- Home
- Skills
- Technical Documentation
- show-receipts
Works with the AI Agents you already use
Show Receipts
show-receipts checks those claims against the evidence you give it — the repo's file list, your notes, your data — before anything goes out.
Free
show-receipts
Example session with this skill installed
(Translated from Chinese) Before this README goes out, check it for anything untrue, exaggerated, wrong or inconsistent with the repo. The repo's file list is attached: one HTML prototype, a PRD and the README itself.
- Read your context and instructions
- Compiled the show-receipts
A real run from the evaluation, translated from the original Chinese report and condensed.
Verdict: fix before publishing. 14 unsupported, 10 unverified, 2 out of scope. The repo is one HTML prototype and a PRD, but the README describes a live system with measured metrics, an npm setup and a full src/ tree.
Must fix (unsupported)
- Lines 6–8 · demo data presented as real. Badges say "success rate 87.4%", "P95 8.6s", "MVP launched". The repo has no backend, monitoring or logs. → Remove the badges, or change them to "target: ≥85% / P95 ≤10s".
- Lines 173–178 · demo data presented as real. A table labels these numbers "current value" and "✅ met", with no data source and no note that they're samples. The targets themselves (≥85%, ≤10s) are fine. → Rename the column "prototype sample value" and add "sample data, not measured".
- Line 230 · something that doesn't exist.
npm install, but there is no package.json. → "Download the HTML file and open it in a browser."
Worth checking (unverified)
- Line 18 · inflation. "The only trusted numbers in the company" is a superlative with no comparison. → "One consistent definition across the company."
Connects securely to your tools. The creator never sees your data.
About this skill
AI-drafted documents tend to overclaim in the same few ways: prototype numbers presented as "current" results, a npm install or src/ folder the repo doesn't have, "studies show…" with no source, two months of work described as "half a year".
show-receipts checks those claims against the evidence you give it — the repo's file list, your notes, your data — before anything goes out.
How it works
- Runs a bundled Python script (standard library only) that mechanically checks paths, command dependencies, license files and repeated numbers.
- Splits the document into checkable claims.
- Looks for evidence for each claim and grades it:
- Backed: evidence found, with file and line.
- Unverified: no evidence either way. External facts are always unverified when offline.
- Unsupported: your materials contradict it, or it describes something of yours that should be there and isn't.
- Suggests a toned-down rewrite for each problem ("current value" → "sample value from the prototype"), rather than just deleting it.
It covers six problem types: demo data presented as real, things that don't exist, numbers or quotes with no source, inflation, capabilities with no evidence, and internal contradictions.
Where it stops
- It checks facts against evidence. It doesn't review design, style or tone.
- It does not browse the web. Without materials, most findings will be "unverified", and the report says so.
- It does not edit your document unless you ask.
What two rounds of controlled evaluation showed
Same document, same model, same prompt, with and without the skill, scored against a frozen answer key:
- It does not make the model find more problems. Opus and Haiku already found nearly all of them without it (Opus 20/20 vs 20/20; Haiku 49/50 vs 50/50).
- It does make the report focused: about 90% of findings are actual claim-vs-evidence problems with the skill, about 70% without.
- With the skill, unverifiable external facts were never marked "must fix".
The full evaluation — design, answer keys, all 18 raw reports — is public on GitHub. The skill is MIT-licensed there too.
中文说明
AI 写的文档,夸大往往就那几种:原型里的数字被写成「当前值」,仓库里没有的 npm install、src/ 目录,说不出出处的「研究表明」,两个月的工作写成「半年」。
show-receipts 在发出去之前,拿你给的证据把这些说法核一遍:仓库的文件清单、你的底稿、你的数据。
怎么做:
- 先跑随附的 Python 脚本(只用标准库),机械地核对路径、命令依赖、许可证文件和前后重复出现的数字。
- 把文档拆成可以核对的说法。
- 逐条找证据并判档:有据(写明文件和行号)、存疑(找不到证据也找不到反证;不联网时外部事实一律算存疑)、无据(材料里有反证,或者说的是你自己的东西,本该有却没有)。
- 每个问题给一个降档的改写建议(「当前值」→「原型示例值」),不是一删了事。
它查六类问题:演示数据冒充实测、写了不存在的东西、数字和引用没有出处、往大里说、能力越界、前后对不上。
它的边界:只核对事实和证据,不点评设计、文风、语气;不联网,没有材料时多数条目会判存疑,报告里也会写明;你不说要改,它就不改你的文档。
两轮对照评测:同一份文档、同一个模型、同一句要求,装和不装各跑一次,和冻结的答案对比。它不会让模型多找出问题(Opus 20/20 对 20/20,Haiku 49/50 对 50/50);但会让报告更聚焦,装了以后约九成条目是真正的说法和证据对不上,不装约七成;装了以后,核不动的外部事实一次都没有被判成「必须改」。完整评测,包括设计、答案和 18 份原始报告,都公开在 GitHub 上,那里也以 MIT 许可开源。
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Technical Documentation.
Works offline: external facts are marked unverified, not checked on the web. Checks facts against evidence only; it does not review design, style or tone. Rules are written in Chinese; the script matches Chinese and English signal words, other languages untested. Repeated-number checks only match identical "number + unit + noun" phrases. The evaluation is small (six cases, one run per arm, a single scorer). Tested in Claude Code; not yet tested through the Agensi connector.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 22 days ago
- Free with an Agensi account
- Software
Python 3
Permissions
Terminal / Shell: runs the bundled scripts/check.py with Python 3 (standard library only). Read Files: reads the document you want checked and the materials you point it to (repo files, notes, data). It does not write files, use the network or read environment variables.