Prompt Proximity Architecture

Prompt Proximity Architecture

Design the cells before writing the strings.

This skill inherits the ethical floor from skills/ETHICS.md. It enforces anti-hallucination and evidence-bound coverage. Anti-spray and human-send are not applicable.

Inputs

Require:

• a measurement charter;
• approved icp_hypotheses.json and buyer_jobs.json;
• target run/review budget;
• required locales, surfaces, and lanes;
• any campaign partition and prior-panel constraints.

If the charter is provisional, design a provisional architecture and name the gaps. Reject a charter that says only “track AI visibility.”

Keep dimensions independent

Each intent cell fixes:

• buyer job;
• information act;
• journey state;
• material constraints;
• persona or buying role;
• locale and language;
• prompt-proximity band;
• expected answer kind.

Add independent tags for evidence grade, partition, lane eligibility, turn form, and optional funnel. Do not make funnel the schema.

Use the supported acts explain, diagnose, plan, generate, compare, recommend, verify, navigate, buy, implement, and troubleshoot. Include only acts entailed by the job evidence.

Use journey states problem_identification, exploration, requirements_building, supplier_selection, adoption, and post_purchase.

Assign proximity

Use the code in prompt_architecture.json and the name in every Markdown summary. The band codes are labels for structurally different prompts, not points on a scale: do not write "B4" in prose, do not order or average them, and do not describe one band as further along than another. See "Display names" in ../build-ai-visibility-panel/references/artifact-contracts.md.

Bands and aided status are correlated by construction — Brand is always target_aided, Category always category_aided, and Problem, Goal and Market are always unaided, so only Shortlist splits. Report the two together rather than as if they were independent findings, and do not present a band breakdown that merely restates the prompted split.

campaign_exposed is an independent flag. Never combine target-aided, competitor-aided, category-aided, unaided, or campaign-exposed cells in one denominator.

Because aided_status is single-valued, a B1 prompt naming both the target and a competitor is target_aided; record the competitor stimulus in the contamination exception and keep the cell in the aided partition. A B1 prompt naming competitors but not the target is competitor_aided.

Do not assign funnel mechanically:

• B0 can be post-purchase, not BOFU.
• B3 can describe an urgent funded decision.
• B1 can be early exploration.
• funnel may be null.

B5 requires fresh dated public evidence and a decay/refresh rule.

The charter's product/capability perimeter is also a coverage dimension. For every named area, create at least one evidence-supported job/cell or a coverage_gap with an exclusion/waiver and the evidence needed. Never silently omit a named area.

Allocate coverage

Cover every dimension that the evidence supports, not every Cartesian combination. Never generate a full persona × locale × act × constraint grid.

For a diagnostic default, target 30–48 unique unaided cells, two variants per cell, and three repeats per wave. For a standard panel, target 60–120 cells. Adjust to the user's budget and keep aided cells outside the unaided quota.

Stratify minimums and targets across:

• B0–B5 where evidence supports each band;
• job, journey, and information act;
• ICP/role and materially distinct locale;
• evidence grade/source type;
• core, rotating, sentinel, control, and aided partitions;
• closed-model, retrieval, consumer-surface, and campaign lanes.

Protect underrepresented but decision-relevant strata. Grade-D cells are rotating discovery until independently validated or explicitly promoted.

When the budget cannot meet required coverage, return an allocation conflict. Do not silently drop a band, role, locale, or control.

Prohibit unsupported cells

Mark a cell prohibited when it:

• changes an informational need into a recommendation solely to elicit brands;
• introduces a category not supported by the job evidence;
• treats a competitor-named prompt as unaided;
• uses target/campaign copy outside the allowed aided lane;
• assumes translation is locale equivalence;
• derives a core cell from current AI answers or target pages;
• exceeds the declared sampling budget without a waiver.

Output

Give a Markdown coverage summary first: required dimensions, planned counts, gaps, conflicts, and why each missing band is unsupported.

Then write prompt_architecture.json:

jsonCopy code
{
  "schema_version": "1.0.0",
  "artifact_id": "architecture-<stable-slug>",
  "created_at": "RFC3339",
  "created_by": "declared agent or human",
  "source_manifest_hash": null,
  "warnings": ["hash_not_computed: compute source_manifest_hash before freeze"],
  "cells": [
    {
      "cell_spec_id": "spec-001",
      "job_id": "job-001",
      "icp_ids": ["icp-001"],
      "proximity_band": "B3_problem_need",
      "aided_status": "unaided",
      "campaign_exposed": false,
      "information_act": "diagnose",
      "journey_state": "problem_identification",
      "funnel": null,
      "persona_id": "role-001",
      "locale": "en-CA",
      "language": "en",
      "material_constraints": [],
      "expected_answer_kind": "diagnosis_and_options",
      "turn_form": "single_turn",
      "lane_eligibility": ["closed_model", "retrieval"],
      "partition": "core",
      "evidence_grade": "A",
      "target_variants": 2,
      "required": true,
      "reason_source_ids": ["source-001"]
    }
  ],
  "prohibited_cells": [],
  "allocation": {
    "core_cells": 36,
    "rotating_cells": 8,
    "sentinel_cells": 12,
    "control_cells": 0,
    "aided_cells": 6,
    "budget_status": "within_budget | conflict | waived"
  }
}

Before handoff, require every cells[].job_id and every cells[].icp_ids[] to resolve exactly. There is one authoritative field for each relationship. Never leave a stale ID and add a compensating alias such as job_id_authoritative; rewrite the cell or stop with an allocation/reference error.

Handoff

Pass the architecture plus a target-free blind brief to realistic-prompt-generation. Keep the contamination register outside the generator's context.