Taste is not a rubric. If you cannot score five answers, you cannot tell whether a pattern from B1 worked, or whether you just liked the third voice.
Why this matters
Power users iterate until they are tired, then ship the last draft because it is last. Or they ask Grok “which is best?” and accept the model’s favorite child.
Write what “good” means while you can still be honest. Generate more than one answer on purpose. Mark numbers you could defend to a colleague. Change the prompt or the mode because a dimension failed. Grok may format the sheet. Grok may not assign the scores. If the model grades its own homework and you rubber-stamp it, you fail.
Teach
Write the rubric first. Five dimensions is enough. Each one needs an observable pass. Adjectives are not dimensions.
| Weak | Usable |
|---|
| Professional | No slang; one ask; no apology paragraph |
| Accurate | Every number appears in my brief or file; else marked assumed |
| Concise | ≤ 150 words or ≤ 12 lines |
| Usable by the reader | The reader can act without a second document |
| Safe | No names I banned; no legal/medical claim |
Score 0 / 1 / 2 per dimension: 0 = miss, 1 = partial, 2 = hold. Total out of 10. Ties are allowed. “I liked #4” is not a tie-break — pick the dimension that matters for this job and say so.
You publish the rubric in the thread before message 2. If you edit the rubric after you see the prose, you are grading your mood. You may add a note after scoring (“I should have scored tone”) but you do not retroactively change numbers to crown a favorite.
Generate five answers under named conditions. Same job, same facts. What changes is the method, so you learn which method earned the points. A set that works on grok.com:
- A1 — Vague. One sloppy line. This is the planted weak answer. You need a floor.
- A2 — ICC only. Intent, context, constraints. Fast.
- A3 — Role-as-reader + ICC. Fast.
- A4 — ICC + two examples (GOOD/BAD). Fast.
- A5 — ICC + critique loop or Think (one or the other — say which).
Stay in one thread so the facts do not drift. Tell Grok not to look at previous answers when writing the next one, knowing it still might. You are comparing methods, not running a laboratory. If Multi-agent is tempting, skip it — five research packets is not this lesson.
You mark. Grok formats. After all five exist, you copy a quote from each and assign 0–1–2 per dimension. Then:
Do not change my scores. Do not add a winner I did not name.
Format the Evaluation Sheet from my marks.
If Grok “helpfully” boosts A5 because it used Think, overwrite it. The sheet is yours.
Act on the sheet. One change you will keep: “examples beat role,” or “Think fixed the banned name.” Put it in the B1 library or B2 standing brief. Do not use this setup for medical or legal advice you will rely on, or for five Imagine pictures.
Buttons move. The 0–1–2 sheet does not.
Worked example
A coordinator evaluates five versions of a 12-line gate-plan text. Rubric before drafts: time first (2), no invented truck (2), radio line present (2), ≤ 12 lines (2), no inspector name (2).
A1 (vague “help with the gate”) scores 3: pep talk, invented second truck, inspector named. A2 (ICC, Fast) scores 7: times right, still buries the radio line. A3 (dispatcher role) scores 8: radio first, still 14 lines. A4 (GOOD/BAD examples) scores 9: length holds. A5 (ICC + critique) scores 9: same total, wins the tie because “time first” is a 2 and A4 was a 1.
She does not crown A5 because it is last. She writes one line into the project standing brief: “Always include a GOOD six-line example; role is optional.” That is evaluation. A vibe would have been “A5 felt smarter.”
Your turn
Choose a real artifact you would send or post: notice, email, parent note, short SOP, vendor chase, briefing. Not a poem. Not the course millings story unless it is truly your job this week. Prepare the facts (no secrets) and, if you can, one GOOD and one BAD example of the same type.
Open grok.com. One new thread named B7 eval [artifact]. You will send the rubric first, then five numbered generation prompts, then your scores. Do not ask Grok which is best until the sheet exists, and even then it only formats.
Exercise b7-l01-e1
| Field | Value |
|---|
| whereToDo | grok.com |
| time | 24 min |
Copy-ready prompts
Message 1 — rubric only. No draft yet.
I am in Grok Mastery (Russell Tech), Pathway B, lesson B7 — score five answers.
Do not write the artifact yet. Repeat my rubric and wait.
Job: [artifact + reader]
Facts I will reuse for all five answers:
[8–15 lines, no secrets]
Rubric (0 / 1 / 2 each):
D1 [name]: 2 if [observable]
D2 [name]: 2 if [observable]
D3 [name]: 2 if [observable]
D4 [name]: 2 if [observable]
D5 [name]: 2 if [observable]
Confirm: "Rubric locked. Waiting for A1–A5."
Session stamp: today's date and "Rubric locked by Grok in this thread."
Then send A1 through A5 as five messages (or one message with five labeled asks if you are short on time — five messages is cleaner). Fill the brackets.
A1:
A1-VAGUE. Ignore the rubric's quality bar. Answer only: Help me with [job in 5 words].
Write the artifact. Label it A1.
A2:
A2-ICC. Fast. Do not copy A1.
Intent: [ ]
Context: [same facts]
Constraints: [length, format, must, must-not]
Write the artifact. Label it A2.
A3:
A3-ROLE. Fast. Role-as-reader: [who rejects this work].
Same ICC as A2. Do not copy A2. Label it A3.
A4:
A4-EXAMPLES. Fast. Same ICC.
GOOD example: [or "none — invent a GOOD/BAD pair from my constraints, labeled invented"]
BAD example: [ ]
Write the artifact. Label it A4.
A5:
A5-LOOP or THINK. I am using [Think / critique loop on Fast].
Same ICC. Write a draft, score it against the locked rubric, rewrite only fails.
Label the final artifact A5. Show the self-score separately so I can ignore it.
Then you score. Paper, notes app, whatever. Then message final:
Do not change my scores. Do not pick extra winners.
My scores (D1–D5 as 0/1/2, then total /10):
A1:
A2:
A3:
A4:
A5:
Winner I name: [A#] because [one dimension].
Prompt change I will keep: [one line]
Format an Evaluation Sheet:
- Locked rubric (short)
- Table: Answer | condition | D1–D5 | total | one quote I said to use
Quotes: [paste one short quote from each A1–A5]
- Winner I named:
- Floor (A1 total):
- What beat vague, in one line:
- Prompt change I will keep:
- Session stamp: today's date and "Evaluation sheet formatted by Grok in this thread from my scores."
Artifact to paste
- The locked rubric and first stamp (proves it existed before drafts).
- The five labeled artifacts (or tight excerpts plus the first two lines of each, if long).
- The Evaluation Sheet with your numbers and the final stamp.
- One sentence from you: which method you will default to for this artifact type, and which you will stop using.
Rubric (for this exercise)
- Pass: Rubric lock stamp is earlier in the thread than A1. Five answers exist under the five conditions. Scores are yours (Grok did not invent the 0–1–2). A1 is not the winner unless you have a specific, written reason. Prompt-change line is specific. Job is yours.
- Fail: Rubric written after you saw drafts. Only two answers. You asked Grok to score and submitted that. All 10s. No vague floor. Offline table. Secrets in the facts.
Common failure
The common failure is grading the vibe — A5 used Think, so it “must” win. If A4 already had a 2 on that dimension, Think bought nothing. Do not add a sixth dimension after you see the prose. Do not skip A1 because it feels rude. If ICC does not beat a five-word shrug, ICC is not written.
What’s next
You have patterns, a project home, files, memory rules, research, small code, and a way to score. The capstone is a library of eight workflows with a before and an after — work you can run on Monday without rereading the pathway.
