dcostenco/prism-coder/prism-startup

Prism Startup

Use this skill once, on the first user turn of a conversation. A greeting such as "hi", "hello", or "ready?" is still a first user turn and must activate it. Do not run it again after startup context has already been shown in the same conversation. This re-entry rule covers startup only; a save refused with context_not_loaded is recovered as described under "Recovering from context_not_loaded" below.

Portable native-skill activation starts when the host processes the first user turn. It cannot display anything before the developer sends that first input. That first-turn behavior is the cross-host contract.

Provisioning contract

This skill ships with the prism plugin, alongside the Prism MCP server registration. The prism connect CLI installs the same protected skill for hosts configured that way; when both are present, the re-entry rule above keeps startup to a single display. Startup must not modify a host's lifecycle configuration. The skill remains available on every subscription tier, and it is intentionally absent from session skill routing so loading context cannot inject this startup procedure recursively.

First action: bootstrap context

Before greeting the developer or answering their task:

1. If session_bootstrap is available, call it with no arguments. Do not guess

a project, role, developer name, or context depth.

1. If session_bootstrap is unavailable but session_load_context is

available, call session_load_context for the current project and omit the level argument. Use the repository or workspace name as the project; use prism-mcp only when no project can be derived.

1. If MCP tools are unavailable but the prism CLI is available, run

prism load <project> --json and omit --level.

1. If the first attempt fails, correct the parameters and retry once. If the

retry and fallback both fail, display exactly: ⚠️ *Session context is temporarily unavailable.*

The level must be omitted in every startup path. Prism resolves the current dashboard setting instead of allowing this static skill to override it.

Context-depth contract

Render only the data returned for the resolved depth:

• quick: greet with the configured name and show the compact current state,

including open TODOs or keywords when present. Do not claim session summaries were absent; quick intentionally does not fetch them.

• standard: also show the last summary and up to five recent sessions.
• deep: also show the complete returned session history, up to fifty entries,

including the extra decisions, TODOs, and changed-file detail supplied by Prism.

Do not make a second load call to expand the selected depth. An empty field is different from a field intentionally omitted at the selected depth.

Greeting and context output

If session_bootstrap returns a ready-to-display greeting/context block, print the complete bootstrap result verbatim as the entire first-turn startup display, before any task answer. Do not summarize, paraphrase, rename headings, reformat, or omit any returned section. Preserve its order and line content.

If no ready-to-display block is returned, render the structured fallback result in this order:

1. Hi, <agent_name>. using the configured agent_name. If it is empty, use

Hi. without inventing a name.

1. **Last session:** <last_summary> when the selected depth returned it.
2. **Open TODOs:** followed by the returned TODOs, when present.
3. **Recent sessions:** for standard results or **Session history:** for

deep results, preserving the returned order.

When the first user input is only a greeting, stop after the verbatim startup display. If only the structured fallback was available, stop after that structured startup display. Do not add another greeting, question, acknowledgement, or other prose. When the first input contains a task, continue directly into that task after the startup display.

Foundational skill verification

Context loading may return entitled [📜 SKILL: <name>] blocks. Use only the skills explicitly reported as provisioned or returned in those blocks. Never guess a package name, request a missing package, or treat a package from an old session as currently entitled. Paid behavioral and engineering packages are delivered by the authenticated manifest; the public startup path relies on the compact safety and evidence contract in the MCP server instructions.

For bug reports or screenshots, run a real diagnostic before explaining the cause. Never report "done", "fixed", or "working" without observable verification.

Recovering from context_not_loaded

Startup runs once, but a later session_save_ledger or session_save_handoff can still be refused with context_not_loaded: after a long idle gap, after a restart, or for a project that startup did not load. That is not a second startup and the re-entry rule above does not apply. Call session_load_context with the same project and the same conversation_id as the refused save, then retry the save once. Do not repeat session_bootstrap or print a startup display for it. If the reload or the retry fails, tell the user; a local note is a fallback after that, never a substitute for the reload.