- Home
- Skills
- Prompt & Skill Engineering
- Evidence-Gated Dev: Prove It Before You Ship
Works with the AI Agents you already use
Evidence-Gated Dev: Prove It Before You Ship
A workflow for coding agents: write numbered requirements, plan risk-based tests, lock them, then prove every requirement with tagged test evidence. Includes a guard script that catches skipped or edited tests.
Free
Evidence-Gated Dev: Prove It Before You Ship
Example session with this skill installed
I need to add a retry mechanism to the payment gateway. It must retry 3 times on 5xx errors but fail immediately on 4xx. Use a contract-first approach with the Verify mode.
- Read your context and instructions
- Compiled the evidence-gated dev
I have initialized the Evidence-Gated Development workflow in
Verify mode. I have analyzed the payment gateway requirements and drafted the initial contract and test plan for your approval.
Requirements Contract & Test Plan
Contract (R-IDs)
| ID | Requirement (Quoted) | Kind |
|---|---|---|
| R-1 | "It must retry 3 times on 5xx errors" | Wanted Behavior |
| R-2 | "fail immediately on 4xx" | Must Never Happen |
Test Plan (C-IDs)
| ID | Risk Group | Requirement | Description | Oracle |
|---|---|---|---|---|
| C-1 | Happy | R-2 | Send 400 Bad Request | "fail immediately on 4xx" - expect 1 call total |
| C-2 | Dependency | R-1 | Mock 503 error for 2 calls, then 200 | "retry 3 times" - expect success on 3rd attempt |
| C-3 | Boundary | R-1 | Mock 500 error for 5 consecutive calls | "retry 3 times" - expect failure after 4 total attempts |
Scope & Protections
- Allowed Path:
src/gateways/payment.js - Protected Paths:
package.json,vitest.config.ts,tests/setup.ts
Next steps
- Review the requirements and test cases above.
- Reply with "Approved" to lock these tests and begin implementation.
- Provide the existing
payment.jsfile if not already in context.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
Evidence-Gated Dev makes your coding agent work like a QA-minded team. It writes down what "done" means, plans the tests, locks them, implements, and then proves the result. Anything it could not prove is reported as a warning, never as success.
Built from the workflow of the open-source AI Department harness and rewritten as plain instructions. No app is required.
The workflow
- Pick a mode. Single is the default; Clarify, Verify or Compare only when a risk signal justifies the extra cost.
- Write numbered requirements (R-1, R-2...) quoted from your own words.
- Plan test cases across 7 risk groups: happy path, bad data and boundaries, authorization, repeats and concurrency, dependency failures, state transitions, data consistency. You approve the plan.
- Write tests tagged [R-x][C-x], then lock them so the agent cannot weaken them to get green.
- Implement inside an agreed scope. Lockfiles, configs and CI files stay untouched.
- Classify every result as pass, fail, skipped, flaky or incomplete. Missing evidence never counts as verified.
- Run an independent review pass. Real bugs become regression tests.
- Get an acceptance report listing proof, blockers and warnings.
Guard script included
scripts/evidence_guard.py (Python 3, standard library only) checks that:
- every required requirement has a tagged test
- no test is skipped, disabled or marked .only
- approved test files were not modified, deleted or moved
It makes shortcuts visible. It does not prove your tests are good, and an agent can still skip running it.
Good for
Payments and billing, state machines, retries and idempotency, permissions, shared APIs, bugfixes you must prove, refactors of code with weak tests.
Not worth it for typos, renames or one-line edits. The skill tells the agent to skip the ceremony for those.
What's inside
- SKILL.md with the 8-step workflow and hard rules
- Reference guides: contracts and test cases, verification and findings, modes, scope and budget
- Templates: task contract and acceptance report
- The guard script
Safety
No network calls and no package installs. The script reads files only under the current folder and writes one file: the lock file you name. Read the source before you run it.
Limits
It guides the agent; it cannot force it. Quality depends on the model following the steps. Examples use JS/TS and Vitest, with naming rules for Java, Python and Go.
Optional companion: the open-source AI Department app (MIT) enforces this workflow in code. See references/ai-department-app.md.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Prompt & Skill Engineering.
Requires Python 3 for guard scripts. Not for rapid prototyping. Manual intervention needed for system config changes. Final acceptance report requires user review of warnings.
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 5 days ago
- Free with an Agensi account
- Software
Git