Guides

ChatGPT or Claude: paste a spreadsheet or connect a program builder?

Paste your training spreadsheet into ChatGPT or Claude, or connect the AI to Phaserr: what you do each time, what comes back and where the answer goes.

Phaserr team

Share

Connect the AI to your training data in Phaserr instead of pasting a spreadsheet into ChatGPT or Claude, and you type only the question: the AI reads the block, the logged sets and your notes itself, and can save a note on the athlete. It reads a pasted sheet about as accurately: in our test a wrong number turned up in a few answers either way.

Pasted sheet Connected
Before Nothing new Data in Phaserr
Each new chat Paste four tabs Type the question
The AI’s answers 2 of 6 had a wrong number 2 of 6 had a wrong number
Afterwards Paste the answer back Note saved on the athlete

The question we asked

Fictional example. John’s squat max is 120 kg. His four-week block has two sessions a week, squat and bench press on Monday and squat and deadlift on Friday, written as percentages of his 1RM and rounded to 2.5 kg. By Wednesday of week 3 he has logged 40 sets with load, reps and RPE, and the coach has written two notes: one that he started a new job with night shifts and about five hours of sleep, one after Monday’s session that his legs “just had nothing”. On Monday his squat was 5×3 at 102.5 kg. He did 3, 3, 2 and 2 reps and skipped the fifth set.

The coach asks: “John missed reps on Monday’s squats. Look at his squat training in this block so far and tell me what happened. Then tell me what you would change for Friday’s squat session, with loads in kg.” Then: “OK, go with your Friday plan. Update his program and add a note saying why.”

We sent both messages to one OpenAI model, GPT-6.2 Sol, and one Anthropic model, Claude Sonnet 5.5, three times with the spreadsheet pasted in and three times connected to the same data in Phaserr. The last section says how.

What you do each time

Pasted. You copy the data into every new chat: four tabs here (maxes, program, log and notes), 65 rows, headers included. Without it the AI has nothing to read. We asked both models the same question with no sheet, and both declined to say what happened and asked for the log. One wrote: “Anything specific I said about Monday would be invented.”

Connected. The coach’s message was the question and nothing else. The model fetched the block, the logged sets and the coach’s notes itself, in 8 to 13 tool calls per answer, none of which failed.

A pasted sheet is a copy from the moment you paste it. Connected, the AI reads what is in Phaserr when you ask, including the sets the athlete has logged through their link.

What the AI answers

Both ways, all 12 answers traced Monday’s missed reps through the earlier sessions and the coach’s note about night shifts and short sleep, and all 12 lowered Friday’s squat. Eleven also pointed out, unasked, that the deadlift on the Friday before was loaded 5 kg above the plan, 125 kg against 120.

Check the numbers either way. Four of the 12 answers contained a wrong number: two pasted, two connected. One pasted answer gave an RPE range that starts at 6 where the log starts at 6.5; the other gave a program percentage as 79 where the sheet says 80. One connected answer worked out three prescribed loads from the percentages itself and got 84, 96 and 87 kg where Phaserr shows 85, 95 and 87.5. Another cited a training max of 108 kg, which his block does not use, and saved it in the note it wrote. All four of those answers came from the Anthropic model. The OpenAI model stated no number wrongly; its answers were shorter and left more out, such as the program’s percentages. Three runs each way are too few to rank the two.

The advice varied more than the reading. The program has Friday’s squat at 3×4 at 90 kg. The 12 proposals were 3×4 at 85 kg seven times, 3×4 at 80 kg twice, 3×4 at 82.5 kg twice and 2×4 at 85 kg once. Nothing in either workflow tells you which is right. That part stays with the coach.

Where the answer goes

The second message asked the AI to update the program and add a note.

Pasted. All six answers returned the changed Friday row and a note, for the coach to paste into the sheet.

Connected. All six runs saved one note on John in Phaserr, with the Friday plan and the reason for it, and told the coach to change Friday’s squat themselves. Read back from Phaserr, each note matched the plan the model had proposed. The reasoning is saved with the athlete without anyone copying it, and the coach changes the session.

How to move a spreadsheet into Phaserr

Do three things in the app first: add the athlete, set their load rounding (2.5 kg here) and create the exercises you log RPE on, with RPE among their fields. Then paste the sheet into a chat with the connected AI, say what you set up, and ask the AI to use those exercises, move everything, keep the loads as percentages of 1RM and log each set’s RPE.

We ran that four times, twice per model. When we read the data back from Phaserr, every run matched the sheet: the three max values, all 16 program rows as percentages that resolve to the sheet’s kilograms, all 40 logged sets with load, reps and RPE, and both notes word for word. A run took 24 to 36 tool calls, none of which failed, and 55 to 80 seconds from the first call to the last.

In two runs without that setup the sheet still arrived, with RPE as a text note on each set and the loads entered in kilograms.

