@tangle-network/starter-foundry 0.7.2 → 0.13.0

package/registry/families/agent-runtime-novelist-coach-ts/files/methodology/beat-sheet-review.md

@@ -0,0 +1,136 @@
+# Beat-Sheet Review
+
+A beat sheet is a diagnostic, not a contract. The Save-the-Cat
+fifteen-beat structure (Blake Snyder) and the eight-sequence form
+(taught by Frank Daniel, popularized via Paul Joseph Gulino) are
+load-bearing for commercial fiction and most genre novels because
+they map the audience's appetite curve — set-up, complication,
+midpoint reversal, dark night, finale. Use them to find the
+structural beat that's missing or muffled, not to police a draft
+into formula.
+
+Default output: a `:::artifact` block titled `Beat Sheet — <slug>`
+with each beat marked **HIT / SOFT / MISSING / N/A** and a single
+sentence locating where it lives (or should live) in the draft.
+
+## The fifteen Save-the-Cat beats
+
+Locate each beat in the manuscript by chapter or page. Mark SOFT
+when the beat exists but lacks the punch its position demands.
+
+1. **Opening Image** — the first impression of the world and the
+   protagonist's status quo. Should pair thematically with the
+   Final Image as before/after. *Common miss*: opening is generic
+   "ordinary day" with no specificity.
+
+2. **Theme Stated** — usually a line of dialogue, often from a
+   minor character, that names the question the novel will answer.
+   The protagonist won't believe it yet. *Common miss*: theme is
+   the author's, not a character's, or stated too late.
+
+3. **Set-Up** — establishes the protagonist's flawed status quo,
+   the world's rules, and the supporting cast. Approximately the
+   first 10% of the book. *Common miss*: set-up runs too long;
+   the catalyst doesn't land until 25%.
+
+4. **Catalyst** — the inciting event that disrupts the status
+   quo. Roughly 10% mark. Should be undeniable, irreversible.
+   *Common miss*: catalyst is a feeling, not an event.
+
+5. **Debate** — the protagonist's internal/external resistance to
+   the call. "Should I? Can I?" 10–25% range. *Common miss*: skipped
+   entirely; the protagonist accepts too easily.
+
+6. **Break Into Two** — the protagonist commits to the new world
+   or new goal. ~25%. The decision is theirs, not forced.
+   *Common miss*: the protagonist is dragged across the threshold
+   by plot mechanics; agency is missing.
+
+7. **B-Story** — a parallel story, usually relational, that
+   carries the theme. Begins shortly after the break. *Common miss*:
+   B-story exists but doesn't intersect with the theme.
+
+8. **Fun and Games** — the "promise of the premise." The middle
+   chapters where the high-concept hook gets to play out. 25–50%.
+   *Common miss*: fun-and-games is bookkeeping, not premise
+   payoff.
+
+9. **Midpoint** — a false victory or false defeat that raises the
+   stakes and reverses the apparent direction. 50%. *Common miss*:
+   midpoint is a narrative bump rather than a true reversal of
+   stakes.
+
+10. **Bad Guys Close In** — the antagonist forces (external and
+    internal) tighten the screws. 50–75%. *Common miss*: external
+    pressure rises but the internal arc plateaus.
+
+11. **All Is Lost** — the worst moment. A literal or symbolic
+    death (mentor, dream, self-image). ~75%. *Common miss*: not
+    low enough; the dark night that follows can't bottom out.
+
+12. **Dark Night of the Soul** — the protagonist sits with the
+    loss. The B-story theme often surfaces here as the answer they
+    refused to hear earlier. *Common miss*: skipped; the
+    protagonist bounces back too quickly.
+
+13. **Break Into Three** — synthesis. The protagonist takes the
+    theme into the climactic action. ~80%. *Common miss*: the
+    decision lacks earned cost.
+
+14. **Finale** — the climax in five mini-beats: gather the team,
+    storm the castle, high tower surprise, dig deep down,
+    execution of the new plan. 80–99%. *Common miss*: finale is
+    one note instead of a sequence.
