Prompts That Route Reliably
Structuring asks so Glean answers consistently
You'll be able to
- Structure a prompt so it consistently routes to the intended answer
- Measure consistency by re-running a prompt and comparing the responses
Read first
Day 7 Lab
- Write a deliberately vague prompt and run it three times.
- Note how much the answers vary in scope, format, and sourcing.
- Rewrite the prompt with explicit scope, an explicit output format, and a named source.
- Re-run the rewrite three times and compare the variance to the vague version.
- Keep the structured version as your template for this kind of ask.
Working Example: A Vague Ask vs. a Routing Ask
A reliable prompt names its scope, its format, and its source, so the request routes the same way on every run. Consistency is something you can measure by re-running. Verified against assistant_skills.
Vague (routes differently each run):
"What's the status of the migration?"
Structured (routes the same way every run):
"Summarize the database migration status from the #db-migration channel and
the migration runbook in Drive. Give me 3 bullets: done, in-progress, blocked."
The three levers that make it route reliably:
SCOPE → "the database migration" (not just "the migration")
SOURCE → "#db-migration channel and the runbook in Drive"
FORMAT → "3 bullets: done, in-progress, blocked"
Measure it: run the structured prompt 3x; the SHAPE of the answer should match
each time. Drift in shape means a lever is still missing.Check for understanding
Take a vague prompt and rewrite it with scope, format, and source so it routes the same way every time.
Check yourself
1. When stacking operators on a 500-result query, how do you know which constraint did the real work?
2. You're exploring an unfamiliar result set and don't yet know which apps have hits. What's the better tool?
3. A trusted document ranks at #8. What's the honest way to investigate the ranking?
4. What three things must a written search recipe include so a teammate reproduces your exact result set?
5. Which question best justifies Glean's multi-step thinking mode over a quick answer?
In 10 seconds
“Consistent answers come from consistent asks. State the scope, the format, and the source in the prompt — then re-run it to measure that you get the same shape back.”
Author a Personal Skill