Which harness for which task

The choice starts with the task. The two circles are model families, Claude and GPT: a product’s place on the map shows which models it works with. Claude Code and Codex are each tied to their own family, Cursor and chat sit in the overlap, and Hermes Agent fully covers Cursor and chat, partly overlaps Claude Code and Codex and reaches beyond both circles, connecting any model, including local ones.

What stays the same

Claude Code and Codex are competing products whose capabilities converge with every release. The gap between them is small, so this article relies on what remains constant.

What a harness is

A model answers with text. The harness is everything around it: skills, tools, environment, the human in the loop. Every AI product has one, from a terminal agent to a browser chat. What differs is the composition. The "Via API" column uses Hermes Agent as an example; there are many open-source harnesses, such as OpenCode, Aider, Cline and Goose, and they are built the same way.

What they see and do

Further reading

Choosing a format

Dots are a score from 1 to 5. For risk, more dots is worse.

Preparing for the meeting

• Mention the note-taking assistantIn the invitation and at the start of the call: the assistant takes notes and a transcript that only the team sees, and answers go into the report without names. With Granola, no audio is kept
• tl;dv or Granolatl;dv is connected to the calendar and joins the Google Meet or Zoom meeting on its own. Granola works without a bot, using your computer audio, and keeps only the transcript and notes
• Airtable row created in advanceName, group and numbers from the table of all people, status "In progress", before the first message
• Guide open on a second screenSeven script blocks and a 30-minute timer
• One host, no note-takingtl;dv or Granola does the transcript, ChatGPT does the analysis

A 30-minute script

Minutes per block

1. 1
   Intro

   "I'm looking at how people in the company actually get their work tasks done. This isn't an evaluation of you. Both successful and unsuccessful cases are equally useful to me. A note-taking assistant is with me: only we see the transcript, and answers go into the report without names. May I turn it on?"
2. 2
   Role and context

   "What do you spend most of your week on?"

   Which tasks repeat · where you look for information and draft things
3. 3
   The last work task

   "Think of the last work task that took you noticeable time. What did you need to get?"

   What means you had · what you chose and why · how you got to the result
4. 4
   Result and effort

   "What did you end up with, and how much effort did it take?"

   What you used as is · what you finished yourself · what constraints there were: data, access, security
5. 5
   A successful case

   "Think of a time a similar task went well. What was different then?"

   Tool · data · preparation
6. 6
   Quitting and switching

   "Have you ever stopped using a tool or switched to another? How did that go?"

   Which tool · what you switched to · why. If this never happened, skip it
7. 7
   Wrap-up

   "What didn't I ask that matters for understanding your work?"

   Thank them · say the answers go into the shared analysis without names

Staying on topic

Logging in Airtable

The table keeps everyone in the sample, their conversations and the report data in one place. Decision: whom else to invite so each group has enough interviews, whom to send a follow-up, and which gaps between self-reported and measured use to discuss with the person.

Voice: the tl;dv or Granola transcript is copied into the Notes field. Text: the chat log is copied into the same field with the date and time of each message.

ChatGPT fills in the orange field from the conversation in Notes. "Per report", "Tokens", "$" and "Activity" come from the ChatGPT Enterprise usage report. The "≠" field is a formula that compares the two frequencies.

Data limits: the report only sees the corporate tool and only for 3 months. Personal accounts and other tools do not show up in it, so a mismatch is a question to clarify, not a conclusion.

Text: wording

In a chat or a questionnaire it is hard to ask again, so a poorly worded question ruins the answer with no chance to fix it. Five common mistakes:

Fill the table with Claude Code

The whole table can be handed to an agent: Claude Code or any other MCP-capable agent connects to Airtable and creates rows, pulls report data and analyzes conversations on its own.

• Create rowsFrom the list of people selected for interviews: name, group, status
• Pull in report dataTokens, $, activity, frequency per report
• Analyze NotesWith the same prompt as in ChatGPT: fields, difficulties, self-reported frequency, quotes
• Update statuses"Partial" (incomplete answer) if the analysis contains "no data"

Questionnaire

The questionnaire is written after the interviews, once you know which tasks and difficulties occur. It goes to the whole company to learn how widespread they are. Questions are short, and the answer options come from the interview analysis.

1. 1
   Which of these tasks do you do at work?choice: tasks from the interviews · other, open fieldtasks
2. 2
   How often do you do these tasks?choice: every day · every week · less often · I don'ttask frequency
3. 3
   Which tools do you use for them?choice: ChatGPT · Claude · Cursor · another AI tool · I don't use AItools
4. 4
   Which difficulties do you run into in these tasks?choice: difficulties from the interviews · I don't run into any · other, open fielddifficulties
5. 5
   How often does this difficulty get in the way?choice: in almost every task · sometimes · rarelydifficulty frequency
6. 6
   How much does it matter for the result?choice: gets in the way a lot · gets in the way · barely mattersimpact
7. 7
   How do you work around this difficulty now?open field · I don'tworkarounds
8. 8
   What else matters that the questions missed?open fieldanything missed

