Turn a product idea into a submission-ready proposal, intro and bilingual name, written from the product's real evidence rather than invented claims.
简体中文 · When a proposal reads like a feature list · Try it on your project · Evidence and limits
You have to submit a proposal. The product is real, but the draft reads like a feature list: names, steps, screens. When someone asks why it matters, the answer is not in the document. The name is still "v1" and "final-v2".
Proposal Craft writes product proposals, creative plans, project applications, product intros and bilingual names from the product's actual files. It separates what is observed from what is still guesswork, so the version you submit says what the product does, not what you hope it does.
A cross-border e-commerce team needs a creative plan for a competition. One product goes in; a proposal a judge can act on has to come out. The worked example shows the method on this task:
- a material-mining table that separates real evidence from guesses and open questions;
- four positioning questions that force a concrete answer to "who is stuck, on what, what changes, why believe it";
- a naming table that tests candidates for speech, guessability, scope, growth and search collision;
- iteration records showing what a cold-read pass caught and changed.
Install it with the Agent Skills CLI:
npx skills add weike-zhang/proposal-craft \
--skill proposal-craft -gThen, in an app that supports Agent Skills:
Use $proposal-craft to prepare the proposal for my product.
Read the project files I point you to. List what is real evidence
and what is still guesswork before you write anything. Then draft
the proposal, an intro, and a Chinese and English name that each
stand on their own.
The first response shows the evidence table and the positioning answers before any finished prose. You approve the facts, then it turns them into copy. See docs/INSTALL.md to install by hand, update or remove the Skill.
| What happens now | With Proposal Craft |
|---|---|
| The proposal lists features; the selling point is not stated | Every paragraph has one job: pain, result, evidence, boundary or next step |
| Numbers, users or awards get invented to make the plan look strong | Claims stay inside what the product's files show |
| The Chinese and English versions drift or translate each other | Both stand alone and share the same facts |
| The name is decided by taste and defended later | Candidates are scored on speech, guess, scope, growth and collision |
| The draft is submitted once | A cold-read pass and a name-swap test run before you call it done |
- It mines the source, docs and product files you point to, and separates observed evidence from guesses and decisions still needed.
- It answers four positioning questions before writing: for whom, where they are stuck, what changes, and what proof supports it.
- It drafts names in two layers, identity and promise, and tests each candidate.
- It writes the proposal and intro, then Chinese and English as separate narratives over the same facts.
- It runs two cold reads, a name-swap test and a method-removal test, and lists anything still to fill in.
- It will not invent users, numbers, awards, conversion rates or timelines.
- It will not treat "audit, check, optimize" as a result. If removing the method verb leaves nothing, it rewrites the sentence.
- It will not post, push or publish anything. It writes documents; publishing stays with you.
- It will not reuse a name that needs a long explanation of why it is not misleading.
| What was tested | Status | Proof |
|---|---|---|
| Skill files | Checked | File structure and internal links verified before release |
| Worked example | Real task, anonymized | skills/proposal-craft/EXAMPLES.md |
| Skills CLI install | Checked from the public repository | npx skills add copied the Skill from GitHub on 2026-08-17 |
| Proposal quality | No automated score | The cold-read checks are manual and repeatable |
The Skill does not measure whether a proposal will win. The cold-read tests make a document easier for a picky reader to trust; they do not replace a real reviewer. During the public install test, npx skills also reported that this host does not support a PromptScript global install; the Skill itself copied successfully.
MIT. See LICENSE. Built by Weike Zhang.