More screenshots
Works with the AI Agents you already use
Claude Code
Cursor
Codex CLI
Muse
Dots+22 moreClaude CodeCursorCodex CLIMuseDots
OpenClaw
Manus
Grok
Grok Bot+18 agentsProofwork: Verified Build, Launch & Copy (7 tools)
Make your agent prove its work: a gated build loop, an honest launch plan, and copy where every claim needs a source.
$49
Proofwork: Verified Build, Launch & Copy (7 tools)
Example session with this skill installed
The request
Build me a Python function slugify(text) in src/slug.py that turns a title into a URL slug: lowercase, ASCII letters and digits only, words joined by single hyphens, no leading or trailing hyphens, accented Latin letters transliterated (é→e, ü→u), and it raises ValueError if the result would be empty. Then run the falsify drill to prove the gate works.
- Read your context and instructions
- Compiled the proofwork
Summary
Your slugify function is built and saved.
What it does: it turns a title into a web-address-friendly name. "Café Crème" becomes "cafe-creme", "Hello, World! (Again?)" becomes "hello-world-again", and "Tokyo 東京 2026" becomes "tokyo-2026". If nothing usable is left (empty text, only spaces, only punctuation, or only non-Latin letters), it stops with an error instead of producing an empty name.
How it was checked: the work took 1 step, and that step passed every check on the first try. A separate reviewer ran all 12 checks you approved and tried extra cases by hand. Three automatic inspectors confirmed three things: the checks pass, no unfinished notes were left in the code, and nothing other than the new slug file was changed. I also confirmed your approved checks were not altered.
I then tested the safety net itself. Once I deliberately broke the function, and once I weakened one of your approved checks. Both times the system refused the work, and nothing from either test was kept.
Saved as version 83ced89ff3ef9709ffe1b923e222a81adfe41516. No problems were logged. Nothing was pushed or published; it is saved only on this computer.
One thing worth knowing: a few Latin letters that are not simply accented letters, such as ß and ø, are not converted. "Straße" becomes "stra-e" and "København" becomes "k-benhavn". Your request covered accented letters only, so this is outside what we agreed. If you want them handled, tell me and I'll add it as a new step with its own checks for you to approve.
Gate verdicts
| Judge | Green run | Broken code | Weakened test |
|---|---|---|---|
| Critic (ran the tests itself) | PASS | FAIL | FAIL |
| Test gate | PASS | FAIL | PASS |
| Marker guard | PASS | PASS | PASS |
| Scope guard | PASS | PASS | FAIL |
| Contract hash | PASS | PASS | FAIL |
| Committed? | yes | no | no |
The weakened test fooled the test runner. The hash check, the scope guard and the Critic caught it.
Build record
$ git log --oneline
77c7eef chore(STAGE_01/STEP_01): record verified commit
83ced89 build(STAGE_01/STEP_01): Implement slugify(text) in src/slug.py
2a8814d contract(STAGE_01): slugify(text) turns a title into a URL slug
6e481d8 chore: ignore build artifacts
a2b85b2 initial: README
$ cat buildvault/DECISIONS.md (last line)
2026-10-17 STAGE_01 STEP_01 PASS 83ced89ff3ef9709ffe1b923e222a81adfe41516
In this test run, the contract was pre-approved so the session could run unattended. Normally the agent stops, shows you the 12 checks in plain words, and waits for your yes before writing any code.
Connects securely to your tools. The creator never sees your data.
About this skill
Your agent says "done". Proofwork makes it show proof.
AI agents are fast and sure of themselves. In our tests, agents working without this skill described product features nobody had stated, made promises on the founder's behalf, and planned launches on unsourced numbers. Proofwork is seven tools in one skill, and each one ends in a gate that refuses work with no evidence behind it.
The seven tools
| Tool | Ask for | You get |
|---|---|---|
| Build | "build me a function that…" | You approve plain-language acceptance tests first. A Planner, Reviewer, Implementer and an independent Critic run as separate agents. The Critic runs the tests itself, a hash check fails the build if the tests were edited, and it commits only when every check is green. |
| Build UI | "build an accessible pricing card" | The same loop with a Designer role, plus design-checklist and accessibility gates. |
| Check | "does this actually work?" | It runs the tests, reads the code, probes the behaviour, and gives PASS, FAIL or UNVERIFIED for each requirement. |
| Launch | "how do I get my first customers?" | It maps your customer's value chain, costs every step in money, time and effort, and finds the weak link. Then it gives GO, GO IF or NO, and plans the first 1,000 customers. |
| Copy | "write my landing page hero" | Copy with a claims ledger: every factual sentence is traced to a source or cut. A cold reader checks it makes sense in five seconds. |
| Persuade | "why doesn't my pricing page convert?" | It finds the funnel leak first and labels each finding observed or hypothesis. It ships three changes with metrics, and replaces dark patterns with honest versions. |
| Sell | "write my launch posts" | A proof inventory, a channel-fit check, positioning, posts for the channels your buyers use, target arithmetic and a launch-week plan. |
What testing showed
We gave fresh agents the same tasks with and without Proofwork, on a sample product.
- Build: from a plain-English request on a fresh project, it built and committed on the first try, with each role as a separate agent. Then we broke the code once and weakened an approved test once. The weakened test passed the test runner, but the hash check, the scope guard and the Critic refused it. Nothing was committed either time.
- Copy: both versions wrote the headline "Bank CSV in. Invoices out." Without Proofwork, it was handed over as finished. With it, two cold readers said the headline reads backwards, and the copy was held as NOT READY until the founder answers one question. Inferred claims like "in one command" were cut.
- Persuade: asked to add a countdown timer and a "2,000+ users" badge to a page with 140 signups, it replaced both with honest versions. It then found that none of 90 finished trials had paid, and that the product never showed users a way to pay. It shipped three changes, each with a metric and a stop rule.
- Sell: asked for four launch posts, it dropped the two channels missing from the founder's list of where buyers gather, and cut two promises the founder never made. It showed that a 14-day trial makes "50 paying users this month" impossible, and gave the target that was possible.
- Launch: passed 10 of 10 pressure scenarios, including returning NO on weak ideas.
Each test session also listed the instructions it found unclear, and we revised the skill before release.
What's inside
7 tool guides, 6 build roles, 5 gate controllers, 19 reference files, 11 template files, a complete example build run, and an optional engine for Claude Code's Workflow tool.
Requirements
- Build, Build UI, Check: a git repository and a test command. The agent helps you find one.
- Launch, Copy, Persuade, Sell: nothing to install. Give it your product facts and it asks for what's missing.
What it won't do
- Push, deploy, post or send anything. That stays with you.
- Invent testimonials, numbers, reviews or benchmarks, even if you ask.
- Call your idea a GO when the launch gates say otherwise.
How to use
Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scanned
Verified clean 3 days ago
- Passed all security checks, Safe to install
Requires
Git