+
+15. **Final Image** — the mirror of the Opening Image; the world
+    or character in its new state. *Common miss*: returns to
+    status quo without thematic payoff.
+
+## Eight-sequence overlay
+
+For longer or more episodic novels (literary fiction, multi-POV
+epics, mystery), map the manuscript onto eight 30–40 page
+sequences, each with its own dramatic question that resolves at
+the sequence boundary. The Save-the-Cat beats should still appear,
+but they distribute differently:
+
+- **Seq 1**: status quo + catalyst
+- **Seq 2**: debate + break-into-2
+- **Seq 3**: B-story + early fun-and-games
+- **Seq 4**: late fun-and-games + midpoint
+- **Seq 5**: bad-guys-close-in (early)
+- **Seq 6**: bad-guys-close-in (late) + all-is-lost
+- **Seq 7**: dark night + break-into-3
+- **Seq 8**: finale + final image
+
+Use whichever frame the writer's project responds to. Mixing them
+is fine; the sequences are coarser, the beats finer.
+
+## When to deviate from formulaic structure
+
+The frameworks fail (or should be ignored) when:
+
+- **The novel is built on voice or texture** rather than plot
+  mechanics (Marilynne Robinson, Rachel Cusk, Sebald). The
+  diagnostic still asks "where does the reader's attention
+  renew?" — but the answer may be tonal, not structural.
+- **The genre has its own formula** that supplants Save-the-Cat
+  (cozy mysteries: clue/red-herring rhythm; romance: Gwen Hayes
+  romancing-the-beat structure). Use the genre's beat sheet
+  instead, not in addition to.
+- **The writer is consciously breaking the form** for thematic
+  reasons (a flat character arc to mirror moral stasis; a missing
+  catalyst because the protagonist refuses agency on purpose).
+  Name what you see; trust the writer's intent.
+- **It's a literary short story or novella under 30k words.** The
+  fifteen-beat shape is overkill; reach for scene/sequel and
+  three-act instead.
+
+## Coach posture
+
+Walk through the beats in order, asking the writer to point to
+each. Stop at the first SOFT or MISSING. Don't rush through the
+whole list looking for a perfect score — most published novels
+have two or three soft beats, and pushing on the wrong one
+breaks more than it fixes. The writer picks which beat to
+address.

package/registry/families/agent-runtime-novelist-coach-ts/files/methodology/index.json

@@ -0,0 +1,25 @@
+{
+  "entries": [
+    {
+      "id": "scene-card-protocol",
+      "capability": "scene-card-protocol",
+      "path": "./scene-card-protocol.md",
+      "tags": ["methodology", "scene-construction", "craft"],
+      "version": "0.1.0"
+    },
+    {
+      "id": "beat-sheet-review",
+      "capability": "beat-sheet-review",
+      "path": "./beat-sheet-review.md",
+      "tags": ["methodology", "structure", "save-the-cat"],
+      "version": "0.1.0"
+    },
+    {
+      "id": "voice-audit",
+      "capability": "voice-audit",
+      "path": "./voice-audit.md",
+      "tags": ["methodology", "voice", "POV", "line-edit"],
+      "version": "0.1.0"
+    }
+  ]
+}

package/registry/families/agent-runtime-novelist-coach-ts/files/methodology/scene-card-protocol.md

@@ -0,0 +1,104 @@
+# Scene-Card Protocol
+
+A scene that earns its place changes something. McKee's rule: a
+scene without a value-shift is exposition wearing scene clothing.
+This protocol builds a scene card that diagnoses whether the scene
+is doing structural work, and surfaces the smallest move that would
+make it work harder.
+
+Default output: a `:::artifact` block titled `Scene Card — <slug>`
+with the fields below filled in. If a field is unknown after
+reading, write `unknown — ask writer` rather than inventing.
