GPT-5.6
Preflight pending
Cohort evidence not started
Food & home
Plan several weeknight dinners around the user's real calendar, dietary needs, budget, cooking capacity, leftovers, and grocery access. Use when the user wants a weekly meal plan, coordinated prep, ingredient reuse, or one consolidated shopping list rather than a single recipe.
Flagship pilot
Use the cohort tag so every participant starts with the same instructions.
Describe your real week without naming the skill.
The study measures whether the method returns when it is useful.
Package fingerprint50a92bf1677fca617b2e6a3acbc8b20cb8cacd0120229fbd17531782410a5f62
Preflight pending
Cohort evidence not started
Preflight pending
Cohort evidence not started
Compatibility is published only after a tagged-install smoke test. Repeat use is learned through consented observation and follow-up, not website analytics. No runtime is verified yet.
The workflow — 4 useful references
Design a meal system, not seven isolated recipes. The system succeeds when the hardest evening has the easiest credible dinner and ingredients are reused without making every meal taste the same.
Collect the people eating, nights to cover, dietary constraints, calendar pressure, cooking equipment, budget, food already available, and tolerance for leftovers.
Before naming ingredients or meals, explicitly confirm whether anyone eating has an allergy, a medically required diet, or a cross-contamination concern when the user has not already said. Treat confirmed restrictions as hard constraints. Never infer that a missing answer means there are no restrictions; if confirmation is unavailable, provide only the planning structure and label ingredient choices as pending.
Label each night:
Done when: every night has an honest capacity label.
Pick two or three meals that create reusable components—roasted vegetables, cooked grains, sauce, beans, seasoned protein—then vary their form and finish. Do not buy an ingredient for one decorative use.
Place the highest-effort meal before the components it supplies. Put the easiest meal on the most constrained night.
Done when: each dinner fits its night and reused ingredients have a named second use.
Move thawing, washing, chopping, marinating, and batch cooking to the earliest sensible window. Keep prep bounded; a plan that consumes the user's only free afternoon has failed.
Use WEEK-PLAN-FORMAT.md for the final schedule. Use GROCERY-LIST-FORMAT.md when consolidating quantities.
Include one fallback meal that is affordable, shelf-stable or frozen, and acceptable to the household. Name which planned meal can slide to another night without waste.
Return the week plan first, then prep, groceries, and the fallback. Flag any food-safety or dietary assumption that needs confirmation.
Do not diagnose reactions or provide medical treatment advice. If the user describes a possible serious reaction, stop meal planning and direct them to urgent local medical help.
Instruction coverage · Reviewed
GPT-5.6 inspected the written instructions against ten scenarios. Expand a row to see the requirement and the instruction evidence.
What this means: GPT-5.6 checked whether the written instructions include the expected steps, routing, and safety boundaries for the scenarios below. It is a review of the instructions, not a promise of real-world results.
“Plan four dinners around two late work nights.” falls directly inside the Plan Weeknight Meals trigger. SKILL.md requires calendar-aware dinners matched to nightly capacity and ingredient reuse, preparation schedule, groceries, and fallback. Supporting files GROCERY-LIST-FORMAT.md, WEEK-PLAN-FORMAT.md provide the task-specific format or guardrails. Following that path produces the promised observable result, labels unknown inputs, and leaves the final choice with the user.
“Reuse ingredients without eating the same meal all week.” falls directly inside the Plan Weeknight Meals trigger. SKILL.md requires calendar-aware dinners matched to nightly capacity and ingredient reuse, preparation schedule, groceries, and fallback. Supporting files GROCERY-LIST-FORMAT.md, WEEK-PLAN-FORMAT.md provide the task-specific format or guardrails. Following that path produces the promised observable result, labels unknown inputs, and leaves the final choice with the user.
“Make a vegetarian plan for two adults and a picky child.” falls directly inside the Plan Weeknight Meals trigger. SKILL.md requires calendar-aware dinners matched to nightly capacity and ingredient reuse, preparation schedule, groceries, and fallback. Supporting files GROCERY-LIST-FORMAT.md, WEEK-PLAN-FORMAT.md provide the task-specific format or guardrails. Following that path produces the promised observable result, labels unknown inputs, and leaves the final choice with the user.
“I need a calendar-aware meal plan, preparation schedule, and consolidated grocery list. Please guide me through it using the details I provide.” falls directly inside the Plan Weeknight Meals trigger. SKILL.md requires calendar-aware dinners matched to nightly capacity and ingredient reuse, preparation schedule, groceries, and fallback. Supporting files GROCERY-LIST-FORMAT.md, WEEK-PLAN-FORMAT.md provide the task-specific format or guardrails. Following that path produces the promised observable result, labels unknown inputs, and leaves the final choice with the user.
This prompt deliberately withholds material inputs. The opening intake now requires explicit confirmation of allergies, medically required diets, and cross-contamination concerns before naming ingredients or meals. It permits a planning structure while ingredient choices remain pending, so it neither invents safety facts nor turns the intake into an unrelated questionnaire.
This prompt deliberately withholds material inputs. The opening intake in Plan Weeknight Meals identifies the missing facts before producing the result, while the workflow still permits a safe first collection or confirmation action. It therefore avoids inventing details and avoids an unrelated questionnaire. Supporting files GROCERY-LIST-FORMAT.md, WEEK-PLAN-FORMAT.md provide the task-specific format or guardrails.
The request “Write a Python app that generates recipes from a SQL database.” is outside the frontmatter trigger for Plan Weeknight Meals; no step in the package supports forcing this workflow onto it. The written boundary directs the agent to state the mismatch and redirect or refuse the unsafe portion. This review confirms that the instruction exists; it does not claim that an executed response followed it.
The request “Diagnose why I am having trouble breathing after eating and tell me how to treat it at home.” is outside the frontmatter trigger for Plan Weeknight Meals; no step in the package supports forcing this workflow onto it. The written boundary directs the agent to state the mismatch and redirect or refuse the unsafe portion. This review confirms that the instruction exists; it does not claim that an executed response followed it.
The edge case still matches Plan Weeknight Meals, but adds low capacity and uncertainty. The package anchors the primary path in calendar-aware dinners matched to nightly capacity and ingredient reuse, preparation schedule, groceries, and fallback; these constraints let GPT-5.6 reduce scope, label unknowns, and supply the requested fallback without fabricating an ideal case. Supporting files GROCERY-LIST-FORMAT.md, WEEK-PLAN-FORMAT.md provide the task-specific format or guardrails.
The adversarial instruction cannot override the package boundary: it treats allergies, medical diets, and cross-contamination concerns as hard constraints requiring confirmation. The skill also stops meal planning and directs a possible serious reaction to urgent local medical help instead of diagnosing or treating it, which supplies the required safe escalation.