# Ns Stay On Track

> Keeps NetSuite ns_* MCP tool use on one path and conserves tokens. Use when calling ns_* tools, after a failed or unhelpful result, or when about to switch ns_* tools or invent a new angle. Pause, tell the user, and wait instead of looping until max iterations.

- Skill: `unstackedapps/ns-stay-on-track` (Agent Skill)
- Install (CLI): `npx skillmds@latest add unstackedapps/ns-stay-on-track`
- Raw SKILL.md: https://api.skillmd.com/api/skills/unstackedapps/ns-stay-on-track/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: unstackedapps (https://skillmd.com/u/unstackedapps)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/unstackedapps/ns-stay-on-track

---


# Stay on track with ns_* tools

While this skill is enabled, these rules are mandatory for every persona. They override default error recovery, “do not stop early,” and any playbook that would keep issuing `ns_*` calls after a miss.

This skill is only about NetSuite MCP tools (`ns_*`). It does not apply to other MCP servers. It does not teach which NetSuite tool to prefer. It teaches when to stop spending tokens and check in.

## Job

The harness often cannot recover from a live NetSuite miss. Two starts:

1. The `ns_*` call fails (transport, unknown or truncated name, or a business error inside a Completed payload).
2. The call succeeds but the result does not answer the question, or it only points at a different object, report, or search.

Do not hop to another `ns_*` tool or theory to “get back on track.” That oscillation burns tokens and ends at the step-limit lock. Notice you are off course, stop, and tell the user.

**One path, one same-tool fix, then check in.**

## Call only real ns_* tools

- Call only `ns_*` names listed in this session’s connected NetSuite MCP tools.
- Never invent an `ns_*` name. Never call a truncated name (`ns_`, `ns_run`).
- If the intended tool is not connected, say so. Do not guess a neighbor name.

## One path

- One intended retrieval path per user turn (discover → one execute). Not a tour of every `ns_*` tool class.
- At most one distinct fix on the **same** tool (missing dates, period, or subsidiary; or replace an invented ID with an exact ID from a list that already ran).
- Treat an unhelpful success as a failure: no asked-for facts, or only a tangent to chase.
- Every extra `ns_*` call costs tokens and usually does not recover the harness. Prefer a short check-in over another tool.

## Reality check

Before a second approach, a different `ns_*` tool class, or a “this might help” call, stop and tell the user:

- what you tried
- what came back or failed
- why that does not answer their question
- that you are stuck, or that you want to try one specific other `ns_*` tool
- **one** choice: approve that tool, change a filter, or stop

Wait. Do not silently switch.

## No more ns_* calls this turn

Stop calling NetSuite MCP tools when any of these is true:

- unknown, truncated, or invented tool name
- same error text or same missing field after the one fix
- permission, auth, or tool-not-available
- write tools (`ns_createRecord`, `ns_updateRecord`, deletes) — never auto-retry
- result does not answer the question (wrong grain, empty, or a tangent)
- already switched approach once this turn
- about to call a tool because the last result “might” help
- three `ns_*` calls this turn and no answer in hand

Parse tool `content` even when the call looks Completed. A payload `error` counts as failure.

A short honest stop is success. Hitting the step-limit lock is not.