Review prompt and tool-description changes in a branch or PR.
NewMaintained
Automatic
◆AI Engineer pack
Last verified 8/29/2026
What it changes
Passed 7 of 15 test cases with the skill, and 10 of 15 without it.
Without skill64%
With skill48%
Checks it passed that the model alone missed
Recommends reviewing runtime prompt diffs before reading test files.
Includes heavy or eval-shaped guidance as a pre-report note category.
States that unit tests and evals should only be cited as evidence of intent.
and 1 more in the full report
3 fixed9 passed either way6 still failing7 broke
Pass-rate lift
-16 pp
Tokens
+9%
Agent turns
0%
Response time
-36%
Model gemini-3.6-flashMeasured 8/29/2026v112 of 25 cases passed with skill · 15 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 3 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 3
Prompt
When analyzing prompt diffs during an audit workflow, what four specific categories should notes be organized into before drafting the final report?
What was checked, without and with the skill
Checked
Without skill
With skill
Includes behavior changes as a pre-report note category.
Passed
Passed
Includes heavy or eval-shaped guidance as a pre-report note category.
Missed
Passed
case-18 · gemini-3.6-flash · 8/29/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md86 linessingle file
Review Prompt Changes
Review prompt-related diffs with a product lens. Focus on what actually changed, whether the guidance belongs in the prompt layer where it was added, and whether any wording is repetitive or overfit to evals.
Scope
Default to main...HEAD unless the user names a different base.
Prioritize runtime files that shape model behavior:
system prompts
prompt builders and prompt fragments
tool descriptions and schema field descriptions
response-formatting instructions
Read evals and unit tests only after the runtime diffs so you can infer the intent behind the change without letting the tests define the review.
In this repo, likely files live under:
apps/web/utils/ai/**
apps/web/utils/ai/assistant/**
apps/web/__tests__/eval/**
apps/web/__tests__/** for prompt-adjacent unit tests
Workflow
Find the changed files with git diff --name-only main...HEAD.
Narrow to prompt-relevant files first. Ignore unrelated UI or plumbing unless it changes what reaches the model.
Read diffs with enough context to see the full prompt block or tool description.
Then read nearby tests or evals to understand what failure mode the change is trying to prevent.
Separate your notes into four buckets:
behavior changes
product-helpful prompt guidance
heavy or eval-shaped guidance
repeated guidance across layers
Prefer exact file and line references in the review.
What To Look For
Product-helpful changes
Capability gating that matches real product states or disabled tools
Moving provider-specific or tool-specific guidance into the relevant tool description