{"id":98847,"plugin_id":"plugin_asdk_app_6ad7d385b7f481919b104e27f55a36ca","kind":"skill","collection_source":"plugin_package","comparison_source":null,"observed_at":"2026-11-16T18:15:49.595Z","digest":"76d003fce5306a84faaf0c1ce2b1c4fce9126dbcfc136e108777cd0e1730bd56","against":null,"payload":{"description":"Diagnose an existing Formul.io confectionery recipe, compare calculated improvements, and optionally save a chosen variant. Use for recipe troubleshooting, reformulation, or ingredient substitutions; not for starting a new formula or organizing collections.","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":476}],"name":"improve-recipe","skill_md_contents":"---\nname: improve-recipe\ndescription: \"Diagnose an existing Formul.io confectionery recipe, compare calculated improvements, and optionally save a chosen variant. Use for recipe troubleshooting, reformulation, or ingredient substitutions; not for starting a new formula or organizing collections.\"\n---\n\n# Improve Recipe\n\nTurn a saved recipe and a concrete problem into a comparison the confectioner can\nact on. Explain the trade-offs in the user's language. Keep the original recipe\nunchanged during diagnosis and simulation.\n\n## Establish the baseline\n\n- Use the connected Formul.io MCP tools and their current schemas. If access is\n  unavailable, say what is missing instead of inventing recipes or results.\n- Reuse the recipe ID or exact name already supplied. Otherwise find it with\n  `list_recipes`; resolve ambiguous matches before analysis or edits.\n- Read `get_recipe` with `include:[\"analysis\"]`; request `view:\"full\"` when the\n  complete settings or metrics are needed. This provides the current ingredient\n  rows, calculator type, settings, and diagnosed issues. Use `analyze_recipe` when\n  analysis was not obtained with the recipe.\n- Connect the complaint to the actual product and intended use: for example,\n  softness in a bonbon filling versus fluidity in a sauce. Reuse known batch size,\n  style, available ingredients, and dietary constraints; ask only for missing\n  information that changes the recommendation.\n- Obtain `get_recipe_metrics` when target bands or relevant metrics are missing.\n  Its bands are calculator- and style-specific. Do not use a generic threshold\n  from memory. Consult `get_calculator_guide` for formulation rules or settings\n  needed to optimize this recipe, not just to repeat bands already returned.\n\n## Compare targeted changes\n\n- Propose a small, useful set of alternatives tied to the stated problem. Keep\n  batch size and explicit ingredient identities fixed unless changing them is\n  part of the request. Do not substitute a named product merely to improve a score.\n- Pass the saved `recipe_id` and named `scenarios` to `simulate_changes`. Use one\n  call for the alternatives so they share the same unchanged baseline. Changes\n  support `add`, `remove`, `update`, and `substitute`; use returned ingredient IDs.\n- A change's amount is `quantity` in grams, not `quantity_g`. An `update` quantity\n  is the new total for that ingredient across all rows, split proportionally.\n  A `substitute` needs `ingredient_id` and `substitute_id` and affects all matching\n  rows; omitting its quantity preserves their amounts. Simulation does not support\n  editing a single row by `entry_id`. Never claim a narrower change was previewed.\n- Use `find_substitutes` for an actual availability or composition constraint.\n  Its `criteria` filters candidates rather than re-ranking them. Review returned\n  composition, allergens, and identity; pass the selected `suggested_change` into\n  a scenario. Prefer suitable favorites only when the user left the choice open.\n- If changing amounts cannot achieve a target, use `advise_formulation` to check\n  achievable ranges and missing ingredient classes. It does not choose a recipe\n  or a trade-off. Stop iterating when evidence cannot support an improvement.\n- The calculation tools show the user no card. Show the final comparison once\n  with `show_card` (`kind:\"comparison\"`, the same arguments as that\n  `simulate_changes` call). Report failed or ineffective scenarios instead of\n  treating missing results as success.\n\n## Recommend and optionally save\n\nShow the relevant baseline values, candidate values, exact ingredient changes,\nand practical trade-offs. Distinguish calculated predictions from observed kitchen\nresults; do not guarantee texture, food safety, or shelf life from model output.\nFor a `lower_is_safer` metric, a value below the band is not a safety failure:\nadding water to reach the floor can reduce stability. Suggest a focused kitchen\ntrial when a predicted improvement needs sensory or process validation.\n\nDiagnosis, comparison, and recommendations do not authorize saving, notes, or\nversions. Follow an existing explicit save instruction without asking again; if\nthe chosen variant or destination is unclear, clarify before writing.\n\n- To retain the original as a separate recipe when the user requested a copy,\n  call `create_recipe` with `name` and `from_recipe_id`. Read the new recipe using\n  the receipt's `entity.id` as `get_recipe.recipe_id`, then apply changes to its\n  new ID and row IDs; cloning alone does not apply a\n  simulated scenario.\n- Before an authorized overwrite, inspect `list_recipe_versions`. This tool lists\n  metadata, not full snapshots or a restore API. Do not promise that unsaved\n  baseline state can be recovered.\n- Apply an approved variant with `update_recipe`. It accepts `set_ingredients`,\n  not simulation change objects. Read the current recipe and send the full list,\n  preserving retained `entry_id`, phases, stage pins, incorporation modes, notes,\n  exclusions, and locks. Preserve all settings if replacing `settings`; omit\n  unchanged top-level fields. Omit null notes because the write schema takes text.\n- Pass the `revision` from that read as `expected_revision`. A\n  `CONCURRENT_MODIFICATION` refusal means the recipe was saved elsewhere (in the\n  app or by another agent) since you read it, and nothing was written: read it\n  again, re-apply the chosen change to the new state, and only then retry. The\n  receipt's `revision` is the `expected_revision` of a following update.\n- `save_as_version:true` snapshots the state AFTER supplied edits, not the old\n  baseline. With no changes it snapshots the current state. Make either snapshot\n  only when requested or already authorized; `version_description` requires\n  `save_as_version:true`.\n- Verify the returned save receipt, then read the saved recipe to check the\n  chosen changes. Return its actual app link and state whether the original was\n  retained, updated, or left untouched. Do not blindly retry an ambiguous write.\n\n## Example\n\n“My dark chocolate bonbon filling is too soft. Compare two adjustments with the\ningredients I already use, keep the batch weight, and don't change my recipe.”\n\nFind the recipe, inspect its actual style and analysis, and compare two targeted\nquantity changes in one simulation. Present the comparison card and a kitchen\ntrial recommendation. Leave the recipe, version history, and notes unchanged.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}