saintpick.

The small-team AI workflow fieldbook

Original editorial frameworks for small teams evaluating AI workflows. These are decision aids and hypothetical examples, not measured product rankings.

1. Decide where automation belongs

QuestionIf the answer is no
Can you define a usable output?Write the task brief first.
Can you inspect the result?Add an explicit review step.
Can you recover from a mistake?Keep approval before the consequential action.

A narrow, reversible trial is a useful starting point. This decision aid is not a guarantee of safety.

2. A task brief with an acceptance check

INPUT: [source + context]
OUTPUT: [format + audience]
MUST KEEP: [facts and terms]
UNKNOWN: mark missing information; do not invent it
ACCEPT IF: [checks a reviewer can observe]

Example: summarize a meeting as decisions, owners and deadlines. If an owner was not stated, label the field UNKNOWN rather than assigning one. This prompt has not been benchmarked against commercial models.

Work through a fictional meeting example and download the review CSV.

3. Cost per accepted output

Invented monthly example, excluding correction time:

ToolSubscriptionAccepted outputsCost each
A$2010$2
B$4040$1
(Allocated subscription cost + correction hours × hourly rate)
÷ accepted outputs

Use the same period and acceptance standard. An accepted output must meet the task brief. Avoid dividing by zero; a trial with no accepted output has not yet produced a usable unit. Account separately for setup costs if relevant.

4. Treat an unknown result as unknown

After a publishing timeout, the remote action may have succeeded. Record an item ID and attempt state, then check the destination before retrying. Confirmed success, confirmed failure and unknown need different handling. This pattern does not imply that every API offers idempotency or perfect reconciliation.

5. Bring a small test pack

Use an easy input, a typical weekly task and a difficult case from your own work. Hold the requested output and acceptance checks steady. Record settings, corrections, failed outputs and time spent reviewing. Three inputs can expose trade-offs but cannot establish a universal winner.

6. Review observations separately from explanations

Record the available views, saves, replies and link clicks in an observed column. Write possible explanations in a separate inferred column. Keep the comparison window consistent, consider sample size and change one factor for the next trial.

Try our reproducible caption structure experiment.