Rich Webb (@realrichwebb) header
Rich Webb (@realrichwebb) profile picture

The PerformanceNinja. Your business runs on people. It should run on systems. | AI & automation builder for service businesses | Automation Geeks ↓

London, UK
Joined September 2023
Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)Photo by Rich Webb (@realrichwebb)
If you work for yourself and struggle to manage your time, read this:
17
51
4
349
73,637
Put a real example beside every calculated metric: inputs, expected result and rounding rule. It gives the next person a way to check the formula without guessing what the original author meant by the column name.
101
The longest prompt is not necessarily the clearest specification. If two instructions conflict, adding a third paragraph may hide the problem. Resolve the contradiction before asking a model to reason its way around it.
145
A useful automation case can be fewer interruptions. Count how often someone must leave focused work to answer a routine status question. The value is not just the minute spent replying, but the broken concentration.
1
222
Before automating a weekly meeting pack, ask which decisions actually use it. If half the pages never change a decision, stop producing them. Making an unused report faster still leaves everyone with an unused report.
1
240
Two people edit the same record while a workflow is running. Which change survives? Use a version check before writing back. A successful API response does not mean you preserved the latest human correction.
231
An agent can finish a task and still miss its deadline. Give time-sensitive work an expiry time, then check it before acting. Sending yesterday's correct answer today can be worse than returning no answer.
1
251
Take one automated document and read it aloud without looking at the layout. Do the headings, links and reading order still make sense? Accessibility belongs in the output specification, not in a later tidy-up.
1
224
Imagine an automated rota that always gives the same person the awkward shift because they are most available. The rules can be correct and the result unfair. Examine who repeatedly absorbs the inconvenience.
288
Testing an agent across languages? Include decimal commas. Use ambiguous date formats. Try local currency notation. Check translated product names. Fluent prose is not proof that the agent interpreted the transaction correctly.
1
316
Before trusting an agent's citations: Check the source exists. Open the cited passage. Match it to the claim. Check the document date. A plausible link is not evidence that the source actually supports the answer.
1
316
Hypothetical example: a delivery booked for 9am is stored as UTC but displayed as local time. Test workflows across a clock change, not just across dates. A valid timestamp can still produce the wrong appointment.
370
Token-based AI pricing separates input from output, and some APIs price cached input differently. A model price comparison needs a representative request shape, not just the cheapest number on the pricing page.
4
1
328
Separate missing, zero and not applicable in your spreadsheets. They answer different questions. Treating all three as a blank can turn an honest unknown into a misleading average when a workflow calculates the totals.
296
Many SaaS APIs retain deleted records as archived items or deletion markers. A sync that only copies active records can leave ghosts downstream. Decide how deletion should propagate before connecting the systems.
276
An automation that asks for more data can make a process worse. Before adding a mandatory field, name the decision it changes. If nobody can explain how the answer affects the work, remove the question.
2
290
A monthly report may take only twenty minutes to assemble. The stronger case for automating it could be a fixed reporting date and comparable definitions, so the next discussion starts with decisions instead of disputes.
1
280
How many times does your team enter the same customer address? Count the copies, then choose one authoritative record. Removing repeated data entry is useful, but preventing conflicting versions may be the bigger win.
206
AI APIs offer structured output modes that can enforce a JSON shape. That helps software read the answer. It does not prove the values are true. Validate business facts separately from field names and data types.
260
Exporting 100 records is not proof that a sync is complete. Follow every page token, compare the fetched count with the source total where available, and test a dataset larger than one page before trusting the workflow.
1
206
An agent reporting 'nothing found' should distinguish an empty result from a failed search. Check whether the query ran, which sources it covered and whether filters excluded records. Absence needs evidence too.
1
211