How to Give AI a Crosswalk Between Old and New Category Names

Summary

• A “crosswalk” is a mapping that tells AI how old category names translate to new ones, including edge cases and “no match” outcomes.
• The main ambiguity is that the same old label can split into multiple new labels (or multiple old labels can merge), so you must specify rules and examples.
• Use a compact, machine-readable format (JSON/YAML/table) plus a short decision policy so the model can classify consistently.
• Add targeted checks: coverage, collisions, forbidden outputs, confidence thresholds, and an “ask a question” fallback.
• Save your crosswalk and prompts so you can reuse them across projects and keep results consistent as names evolve.

When you rename categories (in a product catalog, analytics taxonomy, support tags, CRM stages, or content labels), AI can get stuck on a simple but costly ambiguity: it does not know whether an old name is a synonym, a parent category, a deprecated label, or something that should split into multiple new categories depending on context.

A “crosswalk between old and new category names” solves that by giving the model an explicit mapping plus rules for how to apply it. This article shows (1) the specific ambiguity you need to remove, (2) a reusable crosswalk format you can paste into AI, (3) a ready-to-run prompt, and (4) targeted checks that catch the most common failure modes.

The specific ambiguity: why “just rename it” fails

The title’s ambiguity is not “AI needs more context” in general. It’s narrower:

• One-to-many ambiguity: one old category can map to multiple new categories depending on the item’s attributes.
• Many-to-one ambiguity: multiple old categories collapse into one new category.
• Partial overlap: the old category covered a broader scope than any single new category.
• Same words, different meaning: a label stays similar but the definition changes.
• Unknown/retired: some old categories should not be mapped at all (or should be flagged for review).

Before-and-after example (the ambiguity in action)

Before (old taxonomy):

• “Accessories”
• “Cables”
• “Chargers”
• “Audio”

After (new taxonomy):

• “Power & Charging”
• “Connectivity”
• “Audio & Headphones”
• “Cases & Protection”

The ambiguity: An old item labeled “Accessories” could be a case, a charger, a cable, or something else. If you only tell AI “Accessories is now Power & Charging,” you will misclassify cases and audio accessories. The crosswalk must include rules (what signals to use) and fallback behavior (what to do when signals are missing).

What a good AI crosswalk contains (minimum viable components)

To make the mapping usable by AI (and by humans reviewing AI output), include these components:

• Old label: the exact string(s) you expect to see.
• New label: the allowed output label(s).
• Definition notes: what the label means now (short, operational language).
• Decision rules: how to choose when there are multiple candidates (keywords, attributes, examples).
• Examples: 2-5 “if you see X, map to Y” examples per tricky label.
• Non-matches: explicit “do not map” or “needs review” conditions.

If you want the model to be consistent, you also need one more thing: an output contract (exact fields to return, allowed values, and what to do when uncertain).

A reusable crosswalk format you can paste into AI (JSON)

This format is designed to be both readable and strict. You can paste it into ChatGPT or another model as-is, then update the mapping rows over time.

{
  "new_categories_allowed": [
    "Power & Charging",
    "Connectivity",
    "Audio & Headphones",
    "Cases & Protection",
    "Needs review"
  ],
  "policy": {
    "output_must_be_one_of_allowed": true,
    "if_insufficient_info": "Needs review",
    "do_not_guess_from_brand_alone": true,
    "return_reason": true
  },
  "crosswalk": [
    {
      "old": ["Chargers", "Charging"],
      "candidates": ["Power & Charging"],
      "rules": [
        "If item mentions charger, charging, power adapter, wall adapter, USB-C PD, MagSafe charger -> Power & Charging"
      ],
      "examples": [
        {"input": "65W USB-C wall charger", "output": "Power & Charging"},
        {"input": "Wireless charging pad", "output": "Power & Charging"}
      ],
      "fallback": "Power & Charging"
    },
    {
      "old": ["Cables"],
      "candidates": ["Connectivity", "Power & Charging"],
      "rules": [
        "If cable is primarily for data/connection (HDMI, Ethernet, Thunderbolt, DisplayPort) -> Connectivity",
        "If cable is primarily for charging/power (USB-A to USB-C charging cable, power cable) -> Power & Charging",
        "If unclear, ask for connector type or use Needs review"
      ],
      "examples": [
        {"input": "HDMI 2.1 cable", "output": "Connectivity"},
        {"input": "USB-C charging cable 2m", "output": "Power & Charging"}
      ],
      "fallback": "Needs review"
    },
    {
      "old": ["Audio"],
      "candidates": ["Audio & Headphones"],
      "rules": [
        "If item is headphones, earbuds, headset, speaker, microphone -> Audio & Headphones"
      ],
      "examples": [
        {"input": "Noise-cancelling headphones", "output": "Audio & Headphones"}
      ],
      "fallback": "Audio & Headphones"
    },
    {
      "old": ["Accessories"],
      "candidates": ["Cases & Protection", "Power & Charging", "Connectivity", "Audio & Headphones", "Needs review"],
      "rules": [
        "If mentions case, cover, sleeve, screen protector -> Cases & Protection",
        "If mentions charger, charging, adapter -> Power & Charging",
        "If mentions cable types (HDMI/Ethernet/Thunderbolt) -> Connectivity",
        "If mentions headphones/earbuds/speaker/mic -> Audio & Headphones",
        "If none match -> Needs review"
      ],
      "examples": [
        {"input": "iPhone case - silicone", "output": "Cases & Protection"},
        {"input": "USB-C to HDMI adapter", "output": "Connectivity"}
      ],
      "fallback": "Needs review"
    }
  ]
}