When a spreadsheet is enough

  • You ask rarely. An occasional question needs only a paste; the connection needs the program in Phaserr first.
  • You want your own columns. A sheet lets you add your own columns, formulas and charts.
  • You only want to skip the pasting. ChatGPT can work on a sheet without pasting, through its Google Drive app or its sidebar in Google Sheets, which can update the sheet.[1][2] We tested neither.

In Phaserr

  • Apps. The setup offers ChatGPT, as a custom connector, and Claude Code, Anthropic’s command-line tool. We connected ChatGPT on a Plus plan with developer mode on. OpenAI’s pricing page lists developer mode on Plus and Pro,[3] and its help page lists custom MCP servers for Business, Enterprise and Edu workspaces, subject to workspace permissions.[4]
  • What a connected app can do. Read athletes, blocks, weeks, logged sets, issues, records, notes, exercises, rep schemes and templates. Create a block, add a note, log sets and set a record. Logging a set again replaces what was recorded.
  • Cost. The builder is free. MCP is €6 a month, or included with the AI coach assistant at €12, and a 3-day trial covers both, which is enough to check that your ChatGPT plan connects.

The MCP page has the setup steps for ChatGPT and Claude Code. The AI coach assistant page has the plans. That assistant works inside Phaserr and proposes changes to the current block for you to apply or discard; it was not part of this test. The workout builder page shows how percent loads are built.

How we tested

We ran the test on 19 November 2026 with ChatGPT and with Claude, each connected to Phaserr over MCP, an open-source standard for connecting AI applications to outside systems.[5] The counts on this page come from a scored set of runs on OpenAI’s GPT-6.2 Sol and Anthropic’s Claude Sonnet 5.5: three per model each way and one per model with no data, each in a new conversation. We checked each first answer in that set against nine facts listed before any run. The spreadsheet was four tables of tab-separated text pasted into the message; we did not open Google Sheets itself. Phaserr held the same data, with rounding at 2.5 kg and our own exercises with RPE among their fields.

Frequently asked questions

Why connect an AI to your training data instead of pasting a spreadsheet into ChatGPT or Claude?

You stop pasting, and the AI's note is saved on the athlete. In our test the pasted version meant copying four tabs, 65 rows, into every new chat. Connected to Phaserr, the coach typed only the question, and the model fetched the block, the logged sets and the coach's notes itself. Asked to act on its advice, each connected run saved a note on the athlete in Phaserr with the plan and the reason; each pasted answer came back as text for the coach to paste into the sheet. Connecting did not change accuracy: on OpenAI's GPT-6.2 Sol and Anthropic's Claude Sonnet 5.5, three runs each way per model, a wrong number turned up in two of the six pasted answers and in two of the six connected ones. Check the numbers either way.

Can ChatGPT read my Google Sheet without copying and pasting?

OpenAI describes two ways: a Google Drive app that lets ChatGPT find and reference Sheets files, and a ChatGPT sidebar inside Google Sheets that can read and update the open spreadsheet. The Drive app's availability depends on your ChatGPT plan, location, workspace settings and Google Workspace policies. We tested neither; our spreadsheet runs used pasted text.

How do I move my spreadsheet into Phaserr?

In the app, add the athlete, set their load rounding and create the exercises you log RPE on. Then paste the sheet into a connected AI chat, say what you set up, and ask it to use those exercises, move everything, keep the loads as percentages of 1RM and log each set's RPE. In four test runs the max values, the plan, all 40 logged sets with their RPE and the notes matched the sheet when read back. In two runs without that setup, RPE was saved as a text note on each set and the loads were entered in kilograms.

Which AI apps can connect to Phaserr?

Phaserr's setup offers two: ChatGPT, as a custom connector, and Claude Code, Anthropic's command-line tool. We connected ChatGPT on a Plus plan with developer mode on. OpenAI's pricing page lists developer mode on Plus and Pro, and its help page lists custom MCP servers for Business, Enterprise and Edu workspaces, subject to workspace permissions. Phaserr's 3-day trial is enough to check your own plan.

What can a connected AI app change in Phaserr?

It can create a new block, add a note, log sets and set a record. Logging a set that is already logged replaces what was recorded for it, and a new 1RM changes the percent-based loads for that exercise. In our test, asked to update the program, every run saved the change as a note on the athlete and left Friday's session for the coach to change in Phaserr.

Sources

  1. OpenAI Help Center. ChatGPT for Excel and Google Sheets.
  2. OpenAI Help Center. Google Drive app and setup in ChatGPT.
  3. OpenAI. ChatGPT pricing.
  4. OpenAI Help Center. Connecting custom MCP servers in ChatGPT.
  5. Model Context Protocol. What is the Model Context Protocol (MCP)?

Make time for the coaching.

Add an athlete and build their first program.