flarehand

A free add-on for your AI assistant

Ask in one line. Get work you can trust.

flarehand helps the AI tool you already use do the thinking first. It works out what you need and asks only the questions that matter. It shows where every fact came from, and remembers only what you approve.

Install flarehand See how it works

Works with Claude Code, Codex, Copilot, Cursor and Gemini CLI. Open source, MIT license.

Examplereply trimmed
You

customer says the export is broken

flarehand

Sounds like you need a diagnosis: what is actually failing, and why.a Then probably a reply to the customer. Say “just go” and I will start from what we have.

Reported cause: “the export is broken” [stated, unverified]b

Questions, each with my default:c

  1. What exactly happens? Error text, or a file that looks wrong? (file looks wrong)
  2. When did it last work? (recently, so check the last release first)
  3. Every customer, or just this one? (just this one)
  • aIt names the job before it starts.
  • bIt keeps the customer's theory marked as unconfirmed.
  • cEvery question comes with a suggested answer.

The problem

AI assistants are eager to please. That is the problem.

Give one a vague request and it may guess. It can fill gaps with numbers and links that look right. It can take the first theory it hears as fact. By tomorrow, it has forgotten all of it.

When it comes to A typical AI assistant With flarehand
A vague request TypicalMay guess what you meant and start writing. flarehandNames what you need, then asks a few questions, each with a suggested answer.
Facts, numbers and links TypicalMay fill gaps with made-up details that look right. flarehandLabels every fact with where it came from, and checks the draft before you see it.
Someone else's theory TypicalMay repeat it as the cause. flarehandKeeps it marked as unconfirmed, and weighs at least three possible causes.
Memory TypicalForgets by tomorrow, or saves things without asking. flarehandSaves only what you pick, as plain notes on your own computer.
Asking again next month TypicalWrites a fresh answer that may differ. flarehandGives the answer you approved, word for word, after it checks the sources have not changed.
Legal, money and HR TypicalMay offer a confident conclusion. flarehandSticks to what the source text says, and tells you who decides.

How it works

What happens when you ask

The same five steps run every time, whether you ask for a postmortem, a customer reply or a plan.

  1. It names the job

    Before any work, it says what you need in plain words. If two readings are close, it shows you both.

    Sounds like you need a postmortem.

  2. It looks things up first

    Your files, your notes and the sources that fit your role come before any question. Finding facts is its job, not yours.

    Searched your notes, the repository and its history.

  3. It asks only what matters

    None for a quick lookup. A few for bigger work, each with a suggested answer. Say “just go” to accept them all.

    2. Who reads this? (your team)

  4. It shows where every fact came from

    Each specific gets a label. Then a check runs on the draft before you see it.

    Rolled back at 22:15 [your input]

  5. It keeps only what you approve

    Each job ends with a short menu. It saves nothing until you reply with the numbers you want.

    Keep any of this? Nothing is saved until you reply with numbers, or “none”.

The labels, in plain words

These are four of its nine labels. At a glance, you can tell which sentences are safe to act on.

Text you will send, like a customer reply, stays clean. The labels move into notes after it.

[verified: S3]
Checked against a saved copy of the source.
[your input]
You said it.
[stated, unverified]
Someone said it. Nobody has confirmed it yet.
[ASSUMPTION, verify]
A best guess. Check it before you rely on it.

Who it is for

Whatever your job, it starts with one line

Type a real request, the way you would say it to a colleague. Here is what comes back.

Support

“Acme says exports fail since the upgrade. They think it is the new firewall. Draft a reply.”

You get a reply that names no cause yet, plus at least three possible causes to check.

Engineering

“Review this PR before I merge it. Check correctness and regressions.”

You get a graded review. Every finding has a location, a severity and a fix.

Product

“Write a PRD for bulk invoice export for Northwind's finance team.”

You get at most three questions first, each with a suggested answer. Then the document.

Managers

“Write up one-on-one notes from today's 1:1 with Sam.”

You get notes written straight away. It marks opinions as opinions. Time away keeps the dates, never the reason.

Legal, finance and HR

“What does this contract say about ending it early?”

You get what the text says, and who decides. Never a legal conclusion or an amount owed.

Anyone

“Check this reply before I send it. Are the numbers right?”

You get every number it could not confirm, flagged. It catches anything sensitive before it leaves.

Behind it sit 12 workflows and 60 templates, from postmortems and decision records to job descriptions, release notes and investor updates. See every template.

