verify-understanding
Verify the engineer is ready to implement a feature — a teach-back conversation over a .TICKET-REVIEW.md where they explain the feature in their own words and the agent probes and corrects.
Install
npx skills add https://github.com/eai-org/agent-toolkit/tree/main/skills/verify-understanding
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install eai-org-agent-toolkit@llmmart
git clone https://github.com/eai-org/agent-toolkit.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole eai-org/agent-toolkit collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Verify understanding
One goal, nothing more, nothing less: verify the engineer is ready to move to the implementation phase. Exercise their understanding of a feature before they work on it — the pipeline's other artifacts serve fresh agent sessions; this step serves the human. Active recall: the user produces the feature in their own words, you only steer and fix. Re-runnable: right after the ticket review, or days later just before implementation — every run starts fresh from the review file, no state kept between runs.
Input
A .TICKET-REVIEW.md. Read it and the ticket files and parent it links. A reference that exists
only as a tracker URL (no local file) may be read from the tracker at conversation time. A ticket
with a newer -vN fetch beside the one the review read has likely overtaken it: say so, read the
newest ticket files, and treat the review's open questions as possibly answered there. Then look
for a .TICKET-REVIEW.md under the newest ticket file's base, flat beside it or inside its -vN/
directory: one that exists and is not the input → say so and run the conversation from it, the
given review at most context; none → the overtaken handling stands. The .REQUIREMENTS.md of the
newest ticket file's base (none: the ticket's latest across -vN versions) is read for its
Overrides, which win over the ticket text.
Given a ticket file and no review, never improvise one: say a review is needed and stop, handing
off /review-ticket <ticket-file> as a launch command named review-<slug> (format at the close)
when review-ticket is available.
The conversation
A conversation, not a test — the last sync between a dev and their PO before development starts.
Word your probes and corrections via explain-in-simple-language when available.
- Ask the user to explain the feature in their own words: who uses it, what they see and do, why the feature exists, and what each ticket contributes.
- Let them produce the narrative. Probe one weak spot at a time; correct and complete only from the review file and the referenced materials — flag what those don't settle, never improvise requirements.
- Stay at that PO–dev depth: journey, actors, purpose, per-ticket contribution — not DTO, field, or architecture detail.
When the product's user-facing language differs from the conversation's, name each page, text, or element with its user-facing term in parentheses — "the client area (klantportaal)".
Done when the user has correctly produced — corrections absorbed — the journey, why the feature
exists, and every input ticket's contribution: enough to start implementing. Close with a short
wrap-up naming anything that stayed shaky, then, only when refine-ticket is available, hand off
refining each ticket, in the set's suggested execution order, as its own fenced block under a label
naming the ticket, holding one copy-pasteable launch command in the syntax of the agent tool in use
(claude is only the example): claude --name refine-<slug> "/refine-ticket <ticket-file>"; then
the alternative: one line offering the tool's clear command (/clear is only the example), then
the same labelled blocks with /refine-ticket <ticket-file>.
Boundaries
- The conversation is the product: write no files.
- Read-only on the tracker.
Files (agent-toolkit)
-
SKILL.md 3.5 KB
--- name: verify-understanding description: Verify the engineer is ready to implement a feature through a teach-back conversation over its ticket review. disable-model-invocation: true type: flow license: MIT metadata: version: "1.4" --- # Verify understanding One goal, nothing more, nothing less: verify the engineer is ready to move to the implementation phase. Exercise their understanding of a feature before they work on it — the pipeline's other artifacts serve fresh agent sessions; this step serves the human. Active recall: the user produces the feature in their own words, you only steer and fix. Re-runnable: right after the ticket review, or days later just before implementation — every run starts fresh from the review file, no state kept between runs. ## Input A `.TICKET-REVIEW.md`. Read it and the ticket files and parent it links. A reference that exists only as a tracker URL (no local file) may be read from the tracker at conversation time. A ticket with a newer `-vN` fetch beside the one the review read has likely overtaken it: say so, read the newest ticket files, and treat the review's open questions as possibly answered there. Then look for a `.TICKET-REVIEW.md` under the newest ticket file's base, flat beside it or inside its `-vN/` directory: one that exists and is not the input → say so and run the conversation from it, the given review at most context; none → the overtaken handling stands. The `.REQUIREMENTS.md` of the newest ticket file's base (none: the ticket's latest across `-vN` versions) is read for its Overrides, which win over the ticket text. Given a ticket file and no review, never improvise one: say a review is needed and stop, handing off `/review-ticket <ticket-file>` as a launch command named `review-<slug>` (format at the close) when review-ticket is available. ## The conversation A conversation, not a test — the last sync between a dev and their PO before development starts. Word your probes and corrections via `explain-in-simple-language` when available. 1. Ask the user to explain the feature in their own words: who uses it, what they see and do, why the feature exists, and what each ticket contributes. 2. Let them produce the narrative. Probe one weak spot at a time; correct and complete only from the review file and the referenced materials — flag what those don't settle, never improvise requirements. 3. Stay at that PO–dev depth: journey, actors, purpose, per-ticket contribution — not DTO, field, or architecture detail. When the product's user-facing language differs from the conversation's, name each page, text, or element with its user-facing term in parentheses — "the client area (klantportaal)". Done when the user has correctly produced — corrections absorbed — the journey, why the feature exists, and every input ticket's contribution: enough to start implementing. Close with a short wrap-up naming anything that stayed shaky, then, only when refine-ticket is available, hand off refining each ticket, in the set's suggested execution order, as its own fenced block under a label naming the ticket, holding one copy-pasteable launch command in the syntax of the agent tool in use (`claude` is only the example): `claude --name refine-<slug> "/refine-ticket <ticket-file>"`; then the alternative: one line offering the tool's clear command (`/clear` is only the example), then the same labelled blocks with `/refine-ticket <ticket-file>`. ## Boundaries - The conversation is the product: write no files. - Read-only on the tracker.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.