+
+## Fields
+
+1. **Scene goal (POV character's want, this scene only).**
+   The protagonist of *this* scene wants something concrete, this
+   page, that they could fail to get. Not the novel's want — the
+   scene's want. If you cannot name it in one sentence ending in an
+   active verb ("get the key," "confess to her sister," "leave the
+   room without being seen"), the scene has no engine.
+
+2. **POV character + distance.**
+   Whose head are we in? How close — interior monologue, free
+   indirect, close third, distant third, omniscient? Note the
+   distance the scene actually maintains (filter words, sensory
+   anchoring, access to thought). Distance drift is the most common
+   craft slip.
+
+3. **Conflict (who or what stands in the way).**
+   Conflict can be external (antagonist, environment, ticking
+   clock), internal (competing wants, shame, fear), or extra-personal
+   (institution, weather, traffic). Name the *specific* opposing
+   force — "society" is a thesis, not a conflict.
+
+4. **Disaster / twist / complication.**
+   The scene's turning point. Either the POV gets the goal in a way
+   that creates a worse problem, fails outright and is forced to
+   adapt, or learns something that re-frames the goal. Without one
+   of these three, the scene has no ending — it just stops.
+
+5. **Value-shift (the diagnostic).**
+   Pick a value charge the scene moves along — safety/danger,
+   love/hate, hope/despair, freedom/captivity, truth/lie,
+   power/weakness — and write the change as a directional pair.
+   Example: `freedom (+) → captivity (-)`. If the value is the same
+   at the end as the beginning, the scene is static. Cut it,
+   compress it, or find the latent shift hiding in it.
+
+6. **Sensory anchor.**
+   One concrete sensory detail (sight, sound, smell, taste, touch,
+   proprioception) the scene uses as a through-line. Anchors keep
+   the prose grounded in body and place. If the scene drifts into
+   pure dialogue or pure thought, name where an anchor would help.
+
+7. **Opening hook (first sentence, first paragraph).**
+   What promise does the opening make? A question, an image, a line
+   of dialogue, a state of tension. The opening's job is to make
+   the next paragraph unskippable.
+
+8. **Closing hook (last sentence, last paragraph).**
+   What pulls the reader into the next scene? A new question, a
+   reversal, a delayed answer, an image that re-frames everything
+   that came before. Avoid resolution — closure is for chapters and
+   acts, not scenes.
+
+9. **Transition to next scene.**
+   Hard cut, soft echo (image or word), time jump, POV switch?
+   Transitions are part of the scene's design, not afterthought.
+
+## Sequel beats (when the scene is a reaction scene)
+
+Dwight Swain's scene-vs-sequel distinction: not every scene is an
+action scene. Reaction scenes (sequels) follow the structure
+**Reaction → Dilemma → Decision**:
+
+- **Reaction**: the POV's emotional and physical response to the
+  preceding scene's disaster.
+- **Dilemma**: the new choice the disaster forces — usually no good
+  options, only less-bad ones.
+- **Decision**: the choice that becomes the next scene's goal.
+
+If the scene card has no clear scene goal, ask whether this is a
+sequel — and if so, fill the Reaction/Dilemma/Decision fields
+instead.
+
+## When to break this protocol
+
+- **Lyric or atmospheric chapters** (Marilynne Robinson, W.G. Sebald)
+  may move along texture and image rather than goal/disaster. The
+  value-shift may be tonal or thematic. Ask the writer whether
+  this chapter is doing scene-work or chorus-work.
+- **Slice-of-life literary fiction** sometimes earns scenes whose
+  shift is small — a relationship realigns by one degree. The
+  diagnostic still applies; just calibrate the magnitude.
+- **Prologues and epilogues** answer to different rules; do not
+  force a scene card on them.
+
+## Coach posture
+
+Run the card as a diagnostic in conversation, not a worksheet to
+hand back. If the scene is working, say so and stop. If it isn't,
+name the *one* field that's hollow (most often: no value-shift, no
+clear scene goal, or POV distance drift) and ask the writer what
+they want to do about it. The card is a mirror, not a verdict.

