Chapter 14: Lumen Writing Better Prompts
The quality of a Lumen report is determined by the quality of your prompt. This is not a metaphor. Lumen is not guessing your context — it reads what you give it and uses it. A vague prompt produces a vague report. A precise prompt produces a precise one.
This chapter covers what goes into a high-quality prompt for each workflow type, with examples of weak and strong versions.
Let's begin by understanding what makes a good prompt. Every good Lumen prompt has five elements.
1. Product identity — What the product is, in one to two sentences. Business model, stage, and motion (PLG, sales-assist, marketplace, etc.).
2. Segment definition — Specific, named segments. Not "small and large users." Named tiers with pricing and a brief behavioral description.
3. Available data — What MCPs are connected and what data exists. If you have 14 customer interviews, say 14. If PostHog has 90 days of data, say 90 days. Lumen calibrates evidence quality to what you tell it.
4. Key question — The single most important thing you need to know from this run. Not a list of questions. One focused question shapes the entire report.
5. Constraints — What you cannot do. Pricing changes before a certain date. Engineering capacity limits. Board-level commitments. Regulatory restrictions. Constraints go directly into DecideWell's option evaluation and RoadMap's capacity planning.
Before and After: W1 PMF Discovery
Weak prompt:
/lumen:pmf-discovery
We make a project management tool. Check if we have PMF.
This gives Lumen almost nothing. The segment is undefined. The business model is unspecified. There is no indication of what data is available. The report will be LOW evidence across every section.
Strong prompt:
/lumen:pmf-discovery
Product: Helix — AI-native PM platform for B2B SaaS teams
Segments: Solo (free, individual PMs), Team ($49/seat, product teams 3–15),
Org (enterprise, custom pricing, 50+ users)
Stage: 2,400 MAU, 18 months since launch, PLG motion with sales-assist emerging at Org
PostHog: connected (90-day cohort, 23 validated events)
Interviews: 14 customer interviews available (11 Team plan, 3 Org plan)
Key question: Team plan D30 retention dropped from 72% to 61% over 8 weeks.
Is this a PMF regression, a product quality issue, or both?
Constraints:
- Cannot deprecate free Solo tier before Q3
- No engineering capacity for features requiring > 4 weeks to build in Q2
This gives Lumen everything it needs. The segment definitions are specific. The data availability is stated. The key question is focused. The constraints are explicit.
Before and After: W3 Strategy
Weak prompt:
/lumen:strategy
Help us plan Q3. We want to grow revenue and improve the product.
Every company wants to grow revenue and improve the product. This is not a strategy prompt. It is a mission statement.
Strong prompt:
/lumen:strategy
Product: Helix — B2B SaaS, PLG + sales-assist emerging
Quarter: Q3 2026 (April 1 – June 30)
PMF status: Team 58/100 · Org 43/100 · Solo 29/100 (from W1 run, March 10)
Current NRR: 100% (recovering, was 89% eight weeks ago)
North Star debate: Board wants MAU. We believe Weekly Active Teams
is more predictive of retention and expansion. Need to resolve.
Engineering capacity: 6 engineers, 2 sprints/month, ~78 sprint-points for product work
Known commitments: SOC 2 Type I audit (May 15), 3 customer contract clauses
consuming ~14 sprint-points
Board constraint: $2M ARR target by September 30
Key question: Given the recovery from Q2 churn, should we stay focused on Team plan
depth or invest in Org plan breadth? We cannot do both.
Prior learnings:
- Roadmap-to-sprint handoff is #1 Team plan pain
- Org plan champions need Slack + Jira integration to unlock enterprise rollout
- Solo plan churn to free Notion templates is accelerating
The board constraint, capacity numbers, prior learnings, and the forced trade-off ("we cannot do both") all go directly into DecideWell's structured decision evaluation.
Before and After: W4 Feature Validation
Weak prompt:
/lumen:feature
Should we build AI into the product?
"AI" is not a feature. This produces a generic analysis with no actionable output.
Strong prompt:
/lumen:feature
Feature: "AI Sprint Advisor" — reads the team's sprint backlog and OKR data
and suggests reprioritization to improve OKR alignment.
Suggestions are advisory only; not auto-applied.
Uses AI/ML: Yes — LLM reading internal backlog + OKR data
Target users: Team plan users with 3–10 engineers
Current evidence: W1 identified roadmap-to-sprint handoff as #1 opportunity.
No behavioral data yet on how teams structure sprint backlog in Helix.
Constraint: 4 engineering weeks available if we decide to build
Key question: Build now, run fake-door test first, or buy a third-party API?
Flag any ethics or compliance requirements.
The AI/ML flag is critical. Without it, DataLayer does not run. Without DataLayer, you ship an AI feature without an ethics review.
Before and After: W6 Churn Analysis
Weak prompt:
/lumen:churn
We're losing customers. What should we do?
Strong prompt:
/lumen:churn
Product: Helix, B2B SaaS
Segments: Team plan and Org plan
Current NRR: 94%, recovering from 89% six weeks ago
Connected: PostHog, Stripe, HubSpot
What changed: Roadmap editor regression in March. Fix shipped 3 weeks ago.
Known churn type: Product churn (from prior W2 run)
Previously tried: Personal outreach to top 3 at-risk accounts, 2-email re-engagement
sequence — mixed results
Specific concern: 3 Org plan accounts renewing in the next 45 days.
Synthex Labs ($28,400/yr) renews in 8 days. Health score: 4.1.
Key question: Is NRR stabilizing? What is the fastest path to getting above 100%?
Including "previously tried" is essential. DecideWell explicitly excludes interventions that have already been attempted. Without this context, you may receive a recommendation for the email campaign you just ran.
Universal Prompt Tips
One question is better than five. Each prompt should have a single key question. If you have five questions, run five prompts. A focused question produces a focused report. A list of questions produces a list of answers, none of them deep.
State what data is not available. If you do not have customer interviews, say so. If PostHog data is only 30 days old, say so. Lumen calibrates evidence quality to stated data availability. Silence is treated as data absence.
Name your competitors. MarketIQ does better competitive analysis when you name the competitors you care about. "We compete with Linear, Productboard, and Notion in the enterprise segment" gives MarketIQ specific companies to assess.
Include the number, not the adjective. "Significant churn increase" is vague. "MoM churn jumped from 3.1% to 7.4%" is precise. Use the actual numbers wherever you have them.
State constraints as facts, not preferences. "We prefer not to change pricing" gives Lumen room to recommend a price change anyway. "We cannot change pricing for existing customers before July 1" closes the option. Constraints stated as facts go directly into DecideWell's option evaluation.