How Adams Web Searcher Works
You type a plain-language request — plant, topic, and date range in your own words.
Claude then drives the search itself, using the firm's ADAMS search methodology
skill as its playbook and calling the NRC ADAMS API directly. It works in a
chat: each step appears below the last, and Claude stops at two checkpoints
to get your approval before doing expensive work.
How a search unfolds
- Plan & confirm
- Claude interprets your request, resolves the plant to its NRC docket number(s),
and lays out its search plan — including, for a design-basis question, the five
document "buckets" the methodology calls for. It stops and waits for
you to reply “go” before searching.
- Search & triage
- Claude runs the searches against the NRC ADAMS Public Search (APS) API,
then rates every result (★★★ HIGH / ★★ MEDIUM / ★ LOW / — NOISE) and shows the triage.
It stops again for your approval on which documents to open.
- Read & report
- After you approve, Claude retrieves each document's indexed text, reads it, and
synthesizes a plain-language analysis with linked citations — the core deliverable.
At any checkpoint you can reply with a
correction instead of “go” — change the dockets, the date range, the ratings, or anything
else — and Claude adjusts before continuing.
Data & privacy
- All AI runs on Mat's own paid Claude API tokens, server-side —
no Claude account needed to use this tool.
- The Claude and ADAMS API keys are stored as encrypted Cloudflare secrets and
never reach your browser.
- No searches or results are saved between sessions.
Model choice
Sonnet 4.6 is the default and handles most searches well.
Haiku is for basic functional testing only.