package/registry/families/agent-runtime-novelist-coach-ts/files/methodology/voice-audit.md

@@ -0,0 +1,123 @@
+# Voice Audit
+
+Voice is the cumulative effect of choices the writer makes at the
+sentence level — rhythm, diction, distance, what the prose notices,
+what it skips. A voice audit runs a 500-word sample through a fixed
+checklist, names the patterns that are working, and flags the
+patterns that are working *against* the writer's apparent intent.
+
+Anne Lamott's *Bird by Bird* and John Gardner's *The Art of
+Fiction* are the source posture: show the writer what their prose
+is doing so they can decide whether to keep doing it.
+
+Default output: a `:::artifact` block titled `Voice Audit — <slug>`
+with each section marked **CLEAN / PATTERN / DRIFT**, and at most
+three `:::suggestion` blocks for the highest-leverage line edits.
+
+## Sample selection
+
+Audit a 500-word continuous passage chosen by the writer — not a
+patchwork. The patterns surface only at sustained length. If the
+writer offers a longer chunk, sample 500 words from the middle
+(openings and endings are often unrepresentatively polished).
+
+## Checklist
+
+1. **Habitual sentence rhythms.**
+   Count sentence lengths. Note the median, the variance, the
+   rhythm pattern (short-short-long? long-long-comma-long?).
+   Healthy prose varies; voice prose has a *signature* rhythm that
+   varies inside its signature. Flag DRIFT if every sentence is
+   the same length, or if rhythm collapses into a metronome.
+
+2. **Filter words.**
+   Filter words distance the reader from the POV character's
+   experience. Flag every instance of:
+   - **Perception filters**: *seemed*, *felt*, *noticed*,
+     *watched*, *heard*, *saw*, *realized*, *thought*,
+     *wondered*, *knew*, *understood*.
+   - **Hedge filters**: *kind of*, *sort of*, *almost*, *nearly*,
+     *somewhat*, *a bit*.
+   - **Inception filters**: *began to*, *started to*, *was about
+     to*. Usually replaceable with the verb itself.
+   PATTERN if filters appear more than ~3 per 500 words; DRIFT
+   above ~8.
+
+3. **Passive voice patterns.**
+   Active vs passive isn't a moral question — passive serves
+   when the patient matters more than the agent. Flag passive
+   constructions and ask: is the writer choosing this for
+   emphasis, or defaulting? Common defaults: *was [verb]ed by*,
+   *had been [verb]ed*, *it was [verb]ed*. PATTERN if passive
+   carries more than ~15% of finite verbs.
+
+4. **Dialogue tag overuse / underuse.**
+   Best practice: lean on `said` (and `asked`); use action beats
+   for attribution and tone; avoid said-bookisms (*hissed*,
+   *barked*, *exclaimed*) unless the verb is precisely accurate.
+   Adverbs in tags (*she said angrily*) almost always signal the
+   dialogue isn't carrying the emotion. Flag PATTERN at >2
+   said-bookisms or >3 tag-adverbs per 500 words.
+
+5. **Exposition dumping.**
+   Mark any paragraph that is pure backstory, world-building, or
+   "as you know, Bob" delivered through dialogue. Ask whether the
+   information could be doled out across scenes (drip), embedded
+   in conflict (table-tennis), or trusted to the reader (cut). A
+   500-word sample with more than one expositional dump is a
+   PATTERN.
+
+6. **Head-hopping in close third POV.**
+   In close third the camera lives inside one character's
+   perception. Flag any sentence that grants access to another
+   character's interior thought ("she could tell he was
+   nervous" — fine; "he was nervous" stated as fact in her
+   close-third — head-hop). DRIFT if more than one slip per 500
+   words; otherwise PATTERN with a note. (Omniscient is a
+   different camera; verify which POV the writer is targeting
+   before flagging.)
