api-cost-gate

Purpose

Solo developers and small teams using paid AI APIs (Suno, ElevenLabs, Midjourney, OpenAI, Tripo, etc.) face two failure modes:

1. Surprise bills — automated retries, runaway loops, accidentally-expensive calls (e.g. 10-minute music generation when 30 seconds was intended).
2. Wasted spend — calls made before the use is clear, producing assets that get discarded.

This skill enforces a pre-flight 4-point disclosure before any paid call. The agent cannot invoke the API until the user has reviewed and approved:

1. Call type — exact endpoint / mode / model.
2. Cost estimate — credits / characters / dollars (with current balance if known).
3. Purpose — what we're producing and why.
4. Use plan — where the output will live, how it will be evaluated, what triggers acceptance.

Auto-mode does not override this gate. Even when the agent is running autonomously, paid API calls require explicit user OK each time.

---

Phase 1: Identify the call

From $ARGUMENTS:

• $1 = service name (e.g. suno, elevenlabs, midjourney, openai)
• $2 = brief description (e.g. combat-bgm-30s, npc-dialogue-line, concept-art-key-visual)

If args missing, ask the user what call they intend to make.

---

Phase 2: Compose the 4-point disclosure

Produce a short, unambiguous block:

=== Paid API Call — Approval Required ===

1. Call type:
   <service> <endpoint> — <model> <mode>
   Example: Suno /api/v1/generate — V5, custom mode, instrumental

2. Cost estimate:
   ~<N> <unit>  (current balance: <M>)
   Example: ~12 credits  (current balance: 486)

3. Purpose:
   <what + why, 1-2 sentences>
   Example: Voice-of-Peace BGM bed for chapter 1 radio segment.
            First sample to lock chapter 1 musical tone.

4. Use plan:
   <where it lives + how it's evaluated + acceptance criteria>
   Example: Lands in Tools/SunoAPI/output/ as workshop. If A/B-compared
            against existing anchor (or pilot-anchored if first), promote
            to Documents/AudioDesign/Anchors/ with prompt.txt.

→ Approve? (Y / N / modify)

The 4-point block is engine-agnostic / service-agnostic. The pattern works for any paid API.

---

Phase 3: Wait for explicit user approval

Do not make the API call until the user responds with explicit approval. Acceptable approvals:

• Y / OK / 진행 / approved / go — proceed.
• N / cancel / stop — abort.
• modify <change> — adjust the call per change, re-show the 4-point block, wait again.

Treat ambiguous responses (silence, unrelated message, maybe) as not approved — do not call.

If $ARGUMENTS was originated by the user themselves with full intent (e.g. they typed /api-cost-gate suno combat-bgm-30s and immediately follow with Y in the same message), the gate is satisfied — but only if all 4 points are still presented and acknowledged.

---

Phase 4: Log the approval (optional but recommended)

After approval, append to a project-level cost log:

Documents/api-cost-log.md

| Date       | Service     | Call                  | Cost   | Result                                  |
|------------|-------------|-----------------------|--------|-----------------------------------------|
| 2026-05-03 | Suno        | combat-bgm-30s V5     | 12 cr  | Tools/SunoAPI/output/<id>/ — pending    |

This makes monthly billing reconcilable and surfaces calls that produced no kept output — a useful signal for prompt iteration efficiency.

---

Phase 5: Hand off to the actual call

Once approved, invoke the API call via whatever mechanism the project uses (CLI script, MCP tool, HTTP). Treat the call's output handling as a separate concern — this skill is only the gate, not the entire workflow.

---

When NOT to use this skill

• Free / local calls — local LLM, sandboxed model, no-cost endpoint. Gate adds friction with no benefit.
• Pre-approved batch operations — if the user explicitly approved "generate 10 variations" as a single batch, individual variation calls don't need re-gating; document the batch approval scope and proceed.
• Read-only / metadata calls — fetching credit balance, listing voices, getting account info. These are typically free even on paid services.
• The cost is genuinely trivial and bounded (e.g. <$0.01 with hard ceiling) and the user has standing approval. Document the standing approval explicitly.

---

Generalises across services

Confirmed services where this gate has prevented surprise spend or wasted output:

Service	Unit	Notes
Suno (music)	credits	1 call ≈ 10-20 credits, 2 variants per call
ElevenLabs (TTS / SFX / music)	characters	TTS ≈ 1 char / char, SFX ≈ 100 char/sec, Music ≈ 1000 char/sec
Midjourney / Stable Diffusion (image)	seconds / generations	varies by tier
Tripo (3D)	credits	varies by mesh complexity
OpenAI / Anthropic (text)	tokens	input + output, varies by model

Add your service to the table when adopted. The 4-point disclosure works identically for all of them.