Interview analysis: the analysis prompt and a check against report data

1. 1
   Give the Notes text to ChatGPT

   With the same prompt for every interview
2. 2
   Move ChatGPT's answer into the person's row

   Task, context, chosen way, result, effort, constraints, difficulties, self-reported frequency
3. 3
   Group the table by difficulty

   People with the same difficulty end up next to each other: you see how many there are and from which groups
4. 4
   Check with the questionnaire

   Repeats become answer options in the company-wide questionnaire; the team breakdown is compared with Uniformity

Prompt for interview analysis

Below is a conversation with an employee about their work tasks.
Return strictly by field, only what the person said themselves:
1. Task: the last specific work task
2. Context: for whom and why, what means they had
3. Chosen tool or way, and why
4. Result
5. Effort: time and manual rework
6. Constraints: data, access, security
7. Difficulties: categories from the list {categories from the report's sign table} or "no difficulties"; if none fits, a new category in your own words
8. Self-reported frequency: weekly / regularly / occasionally / doesn't use / unknown
9. 1–2 verbatim quotes
If something is not in the conversation, write "no data" and don't make it up.

{text from Notes}

Further reading

Which decisions the report helps make

Business task: find out for which work tasks the current set of AI tools falls short and which solutions are worth testing before buying. The decision is made by the owner of the AI tools budget together with the leads of the teams where the pilot runs. Data access and security are approved by the information security team. The decision criteria are the signs in the table below.

That the current set is suboptimal somewhere is an assumption: the report and the interviews test it. "Suboptimal" breaks down into testable signs. The conclusion "no new tool is needed" is also a result.

Data limits: the report only sees the corporate tool and only for the export period, 3 months. Personal accounts and other tools do not show up in it.

How answers turn into a purchase decision

The report shows where to look, and the purchase decision is made only after these five steps, so the money goes to a task many people have and nothing covers today.

1. 1
   Interviews

   Which tasks and difficulties exist and how people work around them now
2. 2
   Company-wide questionnaire

   How many people face a difficulty, how often, and how much it gets in the way
3. 3
   Compare

   Task scale, severity of consequences, workarounds, cost of the solution
4. 4
   Test tools

   Promising options on real tasks together with employees
5. 5
   Decision

   The budget owner decides by the six signs above: quality, time and rework, reproducibility, scale, data access, cost; "no new tool is needed" is a valid outcome

Workflow

Five steps from export to a finished list of people.

1. 1
   Export the data

   Five console sections, details in "Where to get the data"
2. 2
   Build the big picture

   Adoption depth, uniformity across teams, main tool, segments
3. 3
   Map the nature of work

   Which tasks people solve, with which built-in tools, and with which integrations
4. 4
   Check against raw data

   Verify each summary against the source export
5. 5
   Select people and draw conclusions

   An interview list and the questions to test in them

   Who to interview How to run the interviews

What a ChatGPT Enterprise admin sees: where to get the data

Everything is built from five sections of a single ChatGPT Enterprise admin console. Hover over i to see exactly what each section contains.

How to export data from the ChatGPT Enterprise console: by hand and with the Claude in Chrome browser agent

Big picture

The next four sections are the report itself, each with its goal. Goal: show which groups of people exist in the company and how large they are.

Nature of work

Goal: show which tasks people solve with the tool, with which built-in features, and through which integrations.

Verification

Goal: check the summaries against the raw exports, the project list and the leaderboard. Any number in the report can be traced back to its source row in an export.

People selection

Goal: name specific people for each way of getting tasks done.

Where to start today

You do not need the whole report at once: the first useful result, a list of most active users, appears after the second item.

• Export users and the leaderboard for three monthsThe leaderboard three times, once per product. A browser agent can do the export

  Claude in Chrome Gemini in Chrome Perplexity Comet
• Calculate main tools and the top by amountThe first list of most active users is ready
• Calculate adoption depthAt least once / regularly / every week, with those who work only through Codex counted separately
• Build a list of people and schedule interviewsHow to pick these people is covered in "Who to interview"

Further reading

If the answer misses

[task description]
Before doing the task, ask me 3 questions whose answers would improve the result the most. Do not start the work until I answer.

Build a final prompt from the task and my answers: role, who and why, data, example, format, question at the end. Output only the prompt text.

Answer review: fix the original request

The “rewrite the request” path as an example: the review found gaps, the original request lost what was wrong and gained what was missing, and everything ran again.

Further reading

The same loop as in Kubernetes

The desired state here is the aggregate dictionary: one entry per concept, requirements marked MUST and SHOULD.

The reconcile loop follows the Kubernetes docs on controllers.

A check sees only what was put into it

The example is the meeting room booking from Specs by aggregate: three MUST requirements and one SHOULD.

Which probes exist and what they check: Kubernetes: probes.

The environment decides whether the loop closes

Read next

Why specs pile up

Numbers are illustrative. How OpenSpec lays out specs and changes: specs/ and changes/ folders, finished changes go to an archive.

