suede-ops-assessment

Files (5)

• CARD.md
• SKILL.md
• openai.yaml · agents/
• evals.json · evals/
• interview-guide.md · references/

SKILL.md (first lines)

# Suede AI Operations Assessment Skills

## Compatibility

Works from operator interviews and authorized process records. No integration is required for an assessment; connected data access must be authorized. Agent estimates must be marked separately from requester-supplied figures.

For a concrete input, expected output, recovery path, and validation check, read
[the worked contract](references/worked-contract.md). Use it to check the handoff
before treating the task as complete.

```text
Iron law: map the floor, not the org chart.
```

Every operation has two maps. The leadership map describes how the business is
meant to run. The floor map describes how the work actually gets done, including
each workaround, each unofficial spreadsheet, and each extra step that exists
because something broke once. A system designed from the leadership map gets
routed around, because the people doing the work can feel the mismatch on day
one.

This skill produces the floor map and the numbers attached to it. It stops
before deciding what to build.

## The requester's numbers are given

Their hours, costs, volumes, error rates, and history are inputs, not claims to
audit. No step here verifies, scores, hedges, or gates on a figure the requester
supplied.

What does get audited is anything **this skill** produces: a benchmark it
reached for, an estimate it filled in, a process step it inferred rather than
heard. Mark every one of those inline as `[assumed]` and list them in the Output
Contract, so a reader can tell the operation's own numbers from this skill's.

Never substitute an industry average for a number the requester has. Ask for
theirs, or record the line as missing.

## Before Starting

1. **Scope**: which processes, departments, or surfaces are in the assessment.
2. **Access**: what the requester has authorized you to read. Work inside it.
3. **Who does the work**: names or roles per process, separated into people who
   perform it and people who manage it.
4. **Self-applied or on behalf**: a founder assessing their own operation reads
   Step 5 differently from an outside team assessing a client's.

Read `references/interview-guide.md` before the first interview.

## Step 1: Interview the floor

Interview at least one person **per process** who performs it, not only the
person who manages it. Where the two descriptions differ, record both and mark
the performer's version as the floor map.

Run it as a walkthrough rather than a survey: the question is what happens next,
repeated until the process ends. The guide carries the full question bank; these
four earn their place in every interview.

- Walk me through this from the beginning, including the steps too small to
  mention.
- What do you type or copy by hand more than once?
- Where does work sit and wait, and who is it waiting on?
- If volume doubled next quarter, what breaks first?

That last one returns more than the rest combined. People work next to the
weakest part of the operation every day and are rarely asked about it.

**Count the exceptions rather than describing them.** When someone says a path
runs "only sometimes", ask how many of the last twenty cases took it. An
exception that turns out to be a fifth of the volume is a main path nobody has
drawn yet.

**Follow up after two days.** People remember the forgotten step, the second
spreadsheet, and the quarterly exception once the interview has settled.

**Gate.** Every in-scope process has at least one performer interview. When the
requester performs the process themselves, their own walkthrough is the performer
interview: record it as such and move on rather than halting for a second
person who does not exist.

**Halt format.** Stop. Name each process covered only by management description.
Offer: schedule the performer interview, proceed and mark that process
`[leadership map only]` throughout the blueprint, or drop it from scope. Wait for
the choice.

## Step 2: Inventory every silo

A silo is any place operational information lives that other systems cannot see.
Catalog all of them: software, spreadsheets, shared drives, inboxes used as
databases, and the messaging channels where decisions actually get made.

Record per entry: what it holds, who writes to it, who reads from it, what it
overlaps with, and its monthly cost.

Cite a proof path for anything claimed to be live: a last-write timestamp, a
seat count, an invoice line, or a recent export. A tool nobody can produce
evidence for is recorded **unknown**, never assumed dead or alive.

**Steps with no system behind them are the finding, not a gap in the
inventory.** A process step that lives only in somebody's memory is
tribal knowledge: record it, name the person the operation stalls without, and
carry it into the ranking as key-person risk.

**Gate.** Every step from Step 1 resolves to one of three: an inventory entry, a
named tribal-knowledge holder, or manual work that depends on no system at all.
Physical and judgment work belongs in the third bucket; recording it as tribal
knowledge invents a risk that is not there.

## Step 3: Quantify the friction in their numbers

Attach a figure to each bottleneck, built from three components. Show the inputs
beside each result so a reader can check the arithmetic.

| Component | Formula |
|---|---|
| Labor | hours per week × loaded hourly cost × 52 |
| Error | cost per occurrence × occurrences per year |
| Throughput | work the team could not take on, priced the way the requester prices it |

From JasonColapietro/suede-creator-skills, MIT.

More from JasonColapietro/suede-creator-skills