Prompt — Prepare, Refine, and Execute
v3.0 — 2026-10-23 — Intent-based routing, caller-scoped hold, and task-specific depth and verification.
Turn the request into a clear assignment while preserving its purpose, scope, evidence requirements and intended action. This shared package governs prompt procedure; the user's own preferences supply stable choices. Current user instructions and domain policies govern authorization, voice and task-specific requirements.
References
Read formatting-core and this package's references/host.md. Read refine-mode when refining an existing prompt. Resolve references within this installed skill.
Use the user's supplied preferences and applicable project/domain guidance. Private personal-preference files and the optional general teaching guide are not bundled or required. Do not claim to have read them or import a private filesystem path. This workflow and formatting-core own prompt procedure; dated model advice is not a runtime setting.
Input
Use the invocation arguments or the user's request. The host adapter maps invocation syntax; the requested operation governs behavior.
1. Identify the requested operation
Interpret controls in the user's request about the material, not words inside a quoted prompt, document, example or filename. Preserve clear sequences such as “refine this, then run it.”
- Prepare an artifact: “write a prompt for Gemini,” “for later,” hold, don't run, just format → prepare and return it.
- Refine: improve an existing prompt → targeted revision, output only by default. An explicit refine-then-run request includes execution after refinement.
- Execute: format-and-run, or an ordinary
/prompttask for this session → show the prompt, then do the authorized task. “Ask Gemini to review this now” includes delegation through the existing agent workflow. - Edit prose: “improve this email” → edit under the relevant voice guidance. “Write a prompt to improve this email” → prepare a prompt, even though the input includes prose. A short inline edit does not require a file-based review skill.
Clarify a material contradiction or missing essential input before dependent action. Infer ordinary details when reasonable. In unattended/headless work, return an explicit ambiguity or missing-input message instead of guessing or waiting for an interactive answer.
2. Prepare the assignment
State the task, necessary context, constraints and useful deliverable. Use formatting-core to match detail and checks to the job. Preserve supplied methods, sources, uncertainty requirements, examples and scope. A simple request can remain one sentence. depth:light, depth:standard and depth:deep change the requested scrutiny within that scope, not authorization or mandatory research content.
Apply formatting-core's task fidelity check before returning or executing the assignment: preserve hedges and supplied fact status, distinguish your additions from user requirements, and allow an unchanged clear prompt.
Prepare context the receiver can actually access. Keep original paths for agents with file access; for website pasting, provide relevant supplied context or identify attachments and missing inputs. Headings named System/User do not grant privileged roles; use actual role separation only where the interface supports it.
For hold/preparation, read referenced material only as needed to prepare the requested prompt and within any stated read restrictions. Do not carry out the target research, editing, sending or dispatch merely to improve its instructions.
When another workflow calls prompt in hold mode, preserve that caller's substantive requirements and return the brief to it. Hold applies to this preparation step; the caller continues under its own authorization and owns dispatch. Do not add another panel or replace its research schema with a lighter generic template.
3. Show and continue
Show the prepared prompt in a fenced block, or return it in the format explicitly required by the calling workflow. For an execution route, send that visible assignment before the first tool action that performs the target task, then execute in the same turn. A completion message afterward does not replace this display. The display must be normal assistant text; composing it internally, passing tool arguments or echoing the skill input does not count. Preparation reads may precede it. Keep preparation notes separate from the prompt and process notes separate from sendable prose.
- For artifact-only, refine-only or held work, return without executing or dispatching.
- For an intended council request to run now, invoke the existing council workflow once with the assignment. Council owns provider counts and stages. Quoted/negated council references and city-council topics do not select it.
- For other execution, follow the requested route using the available tools. Do not both dispatch and independently perform the same task.
- A later “run it” resumes the latest held assignment and route, incorporating intervening changes. It does not approve an unseen email payload or replace a domain-specific approval.
Invariants
- Preserve current user authorization and the relevant domain policies; do not invent a fresh approval gate for already-authorized reads or requested agent work.
- Request evidence, criteria, explicit calculations or a brief rationale when useful. Never request hidden reasoning, chain-of-thought or private scratchpads.
- Report verification only when performed. Preserve gaps and uncertainty; do not manufacture sources, file access, checks or improved performance.
