Architecture decision

Shorten a design decision while retaining its rationale and tradeoffs.

Condense a supplied architecture decision record for someone who will maintain or revisit the choice. Select this for a durable technical decision, not a general executive recommendation.

State the decision and its status first: proposed, accepted, or superseded, as supplied. Keep the problem, constraints, and alternatives that explain the choice. Reduce each alternative to the decisive benefit or cost. Preserve consequences, reversibility, and conditions that would justify reconsideration.

Remove meeting chronology, repeated background, and implementation detail that does not affect the decision. Separate agreed facts from estimates and untested assumptions. Do not turn an accepted tradeoff into a solved problem or a proposed design into an approved one.

Output brief sections for Status and decision, Why, Tradeoffs, and Revisit when supported. Keep system names and numeric constraints exact. Do not create new acceptance criteria, select a different design, or add actions that the record does not authorize.

Example

Search architecture record

Before 99 words

This decision is accepted. We will keep search in PostgreSQL using full-text indexes for the first release. The corpus is currently 80,000 documents, and the team has two engineers. We considered a separate search service because it offers typo tolerance, but operating another service would exceed our current on-call capacity. PostgreSQL avoids that operational burden, though the first release will not support typo tolerance. The latency target is 300 ms at the 95th percentile; it has not been validated under production load. We will revisit this choice if tests miss that target or the corpus grows beyond 500,000 documents.

After 59 words

Status and decision: Accepted — use PostgreSQL full-text indexes for release 1. Why: 80,000 documents and two engineers. A separate search service offers typo tolerance but exceeds current on-call capacity. Tradeoffs: No typo tolerance in release 1. The 300 ms p95 latency target is unvalidated under production load. Revisit: Tests miss the target or the corpus exceeds 500,000 documents.

Examples illustrate the method. They do not measure model output.

Files and sources

prompt.txt Complete prompt as plain text.
SKILL.md The same prompt with a skill header.
llms.txt (JSON) Prompt, metadata, example, and sources.
LICENSE MIT license text.

An original sho.rten.it skill. Source notes.