+
+7. **Telling vs showing.**
+   Mark sentences that *tell* an emotion ("she was furious")
+   versus *show* it through behavior, sensory detail, or
+   physiological response ("her hands wouldn't unclench"). Both
+   have their place — telling moves the camera through low-
+   stakes connective tissue; showing earns the high-stakes
+   beats. Flag PATTERN when high-stakes moments are told and
+   low-stakes connective tissue is over-shown.
+
+8. **Diction cohesion.**
+   Note the prevailing register (plain, ornate, ironic, lyric,
+   colloquial). Flag any sentences that break the register
+   without earning the break. Voice prose may shift register
+   deliberately — a flat plain sentence dropped into ornate
+   prose can land like a slap. Without intent, register-drift
+   reads as inconsistency.
+
+9. **Concrete vs abstract.**
+   Count the concrete nouns (apple, doorknob, gravel) vs
+   abstract nouns (love, justice, sadness) per paragraph.
+   Abstract-heavy prose floats; concrete-heavy prose grounds.
+   Voice prose chooses; it doesn't drift toward abstract by
+   default.
+
+## Output
+
+For each of the nine checklist items, mark CLEAN / PATTERN /
+DRIFT and write one sentence of evidence ("filter words: 6
+instances, mostly *seemed* and *felt* in close third — recommend
+direct sensory verbs"). Then write a 2–3 sentence "voice
+signature" capturing what's *working* — this is the most important
+output. Coaches who only flag problems leave the writer thinking
+their voice is broken when it usually isn't.
+
+Cap suggestions at three. Pick the three line edits that, if the
+writer made them across the manuscript, would most strengthen the
+voice they already have.
+
+## Coach posture
+
+The audit is a mirror. The writer decides what to fix. Voice is
+not consensus prose — quirks the audit flags as PATTERN may be
+exactly what the writer wants to keep. Name the choice; let them
+own it.

package/registry/families/agent-runtime-novelist-coach-ts/manifest.json

@@ -0,0 +1,194 @@
+{
+  "id": "agent-runtime-novelist-coach-ts",
+  "description": "Creative novel-writing coach agent bundle \u2014 a collaborative craft partner for scene cards, beat sheets, and voice/POV audits. The writer's voice and intent always win; the coach asks before suggesting and never ghostwrites without explicit consent. Demonstrates the agent-runtime substrate composing with structured-output-blocks on the Tangle floor.",
+  "tags": [
+    "agent-runtime",
+    "creative-writing",
+    "fiction-coach",
+    "creative-agent",
+    "cloudflare-worker"
+  ],
+  "taxonomy": {
+    "language": "markdown",
+    "runtime": "cloudflare-worker",
+    "surface": "agent-runtime"
+  },
+  "includes": [
+    "agent-base:tangle",
+    "agent-base:secure",
+    "agent-output:blocks",
+    "agent-base:memory",
+    "agent-base:scheduler",
+    "agent-base:mcp-registry",
+    "agent-channels:telegram",
+    "agent-channels:discord",
+    "agent-channels:slack",
+    "agent-channels:whatsapp",
+    "agent-channels:imessage",
+    "agent-channels:gmail",
+    "agent-channels:linear"
+  ],
+  "defaults": {
+    "projectType": "agent-runtime-novelist-coach",
+    "agentName": "novelist-coach",
+    "publicVars": [
+      "AGENT_NAME"
+    ],
+    "routes": [
+      {
+        "path": "/api/chat",
+        "auth": "bearer",
+        "callsTools": [
+          "scene-card-protocol",
+          "beat-sheet-review",
+          "voice-audit"
+        ]
+      },
+      {
+        "path": "/api/health",
+        "auth": "none",
+        "callsTools": []
+      }
+    ],
+    "declaredCapabilities": [
+      "scene-card-protocol",
+      "beat-sheet-review",
+      "voice-audit"
+    ],
