reask

Your agent asked you something and you cannot answer it. You lost the context, the options are jargon, or you cannot see what actually differs between them. Type /reask.

The agent asks the same question again, written for someone who just walked in: what is being built and why this decision matters now, every name from the session explained, the pros and cons of every option (the one it recommends included), its recommendation, and one question you can answer. /reask on its own is enough. The harder the question and the less sure the agent is, the more context and evidence you get; an easy call it is sure of stays short. Asked several questions at once, you can bounce just the ones you cannot answer. Always in your language.

Before it recommends anything, the agent argues against its own first lean and checks the fact that lean rests on in your project. When the check changes its mind, or turns up a better option, it says so in one line.

reask is an Agent Skill: one file, SKILL.md, that Claude Code, Codex, Cursor, Gemini CLI and other agents that read the format can load.

Install

Any agent, with the Skills CLI:

npx skills add Nezhinskiy/reask --global

It asks which agents to install into; --agent claude-code (or --agent '*') skips the question, and leaving off --global installs into the current project.

Claude Code plugin:

/plugin marketplace add Nezhinskiy/reask
/plugin install reask@reask

A plugin's skills are namespaced, so this one answers to /reask:reask. Updates arrive with /plugin marketplace update reask.

Python, for Claude Code:

uvx reask install

or pipx run reask install. This writes ~/.claude/skills/reask/SKILL.md (under $CLAUDE_CONFIG_DIR if you set it); --project writes ./.claude/skills instead and --skills-dir PATH any other location. reask print prints the skill and reask uninstall removes it. Neither command overwrites or deletes a copy you have edited unless you pass --force.

claude.ai and Claude Desktop: download reask-skill.zip from the latest release and upload it under Settings > Capabilities > Skills.

By hand: copy skills/reask/SKILL.md into a reask/ folder in your agent's skills directory.

Example

Mid-task, the agent asks:

Should the price cache key include the locale, or should we normalise currency at read time?

You type /reask, and get:

Where we are and why this comes up. We are putting a cache in front of the pricing API so product pages stop taking 800 ms. The cache works; what is open is what each entry is keyed on, and nothing merges until that is decided. The pricing API already returns a country-specific price for 3 SKUs. Decided: Redis, 10-minute expiry. My assumption: the 40 locales we ship today are the ones to plan for.

A live example. SKU B-2210, one of those 3, costs €49 in Germany and €45 in Austria. Viewed from Berlin (de-DE) and then from Vienna (de-AT):

	Key per locale	Normalise at read
Cache entries	2	1
API calls on first views	2	1
Price shown in Vienna	€45	€49

Options. Key per locale: one entry per SKU per locale. Pros: every country sees its own price; no conversion code. Cons: 40 times the entries; 40 cold-cache API calls per SKU instead of one. Normalise at read: one entry per SKU in euros, converted when read. Pros: 40 times fewer entries; one warm-up per SKU. Cons: shows the wrong price for those 3 SKUs today; a rate-table lookup on every read.

What I recommend. Key per locale. Normalising shows Vienna €49 for B-2210 today, not someday; 40 entries per SKU is about 2 MB at our catalogue size. I would change my mind if those 3 prices turn out to be a data error.

The question. Key the cache per locale (recommended), or normalise currency at read time?

In Claude Code the question arrives with both answers as options to pick, the recommended one first. This example was written by hand to show the format; a real briefing is built from your session's facts.

Why it works

Agents ask good questions in the wrong register: they have the whole session in context and you do not. reask fixes the register, not the question. It keeps the decision yours (it never starts on its default), states each option's downside at full strength, and ties the recommendation to facts from your work, so you can overrule it with one reply.

Contributing and security

Behaviour reports with a transcript are the most useful contribution; see CONTRIBUTING.md. Report vulnerabilities privately, as described in SECURITY.md. Every release is built in CI and carries a build provenance attestation.

License

MIT