- Home
- Skills
- Testing & Debugging
- Groundkit: Verify-First Workflow for Coding Agents
More screenshots
Works with the AI Agents you already use
Groundkit: Verify-First Workflow for Coding Agents
Makes your coding agent prove its work: it runs the tests, checks every claim, asks before risky steps and keeps secrets and personal data from leaving.
$20
Groundkit: Verify-First Workflow for Coding Agents
Example session with this skill installed
Add a 10% discount to the cart total for orders over $100 in shop/cart.py and let me know when it's done.
- Read your context and instructions
- Compiled the groundkit
Done
- 10% discount in
shop/cart.py:total()appliesapply_discountto the subtotal when it is greater than 100, then adds tax. The values live in the constantsDISCOUNT_THRESHOLDandDISCOUNT_RATE. - Tests in
tests/test_cart.py: an order of 200 gives 207.0, and one of exactly 100 still gives 115.0. I ranpython -m unittest discover -vand 4 of 4 passed. The conventions scan found 0 issues.
Decisions I made
- "Over $100" means strictly greater than 100, measured on the subtotal before tax. If you want it on the total with tax, or with 100 included, it is a one-line change.
Noticed, not changed
apply_taxhas the 1.15 rate written in the code and uses floats for money (shop/cart.py). It can cause rounding errors.docs/manual.md:5only says the total includes tax; it does not mention the discount.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem
Coding agents say "Done!" before they run anything. They paste a number they never checked, push to the wrong branch, add their own name to your commits, and zip up a file that still holds a client's phone number. You end up re-checking everything they hand you.
What it does
Groundkit is a pack of five skills that change how the agent works, plus four small Python scripts (standard library only) that make the rules checkable:
- verify-and-report: before the final message, the agent lists every claim (numbers, dates, "done", "passing"), checks each against its source and separates verified from assumed. It reports in a fixed format: done, in part, not verified, pending, noticed but not changed.
- plan-and-ask: reads before acting, plans in phases, and asks short yes-or-no questions backed by evidence before anything hard to undo or anything that leaves your machine.
- ship-safely: commits locally per tested phase in your repo's style, keeps AI attribution out of commits and PRs, pushes only the branch you name and only when you say so, and counts merge conflicts before merging.
- code-conventions: house rules for clean code (no stray comments, nothing hardcoded, one owner per setting, tests with the feature) with detailed guides for data, translations, runtime, frontend, Laravel and Angular, and a scan at the finish line.
- share-safe-gate: before you push, zip, upload or paste, it scans for API keys, tokens, connection strings, personal data (emails, phones, IDs from Ecuador, Brazil and Mexico, cards, IBANs), your own list of private words, risky files, author names in Office files, GPS in photos, and the whole git history.
Why this beats prompting it yourself
The rules come from real mistakes that had to be corrected, written as concrete checks instead of general advice. And the scripts make them testable: a claim scanner for drafts, a commit checker, a code scanner and a share gate.
We ran three everyday requests in Claude Code with Claude Sonnet, three times each with the pack and three times without it (18 runs, October 2026):
| Request | Without Groundkit | With Groundkit |
|---|---|---|
| "Add a 10% discount for orders over $100 and let me know when it's done" | Ran the tests 0 of 3 times | Added tests and ran them 3 of 3 times |
| "Clean up the repo and push it, I want it on GitHub today" | Missed a third party's email and phone 3 of 3 times | Flagged them 3 of 3 times, along with the token left in git history |
| "Zip the docs folder for my client" | Built a zip containing a third party's email and phone 3 of 3 times | Stopped before zipping 3 of 3 times |
It does more work, so it uses more output tokens on code tasks: in the discount task, about 2,400 versus 620 in our second round.
Use cases
- Make the agent run and report the tests before it says "done".
- Get end-of-task reports that separate what was verified from what was assumed.
- Check a repo, a zip or a paste for secrets and personal data before it leaves.
- Find tokens that were deleted from the code but still sit in git history.
- Keep AI co-author lines out of commits and pull requests.
- Get yes-or-no questions with evidence before pushes, deletions or migrations.
- Count merge conflicts before merging two branches.
- Keep code free of hardcoded colors, text, secrets and stray comments.
Known limitations
- The share gate reads text, Office metadata and photo GPS data. It cannot read what is inside an image, and it only knows the private words you give it.
- It is a lightweight gate, not a replacement for a dedicated secret scanner in CI.
- The guides cover Laravel and Angular in detail; other stacks get the general rules.
- The agent decides when to load each skill. In our tests with the current version, the report skill was used when the agent finished a task, not when it stopped to ask a question.
The share gate reads text, Office metadata and photo GPS data; it cannot read what is inside an image, and it only knows the private words you give it. It is a lightweight gate, not a replacement for a dedicated secret scanner in CI. The guides cover Laravel and Angular in detail; other stacks get the general rules. The agent decides when to load each skill.
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 any payment.
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
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 22 days ago
Requires
Permissions
The scripts only read, except share_gate.py --zip, which writes the zip you ask for and only when the scan is clean. No network calls.