prompt-session

Generate a concise prompt that the user will paste into another coding agent session. This skill also applies to follow-up messages to sessions that were previously prompted by this skill (e.g. additional instructions, corrections, or new context for a running session).

$ARGUMENTS may specify a role: architect, implementer, reviewer, investigator, researcher, secretary, or writer. If omitted, ask the user which role the target session will play.

Core rules (what NOT to include)

Assume the target session has the same instructions files, skills, and tools as this one, unless the user says otherwise. Do NOT restate:

• Global conventions (writing style, shell, tool preferences, etc.)
• Tool usage rules (gh for GitHub, search tools over shell, etc.)
• Git commit rules (author, sign-off, commit message style)
• Content of the plan issue (the target reads it via /gh-read-plan)

What to include

• One-sentence goal: what the target is being asked to do
• Plan reference: if there is a plan issue, cite it as owner/repo#N and instruct the target to run /gh-read-plan owner/repo#N at the start. Do not summarize the plan.
• Skill to invoke: if a ready-made skill covers the target's job (see roles below), instruct the target to invoke it. Do NOT restate the skill's rules in the generated prompt; the target loads those itself.
• Recall: if a durable-memory recall skill is configured, instruct the target to run it on the task's own terms before starting non-trivial work. /gh-read-plan recalls on the plan issue's terms; this adds the task-specific angle. Skip for trivial tasks: secretary updates, formatting fixes, mechanical edits.
• Pointers: branch, worktree path, key file paths, sibling checkouts to avoid
• Hard constraints: scope boundaries (no push, no PR create, no amend, no changes outside X, don't delete worktree, etc.)
• Expected output: only specify this when there's no skill for the role, OR when the user wants output that differs from the skill's default.