Why this works: you are not asking the model to “figure out” your taxonomy. You are giving it a constrained set of outputs, explicit decision rules, and a safe fallback.

A reusable prompt: map old categories to new categories (with checks)

Paste this prompt, then paste your crosswalk JSON and your items. It is designed to produce consistent, reviewable output.

You are mapping items from an old category taxonomy to a new taxonomy using the provided crosswalk JSON.

Instructions:
1) Use ONLY the new categories listed in crosswalk.new_categories_allowed.
2) For each item, apply the crosswalk rules for its old category.
3) If multiple candidates match, choose the best match and explain why.
4) If information is insufficient to choose confidently, output "Needs review" and ask ONE clarifying question.
5) Do not invent new categories. Do not output the old category as the new category.

Return JSON array with fields:
- id
- old_category
- new_category
- confidence (High/Medium/Low)
- reason (1-2 sentences)
- clarifying_question (string or null)

CROSSWALK JSON:
[paste crosswalk JSON here]

ITEMS:
[paste items here, e.g. id, title, old_category, optional description/specs]

Example input items

[
  {"id":"A-101","title":"USB-C to HDMI adapter","old_category":"Accessories","description":"Connect laptop to TV"},
  {"id":"A-102","title":"USB-C cable 2m (charging)","old_category":"Cables","description":"Fast charging cable"},
  {"id":"A-103","title":"Protective case for tablet","old_category":"Accessories","description":"Shockproof cover"},
  {"id":"A-104","title":"USB-C cable 1m","old_category":"Cables","description":""}
]

Example output (what “good” looks like)

[
  {
    "id":"A-101",
    "old_category":"Accessories",
    "new_category":"Connectivity",
    "confidence":"High",
    "reason":"The item is an adapter used to connect devices (USB-C to HDMI), which matches the Connectivity rule.",
    "clarifying_question":null
  },
  {
    "id":"A-102",
    "old_category":"Cables",
    "new_category":"Power & Charging",
    "confidence":"High",
    "reason":"The description indicates the cable is for charging, which maps to Power & Charging.",
    "clarifying_question":null
  },
  {
    "id":"A-103",
    "old_category":"Accessories",
    "new_category":"Cases & Protection",
    "confidence":"High",
    "reason":"A protective case/cover is explicitly routed to Cases & Protection in the Accessories rules.",
    "clarifying_question":null
  },
  {
    "id":"A-104",
    "old_category":"Cables",
    "new_category":"Needs review",
    "confidence":"Low",
    "reason":"The item is a USB-C cable but the use (data vs charging) is not specified, and the crosswalk requires that distinction.",
    "clarifying_question":"Is this cable primarily intended for charging/power, or for data/video connectivity?"
  }
]

Targeted checks that catch crosswalk failures early

Once you have a crosswalk and a prompt, the next step is to add checks that detect when the mapping is drifting or underspecified. These are practical checks you can run manually (spot-checking) or by asking the model to self-audit its own output.

Check 1: Coverage (do you handle every old category?)

• List all distinct old categories in your dataset.
• Confirm each appears in the crosswalk old list.
• If not, decide: add mapping rules or route to “Needs review.”

