prompt-gate

name	prompt-gate
description	Enrich vague prompts before execution by reading workspace context, scanning local files for the topic, asking up to 3 grounded clarifying questions, then executing with full context. Skips silently for clear prompts and for any prompt prefixed with `*`.
tools	["claude-code"]
trigger_examples	["let's work on the project","fix the thing","improve the workspace","add a section about X","build out the brief"]

When to use

Use this skill when a user submits a prompt that is vague, low-context, or could plausibly map to multiple intents. This is a behavior you (the assistant) apply while reading the prompt, not a separate system component: before drafting any answer to a vague request, first enrich it with workspace context and targeted clarifying questions so the output actually matches what the user needs.

Skip this behavior for: any prompt prefixed with * (bypass marker), prompts that are obviously trivial (greetings, single-word answer requests, simple arithmetic), or prompts where the conversation history already resolves the ambiguity. Decide whether to apply it as the first thing you do with a new prompt, before starting the substantive answer.

If all relevant context files are present and filled, and the prompt's intent is unambiguous after a quick workspace scan, proceed directly to execution without surfacing any enrichment form.

Triggers

• User submits a prompt without project, file, or scope context — e.g., "let's work on this," "improve it," "can you fix the thing."
• User says "work on," "improve," "fix," "build out," or "add" with no clear object or target.
• Prompt could plausibly map to multiple workflow paths and the correct path depends on user intent.
• User opens a preset workspace with an unfilled context file and submits any non-trivial first prompt.
• * prefix → DO NOT trigger. Execute the literal text after * directly.
• Trivial prompts (greetings, "what time is it?", single-word echoes, math calculations) → DO NOT trigger.

* prefix bypass

Prompts beginning with * skip the prompt-gate evaluation entirely and execute directly. This is a Council convention preserved in Cowork. Use it when you know exactly what you want and do not need enrichment. Example: *draft the project update email runs without any gate, context check, or clarifying questions.

Instructions

Self-evaluation gate

Before Phase 1, decide whether enrichment is needed at all:

• Bypass immediately (no enrichment): prompt begins with *; prompt is a greeting or social opener ("hi", "hello", "thanks"); prompt is a simple factual question with a bounded answer ("what time is it?", "what does X mean?"); prompt includes all necessary context inline ("summarize this paragraph: ").
• Proceed to Phase 1: prompt is vague, references project work without specifying what, or could reasonably map to more than one execution path.

Phase 1 — Context check

Read context/about-me.md, writing-profile.md, and working-rules.md if present in the workspace.

For each file:

1. If the file is missing AND the task is clearly dependent on the information it would contain → emit an AskUserQuestion with chips: Fill now / Skip / Run the wizard.
2. If the file exists but contains unfilled template placeholders (e.g., [your name], [describe your work], [your role]) AND the file is relevant to the task → emit the same AskUserQuestion with chips: Fill now / Skip / Run the wizard.
3. If the file is irrelevant to the task — for example, writing-profile.md exists but the task is a math calculation or a file-search — silently skip that file. Do not surface a bootstrap offer for an irrelevant file.
4. If all relevant files are present and filled → skip Phase 1 bootstrap offer entirely. Proceed to Phase 2.

One offer per relevant missing/unfilled file. Do not stack multiple bootstrap offers in a single pass.

Phase 2 — Workspace research

Use Glob and Grep to scan the active workspace for content matching the prompt's topic:

• PROJECTS/ — project folders and their content
• TEMPLATES/ — any template files relevant to the task
• cowork-profile.md — if present at repo root
• Active workspace folder contents

Document findings internally. The goal is to narrow the prompt's intent and ground the Phase 3 questions in real workspace content. If the research fully resolves the ambiguity — you know exactly what the user wants and what file/project to work in — skip Phase 3 and proceed to Phase 4 directly.

Phase 3 — Clarifying questions (1–3)

Emit 1–3 AskUserQuestion items grounded in Phase 2 findings:

• Every chip option must come from research, not assumption.
• Never ask a question answerable from the context files already read.
• Cap at 3 questions — one AskUserQuestion with up to 3 items is preferred over 3 separate questions.
• If Phase 2 resolved the ambiguity completely, skip this phase.