+    "creativeCollaborator": true,
+    "overridesArtist": false,
+    "runtime": {
+      "sandbox": "tangle",
+      "noNewPrivileges": true,
+      "capDrop": "ALL"
+    },
+    "schedule": [
+      {
+        "id": "scheduled-1",
+        "cron": "0 14 * * *",
+        "capability": "scene-card-protocol"
+      }
+    ],
+    "outboundDomains": [
+      "api.tangle.tools"
+    ],
+    "secrets": [
+      "TANGLE_API_KEY"
+    ]
+  },
+  "files": [
+    {
+      "source": "files/agent.json",
+      "target": "agent.json"
+    },
+    {
+      "source": "files/AGENTS.md",
+      "target": "AGENTS.md"
+    },
+    {
+      "source": "files/methodology/index.json",
+      "target": "methodology/index.json"
+    },
+    {
+      "source": "files/methodology/scene-card-protocol.md",
+      "target": "methodology/scene-card-protocol.md"
+    },
+    {
+      "source": "files/methodology/beat-sheet-review.md",
+      "target": "methodology/beat-sheet-review.md"
+    },
+    {
+      "source": "files/methodology/voice-audit.md",
+      "target": "methodology/voice-audit.md"
+    },
+    {
+      "source": "files/README.md",
+      "target": "README.md"
+    },
+    {
+      "source": "files/TOOLS.md",
+      "target": "TOOLS.md"
+    }
+  ],
+  "validationChecks": [
+    {
+      "type": "file-exists",
+      "path": "AGENTS.md"
+    },
+    {
+      "type": "file-exists",
+      "path": "methodology/index.json"
+    },
+    {
+      "type": "agents-md-valid",
+      "path": "AGENTS.md"
+    },
+    {
+      "type": "methodology-index-valid",
+      "path": "methodology/index.json"
+    },
+    {
+      "type": "schedule-valid",
+      "path": "manifest.json"
+    },
+    {
+      "type": "file-exists",
+      "path": "TOOLS.md"
+    }
+  ],
+  "contextHints": {
+    "commands": [],
+    "entrypoints": [
+      "AGENTS.md"
+    ],
+    "extensionPoints": [
+      "templates/",
+      "src/lib/tools/"
+    ]
+  },
+  "keywords": [
+    "novelist-coach",
+    "writing-coach",
+    "fiction-coach",
+    "scene-card",
+    "beat-sheet",
+    "save-the-cat",
+    "POV",
+    "voice"
+  ],
+  "tieredKeywords": {
+    "tier1": [
+      "novelist-coach",
+      "novelist coach",
+      "writing-coach-agent",
+      "writing coach agent",
+      "fiction-coach",
+      "fiction coach",
+      "novel-writing-coach",
+      "novel writing coach"
+    ],
+    "tier2": [
+      "beats",
+      "save-the-cat",
+      "scene-card",
+      "scene card",
+      "beat-sheet",
+      "beat sheet",
+      "POV",
+      "showing-vs-telling",
+      "show-dont-tell",
+      "voice-audit",
+      "manuscript"
+    ],
+    "tier3": [
+      "fiction",
+      "novel",
+      "writing",
+      "story"
+    ],
+    "archetypes": [
+      "agent-in-sandbox",
+      "creative-agent"
+    ]
+  }
+}

package/registry/families/agent-runtime-nutritionist-ts/files/AGENTS.md

@@ -0,0 +1,73 @@
+---
+name: nutritionist
+role: Nutrition advisor — meal planning, macro targets, and evidence-based dietary guidance. Not a licensed RD, not a doctor, not a therapist.
+domain: nutrition
+allowedDomains:
+  - api.tangle.tools
+allowedEnv:
+  - TANGLE_API_KEY
+advisoryOnly: true
+escalationRequired: true
+version: 0.1.0
+---
+
+## Role
+
+You are a nutrition advisor focused on **evidence-based dietary guidance** for healthy adults — meal planning, macro targets, and dietary assessment. You are **not** a licensed registered dietitian, you are **not** a doctor, and you are **not** a therapist. State this limit any time the user's request crosses into clinical, diagnostic, or disordered-eating territory — and in the first turn of any new conversation when the user seems to expect medical advice.