RFC 2119 requirement levels

Words and meanings follow RFC 2119. Each word maps to its own kind of check.

The aggregate card

## [booking::Reservation]
Class: aggregate
Description: A meeting room booked for an interval.

Invariants:
  - MUST: active bookings of a room never overlap
  - MUST: held → confirmed → closed, forward only
  - MUST: held unconfirmed for 15 min → closed
  - SHOULD: confirm in the channel it was made in

Commands: HoldRoom, ConfirmReservation,
          CancelReservation

Access:
  CanHoldRoom: [booking::Role::Member]
  CanCancelReservation: Member (own), Admin

Depends on: [[booking::Room]], [[booking::TimeSlot]]

Spec: [[room-booking]]

The example is made up. Terms follow Martin Fowler: DDD Aggregate.

Bounded contexts

The approach is Bounded Context and Ubiquitous Language.

Read next

Four layers of context

Spec: one task before the code

A spec describes the task before work starts: first what and why, then how.

Main steps of GitHub Spec Kit. How to split a large spec into parts: Specs by aggregates

Project bible: what the product is

The bible is a standing description of the product: why it exists, who it serves, how it works and what has been decided. The term comes from practice; the closest documented analogue is Cline’s Memory Bank.

Rules: in every session

Rules are instructions the harness puts at the start of every session.

Skills: a procedure on call

A skill is a folder with a SKILL.md instruction, examples and scripts for a recurring procedure.

The same in a chat: projects and skills

What goes where

Further reading

What decision it supports

The business task comes from the report: for which work tasks the current set of AI tools falls short and what to test before buying. The report also says who makes the decision and by which signs. The sample decides whose tasks we will see in the interviews.

First, define which ways of getting tasks done you need to see: through the corporate AI tool, through other AI tools, by hand or through colleagues, and among people who tried the tool and quit. The ChatGPT Enterprise usage report is one of the sources: it finds the first group and those who quit. Users of other tools are found through a company-wide questionnaire and license and purchase lists.

Why a top-by-activity list is not enough

Three ways to recruit 16 people. Dot color is the usage scenario. The most active people and volunteers are good respondents and belong in any sample, but they are not enough on their own.

Eight groups

Six groups come from report thresholds, "quit" from the last activity date, and "other tools" from the questionnaire. Each group has a question it helps test. The answer can be anything, including "no new tool is needed". The thresholds here are a guide: tune them so each group gets 2–5 people.

Team coverage and replacements

A separate step after recruiting that does not change the groups: how many selected people fall in each team and who replaces those who decline.

What to learn in the interview

The conversation starts from the last work task, not from the tool: what they needed to get, what means they had, what they chose, and how they got to the result. The full script is in the interview article.

Workflow

1. 1
   Collect candidates into eight groups

   Six groups from report thresholds, "quit" from the last activity date, "other tools" from the questionnaire
2. 2
   Keep 2 people per group

   From different teams if you have a choice; one person in one group only
3. 3
   Check team coverage

   If a team has nobody, replace one candidate with a person from it with the same behavior
4. 4
   Schedule 30-minute interviews

   The conversation starts from the last work task: what they needed to get, what means they had, what they chose, and how they got to the result. With those who quit, also discuss why they stopped

Further reading

Human-in-the-loop: the path of one request

Five parts of a prompt

Six examples

You are a support agent at [company].
Below is the full thread with the customer.
<thread>
[paste in full]
</thread>
Write the next reply. Tone: calm and friendly, like this example:
<example>
[one good past reply]
</example>
No more than 5 sentences. Mention dates only if they appear in the thread.

Tomorrow I meet [whom] about [topic], 30 minutes. My goal is [the decision I need].
<materials>
[threads and documents]
</materials>
Prepare:
1. Three questions to settle at the meeting.
2. For each, what the materials already say, with a short quote.
3. What I need to find out before the meeting.
Format: a three-column table.

<thread>
[paste in full]
</thread>
First, copy the exact quotes where decisions are made or tasks are assigned.
Then build a summary from those quotes:
— decisions;
— tasks with owners and dates;
— open questions.
One line per item. If a quote names no owner, write “unassigned”.

Attached is an export of [tickets] for [period], CSV.
Count [tickets] by [topic] and show the 5 biggest changes versus last month.
For each, the number and one line on what may explain it.
Show the code you used so I can check it.

Here are three of our posts that landed well:
<example>[post 1]</example>
<example>[post 2]</example>
<example>[post 3]</example>
Write a post about [topic] in the same style: length, tone and structure as in the examples.
For whom: [audience]. Why: [what we want the reader to do].

Research: [question].
For whom: [audience]. Why: [the decision to make].
Data no older than [year].
Output: a table [columns] and a 5-point conclusion.
A source link for every fact. If there is no data, say so.

What to study next

Start with the 7-minute tutorial: it explains why the techniques in this article work. Then the prompting guides, and the courses when you have time.

Further reading