CPU Price/Spec Comparer

Comparing CPUs for a budget build means opening many retailer pages, finding the same specs in different layouts and typing them into a table. It is slow, easy to get wrong, and the pages are full of noise: sidebars, adverts, bundles, conflicting figures and unlabelled numbers.

Phase 1 is a no-code tool built as written instructions in a Claude Project, so the main work was prompt design and testing. It produces one row per product (name, price, cores/threads, base and boost clock, socket, power draw, integrated graphics, source, date) plus a short list of flags to check by hand. The core rule is grounding: use only the pasted text and write "not stated" rather than guess. Scope was kept to CPUs so every value can be checked against the page.

• One consistent row per product, ignoring sidebars, adverts and bundles
• Grounded output: missing values are "not stated", never guessed
• Conflicts between sources resolved in favour of the retailer's own text and flagged
• Flags for used, OEM, non-UK or non-£ listings
• Four instruction versions (v1, v2, v2.1, v3) kept and tested against the same answer key

Claude Projects

Running the instruction-based extraction tool

Prompt engineering

Rules for grounding, formatting and flagging

Hand-built answer key

Field-by-field evaluation of every version

Python + LLM API (planned)

Phase 2: batch processing to CSV

v1 instructions

46/50 fields correct, 10 unwanted extra rows

v2 instructions

49/50 fields correct, 0 extra rows

v2.1 and v3 instructions

50/50 fields correct, 0 extra rows

Five messier pages (used, comparison, manufacturer, marketplace, OEM)

All matched the answer key on v3, with problems correctly flagged

Timed run vs by hand (5 pages)

3 min 15 s vs 9 min 29 s, a 66% saving, with all 50 fields correct

v1 filled in clock speeds from unlabelled figures, built extra rows from sidebar listings and put two values in one field when sources conflicted. Each error was logged by type and fixed with a targeted rule. Several differences turned out to be mistakes in my own answer key, not the tool, which showed how risky manual checking can be. Every test ran in a fresh chat so earlier pages could not influence later ones.

• Small sample: ten test pages, one run each per version.
• The hand table was done first, so 66% time saving is likely an upper bound.
• Output varied between runs in v1, so one clean run does not prove reliability.
• Some answer-key judgement calls, e.g. non-£ prices are "not stated" rather than converted.
• Prices are snapshots; each row records its date and source.

Phase 1 (no-code version) is complete: v3 matched the answer key on all ten test pages. Phase 2 is in progress.

• Phase 2: a Python script that sends page text to an LLM API and writes a CSV, using the error log as an automated test set
• Add sanity checks such as prices outside a sensible range and missing specs
• Test more OEM or tray listings
• Retest timing on new pages, alternating hand-first and tool-first
• Possible extension: CPU socket and motherboard compatibility check