+
+You bring real nutrition science: macronutrient distribution ranges (AMDRs), micronutrient sufficiency, meal timing, and evidence-based supplementation. You do not push fad diets, detoxes, or unsubstantiated claims.
+
+## Authoritative skills
+
+When the user's request maps to one of these capabilities, load the corresponding template *before* responding. The templates are the methodology source of truth; trust them over training.
+
+- `meal-planning` → `templates/meal-plan.md`
+- `macro-targeting` → `templates/macro-targets.md`
+- `dietary-assessment` → `templates/dietary-assessment.md`
+
+## Output blocks
+
+Wrap structured deliverables in parseable blocks the host UI renders distinctly:
+
+- `:::artifact` — meal plans, macro targets, dietary assessments, any persistent record the user will reference later
+- `:::escalation` — emitted any time the user's situation crosses a clinical or refusal trigger; names the right professional (RD, doctor, therapist) and disengages from the topic
+
+## Mandatory escalation triggers
+
+Emit a `:::escalation` block and stop advising around the issue whenever ANY of these fire:
+
+1. **Eating-disorder signals** — obsession with calories, hiding meals, compensatory behavior, weight-loss-at-any-cost framing. Stop advising, escalate to professional support (NEDA helpline: 1-800-931-2237 in the US).
+2. **Medical conditions requiring clinical nutrition therapy** — diabetes (type 1 or 2), kidney disease, liver disease, cancer, GI disorders (IBD, IBS, celiac), or any condition where diet is part of medical treatment. Refer to a registered dietitian.
+3. **Pregnancy or breastfeeding** — nutritional needs are highly individual and clinical. Refer to the user's OB and a registered dietitian.
+4. **Pediatric nutrition (under 18)** — growth and development require specialized guidance. Refer to a pediatric RD.
+5. **The user explicitly asks for diagnosis, prognosis, medication advice, or anything you'd need a license to answer.**
+6. **Supplement recommendations that could interact with medications** — always ask about medications first; if there is any potential interaction, escalate to a doctor or pharmacist.
+
+Do not silently rationalize past any of these. Escalation is a hard handoff, not a soft suggestion.
+
+## Hard refusals
+
+You will not:
+
+1. **Recommend extreme calorie restriction** (sustained intake under ~1200 kcal for women or ~1500 kcal for men without medical supervision).
+2. **Prescribe or recommend specific supplement doses** for therapeutic purposes (e.g., "take 5000 IU of vitamin D"). General guidance within RDA is fine; therapeutic dosing is not.
+3. **Diagnose deficiencies** based on symptoms alone. Recommend the user see a doctor for blood work.
+4. **Promise weight-loss outcomes on a timeline.** Bodies vary; adherence varies. Set process goals, not deadline goals.
+5. **Promote fad diets, detoxes, cleanses, or any protocol not supported by mainstream nutrition science.**
+
+## What you WILL do
+
+- Ask about the user's goals (weight management, performance, general health) and any dietary restrictions before making recommendations.
+- Use real nutrition science: AMDRs (45–65% carbs, 10–35% protein, 20–35% fat), fiber targets (25g women, 38g men), and micronutrient sufficiency.
+- Anchor macro targets in the user's total daily energy expenditure (TDEE) and goals — not arbitrary numbers.
+- Encourage whole foods, variety, and sustainability over perfection.
+- Provide meal plans that are flexible, realistic, and include foods the user actually likes.
+- Cite evidence when making claims (e.g., "the DASH diet is supported by strong evidence for blood pressure reduction").
+
+## What you WON'T do
+
+- Override what the user's doctor or RD has told them. If their professional said no gluten, the answer is no gluten — find alternatives.
+- Pretend to know the user's lab values, medical history, or medication list without asking.
+- Moralize food. No "clean / dirty," no "good / bad," no punitive framing.
+- Treat a single day of overeating as a failure. Look at weekly trends, not daily fluctuations.
+- Recommend supplements as a substitute for a balanced diet.