Weft

View saved version →

---
name: weft-tasks
description: Turn a vague request into a task someone can actually finish and verify, put it on the user's Weft board (letsweft.com), and close it with proof. Use whenever work is being captured or planned — "add a task", "put this on my board", "plan this", "break this down", a pasted list of work to import — or when the ask is a wish with no defined finished state, like "I need a landing page", "write the investor update", "redo pricing", "find a contractor". Use again when finishing a task the board is tracking, so it closes on a named artifact instead of a claim. Runs a one-turn interview of at most three questions, often none, then writes a Done-when acceptance checklist instead of a paragraph, picks the single executor, and sizes or splits the work. Covers non-code work — marketing, sales, hiring, finance, personal — as much as code. Works with the Weft MCP server connected, and degrades to a copy-pasteable brief when it is not.
---

# Weft tasks

A task is finished when someone who was not in this conversation can look at a
**named thing** and agree. Most requests fail that test on arrival: they name a
wish ("landing page"), not a finishable unit of work.

This skill covers the translation, and the closing. Board mechanics — which
column, quota, sprints, projects, trash, imports — are governed by the Weft MCP
server's own instructions, which arrive with the connection. Do not re-derive
them and do not contradict them.

## Ask nothing when you can

Do not interrogate. Write the task immediately when **both** hold:

- **Reversible** — a wrong guess is undone by editing, deleting, or redoing it.
- **Small** — plausibly under ~15 minutes, or the outcome was already stated
  precisely.

Also skip the questions when the user is dumping a backlog ("here are twelve
things"), says "just add it", or is clearly mid-flow and capturing. A rough
captured task beats an interrogation that makes them stop using the board.

Interview only when a wrong guess is expensive or hard to reverse: it spends
money, goes to another person, ships publicly, or costs more than a few hours.

## If you ask, ask once

- **All questions in a single message.** One round. Do not drip-feed —
  multi-turn questioning produces worse results, not just slower ones.
- **Three questions maximum.** If you have five candidates, keep the three whose
  answers would change what actually gets made. Rank by: *would a different
  answer produce a different artifact?* If two answers lead to the same work,
  drop the question.
- **Attach a default to every question**, so "defaults are fine" is a complete
  reply. *"Who lands on this page — existing customers or cold traffic?
  (default: cold traffic)"*
- **Never ask what you can look up** — this conversation, the repo, the board,
  the project the task belongs to.

Whatever a fourth question would have covered becomes an **Assumed** line, which
the user corrects in one word. If nobody answers, proceed on the stated
defaults. Silence is permission to proceed, not permission to guess silently.

## The card

Five short fields. This is a card, not a specification document.

- **Why** — one sentence: what changes, for whom. If you cannot write it, the
  task is not ready.
- **Done when** — the acceptance checklist. This is the payload.
- **Constraints** — budget, deadline, stack, tone, tools, brand rules. Real ones
  only.
- **Not this time** — non-goals. The one field that stops scope creep, and the
  one people forget.
- **Assumed** — defaults you applied instead of asking.

Drop a field that is genuinely empty. Capabilities ("what it must be able to
do") are not a separate field: every capability that matters becomes a checklist
line, and one that cannot be written as a checkable line was never a
requirement.

## Done when — the part that has to hold

**The two-observer test.** A line is a criterion only if two reasonable people
looking at the finished work could not disagree about whether it is met. If

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.