weave-prompt git:20261016.ed0740cA

Immutable. This exact content never changes and is served at /api/v1/blob/01af5c8a59ef5dba.

---
name: weave-prompt
description: Generate a ready-to-paste plain-text task brief (SuperPrompt) tailored to Claude Code, Gemini/Antigravity, or Codex, built from the task and Weave's request contract, working memory, step budget, and fidelity check. Use when handing a task to another agent or session, queuing work for later, or when the user asks for a SuperPrompt or task brief.
---

# Weave SuperPrompt

A SuperPrompt is a self-contained task brief another agent can execute
directly. It is plain Markdown text, never XML, and its wording follows the
target agent's own conventions.

## Generate one

```
weave prompt [claude|gemini|codex] [options] [task...]
```

If `weave` is not on the PATH, run `node "<plugin-root>/cli/weave.js" prompt ...`.
Over MCP, call the `prompt` tool with `task`, `target`, `mode`, `budget`,
`load_state`, `memory_name`, and `cwd`.

Examples:

```
weave prompt claude "fix the login timeout"
weave prompt codex
weave prompt gemini -b 5 "add pagination to /users"
```

- The first word picks the target when it is `claude`, `gemini` (aliases
  `antigravity`, `agy`), `codex`, or `generic`. Without it, the agent running
  the command is detected; unknown agents get the `generic` style. You can
  also use `-f`/`--for <agent>`.
- With no task, the Goal, Constraints, Not requested, Completion criteria,
  Blocking questions, Plan, and Working memory from `.weave/state.md` are used.
  Add `-s`/`--state` to include that context alongside an explicit task.
- `-m`/`--mode <off|lite|full|ultra>` sets the engineering rules (default:
  `.weave/mode`). `-b`/`--budget <1-50>` sets the step budget (default 10).
  `--memory <name>` adds a saved snapshot from `.weave/memories/`.
- Use `--` before a task that starts with `-`.

## What each target gets

Every brief has Task, Done when, Scope, verified context when available,
the existing plan and open decisions when present, engineering rules, work
instructions, fidelity questions, and final report.

- **claude**: points to `CLAUDE.md`, asks to read files before proposing
  changes, ask before editing when a requirement is ambiguous, and keep a todo
  list.
- **gemini**: points to `GEMINI.md`, asks for confirmed facts only, short tool
  output, one direct question when unclear, and places the task at the end,
  after the context.
- **codex**: points to `AGENTS.md`, asks to keep going until the task is
  resolved, search with `rg`, keep diffs minimal, run tests, and give a brief
  final answer.

## Make it useful

The CLI/MCP output is a scaffold. When an agent invokes this skill, do not
merely paste it back to the user. First inspect the task, relevant repository
entry points, and `.weave/state.md` when applicable. Then produce one
ready-to-paste brief in the target agent's style that includes:

1. The user's actual objective, preserved without adding features.
2. Specific acceptance criteria and a practical verification for each.
3. Confirmed repository facts: relevant paths, the current flow, and any
   constraints or invariants the next agent must preserve.
4. The first useful action and an ordered plan for multi-step work. Include
   the existing plan only after checking it still fits the request.
5. The one blocking question, if any, or explicit assumptions otherwise.

Use the generated scaffold for target conventions and engineering rules;
remove repeated or irrelevant boilerplate. Distinguish verified facts from
assumptions. If there is no repository context, do not invent paths or
acceptance criteria that cannot follow from the task. Ask for missing
information only if a correct brief cannot be written without it. Present
only the finished Markdown brief, without XML or commentary about generation.

## When to use this vs. the `weave` skill

For a task you are executing yourself in this session, follow the `weave`
skill directly. Use `weave-prompt` when the work goes to another agent or
session, or when the user explicitly asks for a SuperPrompt.