Check 2: Collision (are two old categories mapping to the same new category unintentionally?)

• Many-to-one merges are fine when intentional.
• Flag collisions where the old categories had meaningfully different definitions (you may need to split the new taxonomy or add rules that preserve the distinction).

Check 3: Forbidden outputs (is the model inventing categories?)

• Require the output to be one of the allowed new categories.
• If you see “Other,” “Misc,” or a slightly misspelled category, tighten the output contract and re-run.

Check 4: Confidence gating (when should it refuse?)

• Define what “Low confidence” means in your context (missing attributes, ambiguous title, conflicting signals).
• Require “Needs review” when confidence is Low, plus one clarifying question.

Check 5: Drift check (did the meaning of a label change?)

• If “Audio” used to include speakers and now “Audio & Headphones” excludes speakers (or vice versa), write that explicitly in the definition notes and add examples.
• When definitions change, update the crosswalk rules; do not rely on the label text alone.

Two crosswalk patterns you can reuse in real taxonomies

Pattern A: Split one old category into multiple new categories

Use this when the old label was a bucket (“Accessories,” “General,” “Other”). The key is to route by observable signals (keywords, attributes, fields you actually have).

• Define 3-6 candidate new categories.
• Write rules in priority order (most specific first).
• Include a “Needs review” fallback when signals are missing.

Pattern B: Merge multiple old categories into one new category (with traceability)

Use this when you intentionally simplify. The risk is losing traceability for reporting or audits. If you still need traceability, keep the old category in output as a separate field (do not try to encode it into the new label).

• Map old categories A/B/C to new category X.
• Return both old_category and new_category in the output.
• Add a reason string so reviewers can spot odd merges.

Where CopyCharm fits: saving and reusing your crosswalk + prompts

When category names change, the hard part is not writing one prompt once. It is keeping the mapping, rules, and examples consistent across weeks of work and across different tools (spreadsheets, ticketing systems, docs, and AI chats).

CopyCharm is a Windows desktop app that saves copied text locally so you can search past clips, favorite important clips, and separately save reusable prompts. A practical workflow for crosswalk work looks like this:

• Save: Copy your crosswalk JSON, your “mapping prompt,” and a few gold-standard examples (inputs + correct outputs). Save the prompt separately so you can reuse it without re-copying from old chats.
• Find: When you are about to run another batch, search in CopyCharm for “crosswalk”, the old taxonomy version name, or a specific tricky label (like “Accessories”). Pull up the exact mapping rules you used last time.
• Reuse: Paste the saved prompt and crosswalk into your AI tool, then paste the new batch of items. If you discover a new edge case, update the crosswalk text you keep and reuse going forward.

If you want ChatGPT to help you retrieve the crosswalk text without hunting through old conversations, CopyCharm also has an authenticated ChatGPT connector backed by optional AI Access sync and a read-only MCP service. After eligible account authorization and AI Access sync, ChatGPT can search and retrieve supported Synced Data; it cannot access unsynced local CopyCharm data. That means you can keep working locally day-to-day, and only sync the categories you choose (Favorite Clips, Saved Prompts, and optionally Other Clips within a selected time range) when you want ChatGPT retrieval.

For Claude, Gemini, Cursor, email, documents, and other applications, the workflow is manual: search or retrieve the crosswalk in CopyCharm, then copy/paste it into the destination application.

Try CopyCharm for saving and reusing your crosswalk prompts

Common mistakes (and how to fix them)

Mistake 1: Only providing a two-column mapping

A two-column list (old -> new) fails when any old label can split. Fix it by adding candidates + rules + examples, and by allowing “Needs review.”

Mistake 2: Letting the model output “close enough” labels

If your new taxonomy has strict names, require the output to be one of an allowed list. This prevents silent typos and near-matches that break downstream systems.

Mistake 3: No explicit fallback behavior

If you do not define what to do when information is missing, the model may guess. Fix it with a confidence gate and a single clarifying question.

Mistake 4: Rules that depend on fields you do not have

If your dataset only has “title” and “old_category,” do not write rules that require “material,” “voltage,” or “department” unless you can supply them. Keep rules grounded in available signals.

Frequently Asked Questions

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Related Guides

• What Is an AI Context Pack?
• How to Prepare Better Context for ChatGPT
• How to Organize Copied Text for AI Work
• Source-Labeled Context for AI: Why It Matters
• ChatGPT Context Pack for Consultants