STALE), or one whose first non-zero charge was substantially smaller than later charges (trial→paid).

No bank login, no Plaid, no cloud. It reads a file you already have, is pure standard library, and is deterministic — the same statement always gives the same answer. A ruler for your recurring spend, not a budgeting app.

When to use

  • The monthly "where does it go?" question. Export a statement, run the tool, see the whole subscription stack ranked by cost in one screen.
  • A cancellation sweep. The STALE and trial→paid flags are your cut list — the tool even totals what cancelling the stale ones saves per year.
  • A budget guardrail on a schedule. Run it monthly with --budget N; it exits non-zero when detected recurring spend creeps past your line, so a cron job or CI step can ping you when subscription creep sets in.

Not for: categorizing one-off spending, forecasting, or replacing a real budgeting app. It answers exactly one question — what am I paying for on repeat?

The method

  1. Export a statement. From your bank or card, download transactions as CSV (any date range; the more months, the better the cadence detection). Most exports have a date, a description/merchant, and an amount column — the tool sniffs these by header name or by content, so you usually need no flags.
  2. Run it.
    python scripts/subscription_audit.py statement.csv
    # or pipe it:  cat card.csv | python scripts/subscription_audit.py -
    
  3. Read the stack. Subscriptions are listed heaviest-first with their monthly cost, per-charge amount, cadence, and last-seen date. The footer gives the total monthly and yearly burn.
  4. Review the flags. ⚠ STALE means no charge appears within the configured interval; trial→paid means the first non-zero charge was smaller than later charges. The summary estimates the yearly cost of the stale entries. Confirm each charge with the provider before cancelling it.
  5. Put it on a leash. Add --budget 80 to fail (exit 1) when recurring spend passes $80/mo, and drop that into a monthly cron so creep gets caught early.

Useful flags: --json (machine-readable, for a cron/agent to consume), --min-charges N (how many hits before a merchant counts as recurring; default 3), --stale-days N (grace period past the expected next charge before flagging), --currency € (display symbol).

Anti-patterns

  • Feeding it two months and expecting miracles. Cadence detection needs at least three hits of a merchant. A short statement will under-report — export a longer range.
  • Trusting STALE as "already cancelled". Stale means no recent charge in this file — it might be an annual plan that just isn't due, or a charge outside your export window. It's a prompt to check, not proof it's dead.
  • Treating the monthly total as your whole budget. This is recurring spend only; it deliberately ignores groceries, gas, and one-offs. It's not a full budget.
  • Cancelling on the merchant string alone. github and github sponsors may be two different things collapsed by name normalization — glance at the example description before you cancel.

Example

$ python scripts/subscription_audit.py statement.csv --budget 100 --stale-days 20
Subscription audit
======================================================================
  monthly      each  cadence    last          merchant
----------------------------------------------------------------------
$  54.99  $ 54.99  monthly    2026-06-15    adobe creative cloud
$  15.49  $ 15.49  monthly    2026-06-03    netflix
$  10.99  $ 10.99  monthly    2026-07-05    spotify usa ny
$  10.00  $ 10.00  monthly    2026-03-28    planet fitness  ⚠ STALE
$   4.00  $  4.00  monthly    2026-04-18    github  ⚠ STALE
----------------------------------------------------------------------
$  95.47  monthly across 5 subscriptions  (≈ $1,146/yr)

⚠ 2 look forgotten (no charge in a while): cutting them saves $14.00/mo (≈ $168/yr).
    · planet fitness — last seen 2026-03-28 (99d ago), $10.00/mo
    · github — last seen 2026-04-18 (78d ago), $4.00/mo

budget $100.00/mo  →  OK ($95.47 detected)

Groceries, gas, a flight, and an Amazon order in the same statement were correctly ignored — only the things that recur on a rhythm were surfaced.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.