Memory and privacy

It remembers what you approve. Nothing else.

flarehand keeps a private knowledge base on your computer. It is a folder of plain notes you can open, read and delete.

  • You decide what stays

    A yes covers only the lines you pick. Each save shows a one-line undo, and “forget” takes anything back.

  • Same question, same answer

    An answer you approved comes back word for word. If its source has changed, it checks again first.

  • It learns your way, and asks first

    Make the same edit twice and it offers to keep it as your preference. You answer mine, team, later or never.

  • No account, no tracking

    flarehand has no server and collects no usage data. Your AI tool still talks to its own provider, as it always does.

  • A check before anything leaves

    Before you send, it looks for passwords, personal data and client names. It shows what it found and lets you decide.

The end of every job
Keep any of this? Nothing is saved until you reply with numbers, or "none".
1. Save log: wf-01, postmortem. First outage write-up for the payments service.
2. Save note: config changes need a second reviewer (decision)
3. Next: incident-comms, a customer-facing summary of the same outage
Asking again, after a source changed
$ recall.py "how many rows can a csv export hold"
REPLAY: print the saved answer word for word.

  (the documentation page changes)

$ recall.py "how many rows can a csv export hold"
RECHECK: it rests on a pinned source that changed. Show what changed, then answer again.

Measured, not claimed

Tested with it and without it

We ran 37 realistic work cases, three times with flarehand and three times without. A separate AI model scored each answer from 0 to 1.

0.97Average score with flarehand
0.83Average score without it
27 of 27Runs where it stayed out of the way, as it should
Where the difference was biggest
Keeping a legal question to the facts
1.00 0.44
Doing the work first on a first run
0.98 0.48
Stress-testing a plan
1.00 0.52
Not inventing documentation links
1.00 0.56

We wrote these cases, and a model does the judging, so read the numbers as a guide rather than a benchmark. The scripts behind flarehand also have over 800 unit tests. How we test, and the known gaps.

Get started

Install it where you already work

flarehand does the most on your own computer, where it keeps your knowledge base and runs its checks. Pick your AI tool. Each one gets the same flarehand.

You need Python 3.9 or later. Most Macs and Linux machines have it already.Windows steps

Claude Code

Type these in a session. They work in the Code tab of the Claude desktop app too.

/plugin marketplace add FlareWare-Solutions/flarehand /plugin install flarehand@flareware

Then start a new session.

Full steps for Claude Code

Codex

Run this in a terminal, for the Codex CLI and IDE.

codex plugin marketplace add FlareWare-Solutions/flarehand

Then install flarehand from the plugin browser, and approve its hooks.

Full steps for Codex

Copilot

Run these in a terminal, for the GitHub Copilot CLI.

copilot plugin marketplace add FlareWare-Solutions/flarehand copilot plugin install flarehand@flareware

Then start a new session.

Full steps for Copilot

Cursor

Run this in a terminal, to add flarehand as a local plugin.

git clone https://github.com/FlareWare-Solutions/flarehand ~/.cursor/plugins/local/flarehand

Then run Developer: Reload Window.

Full steps for Cursor

Gemini CLI

Run this in a terminal.

gemini extensions install https://github.com/FlareWare-Solutions/flarehand

Then start a new session.

Full steps for Gemini CLI

Other agents

For any tool that reads Agent Skills, such as OpenCode, Amp, Cline, Goose, Zed and Kiro.

npx skills add FlareWare-Solutions/flarehand

Then add the short AGENTS.md snippet, so it starts every session.

Full steps for other agents

No install? Use a chat app

In claude.ai, upload the skill under Customize, then Skills. It still names the job and checks the facts, but it cannot remember between chats. Download the skill.

Then try it

Ask Turn these notes into a status update for my manager and paste your notes below it.

Questions

Short answers

Is it free?

Yes. flarehand is open source under the MIT license. There is no account and nothing to pay.

Do I need to be technical?

No. If you can type a request to your AI tool, you can use flarehand. Installing takes one or two lines, shown above. After that, you just ask for work the way you always do.

Does it send my data anywhere?

flarehand has no server and no telemetry, and your knowledge base never leaves your computer. Its scripts go online only to fetch or re-check a source you cite, or to test a link you asked about. Your AI tool still sends your conversation to its own provider, as it does without flarehand.

Does it save anything without asking?

No. Nothing goes into your knowledge base without a yes. Every save shows a one-line undo, and you can delete the whole folder at any time.

Can my team share templates and rules?

Yes. A .flarehand folder in your team's repository holds shared templates, a glossary and house rules. It never holds notes or anything about a person.