Prepare prompt

Write for a session holding only this text — no chat history, no memory of what was meant — and make it as short as that reader allows: nothing it needs is cut for brevity, nothing else stays. The reader is a fresh interactive session the user pastes it into; a subagent brief only when the invocation asks for one, under the same rules.

Input

The user's goal, in a line, plus this session. Draft at once; ask only what neither settles, one question at a time, with a recommended answer.

Content

A suggested order, every part optional: a part is never filled to exist, and two lines are a complete prompt when two lines cover the case.

1. Situation — what exists that the next session builds on, and what is open or was missed, each with its identifier (ticket URL, branch, commit, PR, task ids).
2. Task and deliverable — what to do, in what form to report or produce it (a chat report with a given structure, a file, a code change) and, when the work continues, the check that says it is done (a command, a criterion, a file to exist).
3. Latitude and boundaries — what it may decide alone, when deviating is fine, what it must not touch (read-only, no git or tracker writes, files to ignore).
4. Facts it cannot derive — paths, the planning directory, docs to read, tool quirks with the workaround, a command that saves time.

The deliverable form and the write boundaries are the usual gaps: check whether the case needs them.

What stays and what goes:

• A fact the next session builds on takes one line, never how it was reached. Anything it will redo anyway, and any narration of this session, goes.
• When its job is to check, verify or review something, this session's own conclusions about that thing go entirely — they would steer the check. When its job is to continue the work, the decisions it must build on stay, one line each.
• Point, never paste: what exists in a file or commit is referenced precisely (path and section, hash), never copied. Only what lives nowhere — a decision made in the chat, a tool quirk, a useful command — is written out. Never a credential: one the next session needs is named by where it lives (env var, config path); one that lives nowhere, such as a key pasted in chat, is left out and the prompt says to get it from the user.
• When an installed skill covers what the next session must do, the prompt opens with its slash command and the rest follows as its arguments (/refine-ticket 1234 …) — not a description of the procedure, nor prose like "run /refine-ticket", which cannot load a skill that blocks model invocation. Verify it exists first. A subagent cannot run a slash command: its brief names the skill to invoke instead, and a task whose skill blocks model invocation stays with an interactive session.

Form

Plain imperative prose in paragraphs, no headings, bullets for lists. Agent-facing, so no conversational voice. Print it as one fenced block, commands and paths inline in backticks (never the opening slash command), never a nested fence; then a --- line before any note to the user; every revision reprints the whole block.

Check before printing

• Every path, id and command in the prompt exists or works — verified by a read-only lookup, not remembered.
• Every fact stated as settled was verified in this session or stated by the user; an inference is marked as one ("probably", "unverified") or cut.
• Nothing narrates this session; for a checking job, nothing states this session's conclusion about the thing checked.
• A session holding only this text can act without asking what was meant.

Done when all four hold.

Boundaries

• The printed prompt is the product: write no file unless asked.