Persistent instructions are the always-on context you give your artificial intelligence (AI) client once so you stop repeating yourself. Clients name this differently — project instructions in Claude, custom instructions in ChatGPT, rules in Cursor — but the mechanic is the same: a short block that stays in context on every turn.
Two or three of these change how the Fullcast Model Context Protocol (MCP) server feels to use. Each instruction below is written to be copied as-is. Replace the bracketed values with yours.
Where these go
Use whichever field your client holds persistent instructions in. Fullcast supports Claude, ChatGPT, Agentforce, and Cursor — refer to your client's own documentation for where that field lives.
Keep them short. Instructions are in context on every turn, so length here costs you on every question, including the ones where the instruction is irrelevant. Anything longer than a paragraph belongs in a skill.
Note
For setup steps, refer to Set up the model context protocol (MCP).
Start with two
Pin your plan and require attribution
Always work in the [PLAN NAME] plan unless I name a different one. Do not
ask me which plan to use.
State which plan each result came from. If a lookup resolves against a
different plan than the one I named, say so explicitly before giving me
the answer. Ledgers are global and span plans, so commission and crediting
objects may not live in the same plan as territory objects.Why: A plan is the container for a go-to-market strategy in Fullcast — distinct from a compensation plan, of which one go-to-market plan may have several. Tenants accumulate them: prior years, model plans for scenario work. Naming yours stops you answering the same question every turn. The MCP does default to the plan you last worked with in the user interface (UI), but that default moves as you work in the app.
The second half is the guardrail, and it matters more than the first. Some lookups resolve outside your named plan no matter what you set, because commission and crediting objects live where they live. In live testing, commission questions asked in one plan silently resolved against another and came back labeled with the wrong plan. One session concluded a role received no credit anywhere, while an active crediting rule for that role existed in a plan the model had not looked at. Pinning the plan tells the MCP where to start; attribution tells you where it ended up.
Set your write posture
Before any write, show me the scope, the record count, and the records
affected, then wait for me to confirm. Never commit in the same turn as a
proposal. Tell me when an action has no proposed state to review.Why: Ask for the review step rather than a specific mode, because the platform decides the mode. Account, people, and product moves already default to proposed. Coverage changes take effect the moment they are called — there is no proposed state to review first. Ledger commits are irreversible: there is no uncommit, and a committed row is immutable.
Note
This instruction is a working convention, not a control. Fullcast enforces what you are permitted to do through role-based access control (RBAC) — the MCP acts as you, and a change your role cannot make fails at call time. It does not enforce an approval process on top of that. If a change must be impossible for someone, that belongs in their Fullcast permissions.
Worth adding once those are working
Report timestamps in your time zone
Report all timestamps in [TIME ZONE], not coordinated universal time (UTC).Every timestamp comes back in UTC by default, and converting them mentally is a small tax you pay dozens of times a day.
Tip
If a converted timestamp does not match what the Fullcast UI shows for the same action, raise it with support rather than assuming the MCP value is wrong.
Set your default output format
For any analysis longer than a short answer, produce an HTML file with
charts rather than a long text response. Include a summary at the top.Long text blocks are hard to read and harder to act on. Asking for HTML once, permanently, is the difference between an answer and a document you can send to someone.
Cap the length
Be concise. No preamble, no restating my question, no closing summary of
what you just did. Lead with the answer.Larger models get wordier as they get more capable. This instruction is worth more on a large model than a small one.
A starter block
Paste the two essentials, work with them for a few days, then add from the section above as you find you want them.
Always work in the [PLAN NAME] plan unless I name a different one. Do not
ask me which plan to use. State which plan each result came from, and if a
lookup resolves against a different plan, say so before giving me the
answer — ledgers span plans.
Before any write, show me the scope, the record count, and the records
affected, then wait for me to confirm. Never commit in the same turn as a
proposal. Tell me when an action has no proposed state to review.What to leave out
Long procedures. A step-by-step workflow belongs in a skill, where it loads only when relevant. In your instructions it costs you context on every turn.
Anything the MCP already knows. Fullcast terminology, the relationships between objects, and the traps in each domain are built into the MCP's own reference tools. Ask it rather than transcribing it.
Anything sensitive. No credentials, no tokens, no personal data. Instructions describe how you want to work.
Rules you have not tested. Add one or two, work for a few days, then add more. A dozen untested instructions interact in ways that are hard to untangle when something goes wrong.
Anything that applies to only one project or region. That is either a skill or a separate project.
Reviewing what you have
Instructions accumulate quietly and go stale. Two habits keep them useful:
When you catch yourself correcting the same thing repeatedly, check whether an existing instruction is contradicting the correction rather than adding a new one on top.
Reread the block after any plan restructure. An instruction that names a plan or territory that no longer exists fails in ways that are hard to attribute.