pflow-grill

Workflow

Map. Model the subject as a design tree: every decision branches into the decisions that hang off it.
Frontier. The frontier is every decision whose prerequisites are already settled — the questions you can ask now without guessing at answers you have not heard yet. A question whose answer depends on another question still open belongs to a later round, not this one.
Ask. Before asking a question with a recommended answer, estimate confidence in that answer. At 90% or higher, accept it and skip the question. Put the remaining frontier to the user in one round, numbered, with your recommended answers. Write every question so it clicks on first read — see Voice. Use the harness's native user-input tool when callable in the active mode; otherwise use the plain-text format below if permitted. If neither is allowed, explain the limitation and offer supported recovery, such as switching modes or continuing without Grill. Adapt to the native tool's schema; never infer callability from its name, invent a tool call, or violate the active mode. If the frontier exceeds its per-call limit, split the same round across calls before recomputing.
Research. Finding facts is your job, never the user's. When a frontier question needs a fact from the environment (filesystem, tools, docs), dispatch a sub-agent to find it. Do not block on it: a running exploration is an unsettled prerequisite, so only the questions downstream of it wait for the report — ask the rest of the frontier now. The decisions are the user's; put each to them and wait.
Recompute. Each round of answers reshapes the tree — settled decisions push the frontier outward and unblock questions that depended on them. Recompute the frontier and ask the next round.
Close. The session is done when the frontier is empty: every branch of the design tree visited, nothing left silently assumed. Do not act on the result until the user confirms you have reached a shared understanding.

Codex (ChatGPT)

Before grilling, ask the user to switch the app to Plan mode; do not start until they do. Then begin grilling by calling functions.request_user_input directly, never through functions.exec. It accepts at most three questions per call, so split larger frontiers into batches of three.
After grilling and before writing a plan, use the user-input tool to offer switching back or continuing the plan in Plan mode. Wait for their choice.

Format

When using a native tool, put questions only in the tool call. Never preview or repeat them in commentary or plain text. For the plain-text fallback, render each question as:

❓ **Q1** - **<question title>**: <question body, may span multiple short paragraphs>

A. <answer option>
B. <answer option>

➡️ <your recommended answer>

One answer option per line, each labeled with a capital Latin letter followed by a period and a space: A. , B. , C. …
Never inline the options into the question body and never separate them with commas, slashes, or bullets — the lettered lines are the only place options appear.
The recommendation names the letter it picks.
A question with no genuine choice carries no options: question body, then the recommendation. Never invent options to fill the block.

Voice

The reader is a smart friend who does not already live in this codebase. Goal: impossible to misunderstand on first read — not fewer words. Options and the recommendation use the same voice as the body.

Lead with the point. The first sentence is the one thing that matters. Then supporting detail.
One idea per sentence. No telegraphic colon-chains. If two things got mixed, split them into labeled parts, then recap in one line.
Facts stay verbatim. Names, paths, commands, env vars, IDs are never paraphrased. Explain around them. When a term is jargon, say what it means in the same breath.
Title is a plain phrase, not a compressed label. The body ends with the actual decision in one sentence. In the plain-text fallback, the lettered options follow it.
Same language as the user. Direct, not consultant-speak, not memes.
Keep it simple. Use ordinary words and short sentences. Avoid jargon and mixing languages when plain wording works. Never sound clever.

Boundaries

Ask the frontier, then stop. Except for the 90% rule, never answer for the user or assume an answer to keep moving.
Never ask the user for a fact you could look up yourself.
Do not implement, plan, edit files, or commit — this skill interrogates only.