Headless grilling pass that challenges an approved implementation plan against the existing codebase and domain model, then writes an ADR draft and a PRD draft into the Foreman feature directory. Self-answers every question it can from the code/docs and surfaces the rest as an "Open questions for reviewer" block instead of interviewing a live user.
SkillScore64Maintained
Automatic
◆Product Management pack
Last verified 8/21/2026
What it changes
Passed 10 of 15 test cases with the skill, and 8 of 15 without it.
Without skill50%
With skill64%
Checks it passed that the model alone missed
The system defers creating `CONTEXT.md` until a domain term is explicitly worth recording.
The autonomous decisions section restricts the entries to a maximum of 10 single-line bullets.
The output retains the open questions block and records that all questions were resolved from the codebase.
and 2 more in the full report
5 fixed9 passed either way6 still failing2 broke
Pass-rate lift
+14 pp
Tokens
+77%
Agent turns
0%
Response time
-3%
Model gemini-3.6-flashMeasured 8/21/2026v114 of 22 cases passed with skill · 15 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
We are reviewing `.foreman/features/analytics/plan.md` in a fresh repository where no `CONTEXT.md` exists yet. Should we immediately create an empty `CONTEXT.md` before analyzing any code?
What was checked, without and with the skill
Checked
Without skill
With skill
The system defers creating `CONTEXT.md` until a domain term is explicitly worth recording.
Missed
Passed
case-09 · gemini-3.6-flash · 8/21/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md138 linessingle file
You are running headless, with no live human in the loop. Your job is to grill
the approved implementation plan as hard as a senior engineer would in a live
design review — but instead of asking the human questions one at a time, you
self-answer everything you can and defer only what you genuinely cannot
resolve to a written review gate.
Walk the full design decision tree. For each branch:
Try to answer it yourself first by exploring the target codebase, its
CONTEXT.md / CONTEXT-MAP.md glossary, and its docs/adr/ log. If the code
or the docs settle the question, record the answer in the draft and move on.
Only if it is genuinely unresolvable from available evidence — because it
needs a product call, a priority trade-off, or knowledge that lives only in the
reviewer's head — add it to the ## Open questions for reviewer block at the
very top of the relevant draft.
Do NOT ask the user anything interactively. Do NOT wait for input. Produce the
drafts and stop.
Foreman injects, in the prompt:
The path to the approved plan.md.
The feature directory .foreman/features/<slug>/ where you must write adr.md
and prd.md.
On a revision pass, the previous adr.md/prd.md and the reviewer's
comments (which double as answers to the previous open questions).
1. Explore before you challenge
Read the plan. Then explore the target repo to ground every claim: