Reference cabinet / Eight field notes + one tool

Understand the system.
Then test the fit.

Move from foundations to practical use, architecture, evaluation, safety, and cost. Every page says what kind of evidence it uses and what it does not establish.

Reading map connecting foundations, workflows, reliability, and economics. — **The cabinet at a glance**Four linked routes from understanding the system to measuring accepted results.

Search the cabinet

9 entries

Foundations

Build the mental model.

The product is more than the model. These notes make the layers and instructions explicit.

03 / FOUNDATIONS↗

How language models, context, and memory fit together

A practical map of generation, working context, retrieval, tools, and systems of record.

Evergreen explainer

04 / PROMPTING↗

Write the prompt as a testable contract

Define the objective, evidence boundary, output, unknown state, and acceptance checks.

Reusable method

Workflows & systems

Ground the work.

Choose the smallest architecture that can meet the task and preserve a review trail.

05 / RESEARCH↗

A source-grounded research workflow

Move from discovery to claim ledger, synthesis, counterevidence, and publication checks.

Step-by-step workflow

06 / SYSTEMS↗

Retrieval, tools, workflows, or agents?

Match system complexity to path variability, external evidence, and action risk.

Architecture guide

Reliability & safety

Test the boundary.

Find the failed layer, use representative cases, and keep authority outside the model.

01 / EVALUATION↗

A model card is the start of the test

Read claims with their task, conditions, limits, and missing coverage—then build a representative check.

Method note

07 / TROUBLESHOOTING↗

Diagnose AI failures before rewriting the prompt

Separate model, retrieval, OCR, schema, tool, permission, and interface failures.

Diagnostic order

08 / SAFETY↗

Safe AI workflows: data, tools, and approval

Set data, permission, confirmation, recovery, and human-review boundaries.

Risk guide

Economics

Count usable results.

Request price is one input. Retries, review, and acceptance determine the practical cost.

02 / ECONOMICS↗

Count accepted results, not just tokens

Measure retries, review time, and the share of outputs that pass the acceptance rule.

Method note

TOOL / ECONOMICS↗

Accepted-result calculator

Use your own request costs, review time, retry rate, and acceptance rate. Calculation stays in your browser.

Interactive tool