krishpunjabi.com ~/skills

← all commands

/spend-gate

/spend-gate repo and safety

Gate any billable action (GPU pods/instances, paid APIs or tiers, cloud credits) behind a cost estimate, a live balance check, and an explicit yes for that action. Use before any create/deploy on a paid provider (runpodctl, vultr-cli, aws/gcloud/az, modal, provider MCP create/deploy/run/start/update), any paid API call, or when spend-guard blocks one.

  • author Krish
  • invocation model + user
  • prereqs None.
  • tested lightly used
  • suggested model · effort sonnet · medium

Install

Install just this, from a clone of the repo. Anything it needs comes with it.

./scripts/install-skills.sh --skill spend-gate   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Skill spend-gate   # Windows PowerShell

Needs: handoff-auto (installed with it).

What it does

Stops Claude from running a billable action (spinning up a GPU pod, calling a paid API, buying a paid tier, spending cloud credits) without first:

  • Estimating the cost.
  • Checking a live balance where possible.
  • Getting your explicit yes for that one action.

When to use

  • Anything that provisions cloud compute: runpodctl create, vultr-cli instance create, aws ec2 run-instances, gcloud compute instances create, az vm create, and equivalent MCP create/deploy/run/start/update tools.
  • Paid deploys (modal deploy) and per-token/per-request paid API calls.
  • Not for read-only checks (listing pods, checking a balance, viewing pricing) - those are how the skill informs its own estimate, not what it gates.

How to invoke

Claude runs it on its own before a billable action. You can also run it directly:

/spend-gate

How it works

  1. Classify - decide if the action about to run is billable; when ambiguous, treat it as billable.
  2. Estimate - unit price x hours/tokens = total, citing a price source (URL or tool output, dated).
  3. Check live balance/credit - read-only calls only, never a call that provisions or spends.
    • No free balance read → say so and ask you for the balance.
    • The step is never skipped.
  4. Gate - show the number, wait for an explicit yes to this action.
    • A budget cap or an earlier "go ahead" doesn't count.
  5. Log - append the approved amount, purpose, price source, and resource id to skilleddocs/spend.md.
  6. Run + remind - run the approved action, then remind about teardown or an auto-stop/idle-timeout for anything left running.
  7. Approval token - if hooks/spend-guard.py blocks the action, write the approval it expects (see HOOK.md) only after step 4's yes.

Design principles

  • A cap is not a yes. Nothing here treats a standing budget as blanket approval - every billable action gets its own ask, even a repeat of one already approved.
  • Estimate before you act, not after. The gate has to show a number before the user says yes, not a bill after the fact.
  • Prefer a hard stop over a guess. If a price or balance can't be verified live, the skill says so and asks rather than estimating from a remembered figure.

Shared context for these principles: Matt Pocock's video.

Use cases

  • About to provision a GPU pod for a training/eval run.
  • About to call a paid API in a loop (batch scoring, embeddings) where the per-call cost adds up.
  • Reviewing whether a repo's existing spend approvals in skilleddocs/spend.md still look right before continuing a paused job.

Tips

  • Cite the price source every time, even for a provider you've priced before - prices drift, and the log should show what you actually checked.
  • If the provider's balance endpoint is unreadable or rate-limited, say so in the gate prompt instead of skipping step 3 silently.
  • Pair with hooks/spend-guard.py (see hooks/README.md) if you want a hard stop at the tool-call level, not just a reminder in the prompt.

Example

/spend-gate

Claude is about to run runpodctl create pod --gpuType RTX4090.

  • It reads the live rate (illustrative: on one day a 4090 listed $0.89/h secure cloud, $0.34/h community; prices drift, so never reuse these).
  • It estimates a 2-hour secure job at ~$1.78.
  • It checks the balance via a read-only billing call (or says none is readable and asks you).
  • It shows both numbers, and waits.
  • You reply "yes, go ahead."
  • Claude logs the row to skilleddocs/spend.md, creates the pod, and reminds you to stop it (or that it will auto-stop) once the job finishes.

hooks/spend-guard.py - optional PreToolUse hook that enforces the same gate at the tool-call level.

  • It blocks a matching command/MCP call unless an approval file exists.
  • The approval file is written after the explicit yes at step 4, Gate; see HOOK.md.

Prereqs

None.

  • Reads whatever pricing/balance tools the billable provider already exposes (its CLI, MCP server, or pricing page).
  • Nothing extra to install for the skill itself.

Sample workflows

  • Run the expensive tests
    gap-scan → you → spend-gate → you → spend-guard → you
    Install just this one: ./scripts/install-skills.sh --workflow paid-tests or .\scripts\install-skills.ps1 -Workflow paid-tests
  • Push a score up until it plateaus
    score-loop → you → spend-gate → you → you → switch → remind → handoff-auto → you
    Install just this one: ./scripts/install-skills.sh --workflow climb-a-score or .\scripts\install-skills.ps1 -Workflow climb-a-score

Source

Use cases & tips from the community