agentleFS
lglucaslglucas/ai-dev-operating-system/.claude/skills/cost-watchdog/SKILL.md

cost-watchdog

Skill12 starsChanged 48 days agoMITView on GitHub

Flag potentially expensive technical choices BEFORE they are committed — AI model selection, hosting tier upgrades, third-party services with metered pricing, and DB patterns that explode with scale. Preventive counterpart to `usage-monitor`, which tracks real spend after launch. Triggers when the user says "posso usar o Opus pra isso?", "qual plano eu contrato?", "isso vai ficar caro?", "quanto custa rodar isso?", or whenever a paid tier or metered API is about to enter the stack.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before introducing any third-party service (LLM provider, DB, queue, monitoring, search, image, email).

    line 10
  2. Distinctive:

    Before suggesting a paid tier of any service the user is currently on free tier of.

    line 11
  3. Distinctive:

    When proposing model selection (Opus vs Sonnet vs Haiku tier, or the equivalent tiers of another provider).

    line 12
  4. Distinctive:

    When implementing patterns that scale poorly: N+1 queries, polling loops, full-text scans, unbounded retries.

    line 13
  5. Distinctive:

    When the user says "está caro" / "está cobrando demais" / "fiquei sem créditos".

    line 14
  6. Distinctive:

    Loops are dangerous: If a feature calls the LLM in a loop, calculate worst case before shipping.

    line 22
  7. Distinctive:

    Streaming + early termination: when user might stop reading.

    line 23
  8. Distinctive:

    Don't enable database point-in-time recovery before paying users exist.

    line 29

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Cost Watchdog

When to run this skill

  • Before introducing any third-party service (LLM provider, DB, queue, monitoring, search, image, email).
  • Before suggesting a paid tier of any service the user is currently on free tier of.
  • When proposing model selection (Opus vs Sonnet vs Haiku tier, or the equivalent tiers of another provider).
  • When implementing patterns that scale poorly: N+1 queries, polling loops, full-text scans, unbounded retries.
  • When the user says "está caro" / "está cobrando demais" / "fiquei sem créditos".

Cost flags to watch

LLM costs

  • Default to the current Haiku or Sonnet tier for routine work, and lower effort before dropping a tier. Use Opus only for tasks that justify it (deep reasoning, business strategy, architecture). Check the provider's current model and pricing page before quoting a number.
  • Always cache prompts that have a stable system prefix (saves up to 90% of input tokens).
  • Loops are dangerous. If a feature calls the LLM in a loop, calculate worst case before shipping.
  • Streaming + early termination when user might stop reading.

Hosting

  • Vercel Hobby (free) is fine until ~100 users. Pro ($20/mo) only when you hit limits.
  • Supabase Free tier: 500 MB DB, 1 GB storage, 50K monthly active users. Plenty for prototypes.
  • Don't enable database point-in-time recovery before paying users exist.

Database patterns

  • Unbounded SELECT without pagination → flag.
  • Realtime subscriptions on noisy tables → flag.
  • Full-table scans in serverless functions → flag (cold-start tax + bandwidth).

Third-party services

  • Stripe: percentage-based, no flag below revenue.
  • Resend: 3K emails/month free. Flag transactional volume design.
  • PostHog free tier: 1M events/mo. Flag if event firing in loops.
  • Sentry free: 5K errors/mo. Flag noisy retry catch blocks.

Vibe coder explanation

Cloud services are like a cab meter. Some run flat-rate (you pay even idle), some run per-trip (per request), some are free up to a limit then suddenly expensive. Before you ship, I'll show you which meter is running and what the worst-case bill looks like if your idea goes viral.

Output

💰 cost-watchdog flags

[HIGH]  Using Claude Opus inside a loop that may run 50+ times per user action.
        Estimated cost: $0.45 per user action. At 1000 users/day = $13,500/month.
        Suggest: switch to the Sonnet tier or the Haiku tier; check current provider pricing before quoting the per-action cost.

[MED]   Realtime Supabase subscription on `messages` table without filter.
        Will fire on every insert across all users.
        Suggest: filter by chat_id.

[OK]    Stripe checkout: percentage-based, no fixed cost.
[OK]    Vercel deploy: still on Hobby free tier.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.