Phase 4 — Execute

Proceed with the enriched understanding gathered in Phases 1–3. Do not re-surface resolved questions. Do not repeat context the user already confirmed. Execute the original intent with full fidelity to the enriched context.

Output format

This skill produces enrichment output as AskUserQuestion chips (Phases 1 and 3) or no visible output (Phases 2 and the trivial-skip path), then proceeds to the user's original ask in Phase 4. The skill does not produce a standalone artifact — its output is the enriched execution in Phase 4. A user who benefits from this skill should see better output, not a visible enrichment step.

Quality criteria

• Max 3 questions in Phase 3 — never more.
• Every chip option comes from workspace research, not from assumption or generic templates.
• * prefix is honored without exception — no enrichment runs when * is present.
• Trivial prompts are detected and bypassed without user friction.
• Phase 1 bootstrap offer fires only when a relevant file is missing or unfilled — not on every workspace open.
• The skill never modifies user files directly. Context files (context/about-me.md, writing-profile.md, working-rules.md) are read-only. Only the user can fill them (via "Fill now" chip).
• If all context is present and the prompt is unambiguous after Phase 2, the user experiences zero visible enrichment — Phase 4 runs directly.

Anti-patterns

• Firing on trivial prompts ("what time is it?", single-word echoes, greetings) — this creates friction without value.
• Asking more than 3 questions in Phase 3. Cap is hard.
• Asking a question answerable from a context file that was already read in Phase 1.
• Asking the user to retype context already provided in context/about-me.md, writing-profile.md, or working-rules.md.
• Modifying context/about-me.md or any other user file directly — always emit chips and let the user choose.
• Running enrichment on a *-prefixed prompt. The bypass is unconditional.
• Surfacing a bootstrap offer for a file that is irrelevant to the current task (e.g., offering to fill writing-profile.md when the task is a calendar lookup or math problem).
• Stacking multiple bootstrap offers in a single AskUserQuestion pass — offer one at a time if multiple files are missing.

Example

Scenario: User opens examples/research/ workspace, types "let's work on the lit review."

Phase 1 — Context check:

• context/about-me.md: present, filled (name and role filled in). No bootstrap offer.
• writing-profile.md: present, filled. No bootstrap offer.
• working-rules.md: present, filled. No bootstrap offer.
• All three fine — skip Phase 1 offer, proceed to Phase 2.

Phase 2 — Workspace research:

• Glob PROJECTS/*.md → finds three project folders: china-supply-chain/, urban-heat-study/, bioethics-review/.
• Research finding: three distinct projects, each with its own lit review context. Ambiguity not fully resolved — proceed to Phase 3.

Phase 3 — Clarifying question: Emit one AskUserQuestion: "Which project's lit review?" Chips: China supply chain / Urban heat study / Bioethics review / Other

Phase 4 — Execute: User picks "Urban heat study." Proceed with the lit review work in PROJECTS/urban-heat-study/, applying research-synthesis or literature-review skill as appropriate.

---

Trivial bypass example:

User types "what time is it?" → self-evaluation gate: trivial factual question → skip all phases → answer directly.

* bypass example:

User types "*draft the executive summary for the China supply chain project" → * prefix detected → skip all phases → execute directly.

Writing-profile integration

This skill does not produce content exceeding 100 words, so writing-profile.md voice preferences are not applied to the skill's own output (the AskUserQuestion chips and the self-evaluation logic are structural, not content). However, when the skill resolves to Phase 4 and the downstream task is a content task (writing, summary, email, brief), the standard writing-voice rule from global-instructions.md — the existing ## Writing voice block in every preset — applies as usual to that downstream output.

Example prompts

• "let's work on the project" → fires (vague; no object or scope).
• "what time is it?" → bypasses (trivial factual question; no enrichment value).
• "*draft the email" → bypasses (* prefix; unconditional bypass).
• "fix the thing" → fires (vague object reference; needs workspace research to resolve).
• "summarize this paragraph: " → bypasses (trivial; full content provided inline).

---

*Pattern from addyosmani/agent-skills skills/context-engineering/SKILL.md @ commit 9534f44c5448086fcc0046f9d83752c654c81930. Copyright (c) Addy Osmani. Licensed under the MIT License.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED. See THIRD-PARTY-NOTICES.md for the unabridged license text.*