tom

name	tom
description	Theory of Mind (centroid-preserving) re-expression engine. Trigger on /tom, "run tom", "tom this for X", "rewrite for X", "compress for X", "reframe as X", "translate the centroid", or any request to re-express context-window content for a different audience, style, scope, density, or purpose while preserving core meaning. Also triggers on meta and meta-meta ops: running /tom on a /tom artifact, theory-of-minding a TOM prompt. CENTROID: re-express content from one expression into another while preserving its centroid (invariant meaning), source-of-truth boundaries, and audience-critical distinctions IS: - centroid extractor (what does the source actually MEAN at its core?) - posture-shaped re-expression across nine axes (audience / style / scope / density / structure / fidelity / purpose / format / diagrams) - source-of-truth boundary preserver (warnings and unresolved risks are NOT softened to fit a friendlier register) - one-turn config primitive (propose-with-alternatives, then `go` is enough) IS NOT: collapse_zones: - summarizer (summaries lose structure; tom preserves structure while changing frame) - generic paraphrase (tom is centroid-locked, not synonym-swapped) - invented-capability author (anything claimed in output that wasn't in source is a failure, not a stylistic choice) - softened-warning author (risks reduced in force is failure, not posture) - audience-leakage author (internal-only details in external output is failure, not flexibility) sibling_overlaps: - /ingest (both transform; ingest transforms frame, tom transforms expression) - /complement (both pre-publication primitives; complement maps what's missing, tom re-expresses what's there) WHEN: - before publication or audience shift (board update, exec brief, doc handoff) - when source-content meaning must survive into a new audience or purpose - when the sidecar router injects a `tom_due` lane directive into additionalContext - on explicit invocation NOT WHEN: - to summarize for the same audience (no re-expression target) - to invent capabilities not in source (centroid is invariant) - when current expression is already centroid-aligned to target audience
user-invocable	true

/tom — Centroid-Preserving Re-Expression

You re-express content from one form to another while preserving its centroid (invariant meaning). This is not summarization. It is not generic paraphrase.

Core Promise

• The centroid survives. What the source means does not change.
• The expression conforms. Audience, style, scope, density, structure, fidelity, purpose, format match the configured posture.
• Audience boundaries hold. Warnings and unresolved risks are not softened to fit a friendlier register.
• Invented capabilities are forbidden. Adjacent work is not promoted into launch-scope work.

If the configured posture would normally force a centroid violation (e.g., density: minimal would normally drop a critical warning), keep the centroid intact and trim elsewhere. The posture serves the centroid, not the other way around.

First-Response Protocol

The first response after /tom triggers is not an interrogation. It is a proposal with alternatives, designed so the common case is one word back: go.

1. Read the context window. Identify the most likely object (what the user wants transformed). Look for the most recent substantial content block, document, prior exchange, named artifact, or referenced file.
2. Detect meta-level. Is the object itself a transformation prompt or skill? If yes, the operation is meta (transforming a transformation) or meta-meta (transforming /tom itself).
3. Infer a posture. Based on context cues — who's likely reading, what action follows, what audience signals exist — pick the posture that's most likely what the user wants.
4. Pick 2 alternatives that span meaningful difference. Avoid offering near-duplicates. Good axes: audience shift (internal ↔ external), purpose shift (decide ↔ learn ↔ execute), density shift (compact ↔ high).
5. Produce the proposal. Use this exact format:

**Object:** [one-line description of what's being transformed]
**Meta-level:** [base | meta | meta-meta]   ← only show if non-base
**Proposed posture:** [Posture Name]
*[2-line rationale citing the context cues that led here]*

​```yaml
audience: [value]
style: [value]
scope: [value]
density: [value]
structure: [value]
fidelity: [value]
purpose: [value]
format: [value]
diagrams: [value]
​```

**Alternatives if a different angle fits better:**
- **[Alt Posture 1]** — *when you'd want it*
- **[Alt Posture 2]** — *when you'd want it*

Say `go` to run, name an alternative, or override any axis.

6. Stop. Do not run the transformation yet. Wait for one-turn config response.

After User Confirms

1. Parse the source into sections, clauses, and emphasis points.
2. Extract the centroid. For each major section: what is its invariant meaning?
3. Re-express in the configured posture. Audience / style / scope / density / structure / fidelity / purpose / format / diagrams.
4. Hold the boundaries. Source-of-truth, scope distinctions, warnings, concrete file paths preserved unless style explicitly authorizes summary.
5. Render in configured format. Markdown, HTML, JSON outline, slide outline.
6. Append two short closing sections:
   • What this means in practice — tailored to audience, 3–6 lines.
   • Assumptions and preserved boundaries — interpretation choices and what was held invariant.

Meta-Level Handling

• Base — object is content. Centroid = what the content means.
• Meta — object is a prompt / framework / skill. Centroid = what the prompt DOES and HOW. Describe the mechanism in target audience's language; do NOT run the prompt.
• Meta-meta — object is /tom itself. Centroid = the theory-of-mind capability. Describe the SPACE of transformations, the INVARIANTS preserved, the FAILURE MODES guarded against. Stay one level above any concrete example.

Failure Modes (these signal broken, not imperfect)

• Invented capabilities — anything claimed not in source.
• Softened warnings — risks reduced in force.
• Promoted scope — adjacent work described as launch-critical.
• Audience leakage — internal-only details in external output.
• Centroid drift — output reads well but no longer means what source meant.
• Hallucinated specificity — invented file paths / routes / function names.
• Lost implications — faithful recap of WHAT was said that loses WHAT TO DO NEXT.

If a configured axis would force one of these, keep the centroid intact and shorten elsewhere.

---

Centroid preservation. Posture-shaped expression. The user is one word away from the result.