The Prompt Lab — Retrieval Priming
Learn the retrieval priming prompting technique with concrete before/after examples.
Retrieval Priming
The Technique
Retrieval Priming involves explicitly activating the specific domain knowledge, frameworks, or mental models you want the model to draw from before stating your actual request. Because language models generate responses by predicting what follows the context you’ve built, seeding that context with the right conceptual vocabulary steers the model toward higher-quality, more precise outputs — without requiring you to write the answer yourself.
The Naive Prompt
Write a postmortem for a database outage that took down our checkout
service for 47 minutes last Tuesday.Why It Falls Short
This prompt gives the model the topic but no signal about what kind of postmortem matters here — it could produce a shallow incident log, an executive summary, or a blame-focused narrative. Without domain context, the model defaults to a generic template that likely misses the analytical depth engineers actually need. You’ll get structure without substance.
The Improved Prompt
You're working within the Site Reliability Engineering tradition. Key
concepts relevant to this work: blameless postmortems, the "five whys"
root cause method, contributing factors vs. root causes, detection lag,
MTTR, action items with DRI ownership, and the principle that systems
fail — not people.
With those frameworks active, write a postmortem for the following
incident:
- What happened: A PostgreSQL primary node ran out of disk due to
unrotated WAL logs, taking down our checkout service
- Duration: 47 minutes (14:03–14:50 UTC, last Tuesday)
- Detection: Customer support ticket, not internal monitoring
- Resolution: Manual WAL cleanup + failover to replica
- Affected users: ~12,000 attempted checkouts, estimated $34K lost revenue
Structure the postmortem with: Executive Summary, Timeline, Root Cause
Analysis (five whys), Contributing Factors, Detection Gap Analysis, and
Action Items (each with owner role and priority).Why It Works
The opening paragraph doesn’t just set a role — it explicitly loads the conceptual vocabulary the model should reason through. When “detection lag,” “blameless culture,” and “contributing factors vs. root causes” appear before the request, the model treats them as active lenses rather than background noise. The incident details then give it enough specificity to produce analysis that feels written by someone who has actually run an SRE retrospective, not someone who Googled “postmortem template.”
When to Use This
- Specialized professional documents where generic outputs are actively harmful — legal memos, medical summaries, financial analyses, or engineering postmortems like this one. Frontier models like Claude Opus 4.8 and GPT-5.5 have the knowledge; Retrieval Priming ensures the right slice of it activates.
- High-stakes first drafts where you don’t have time for multiple revision rounds. Priming the framework upfront collapses the iteration cycle by making the first output structurally and conceptually sound.
- Domain-crossing tasks where the model might default to a surface-level interpretation — for example, asking a model to analyze a supply chain issue “through the lens of queuing theory and bottleneck economics” before describing the problem pulls the response away from vague recommendations and toward tractable analysis.
Next edition: another technique from the prompting toolkit. Have a before/after you’d like deconstructed? Reply to this newsletter.