Requirements-as-Code (Authoring + Scoring)

Looking for the full Requirements-as-Code guide?

This page documents the original authoring + scoring bridge (draft_requirements, score_output). The modern workflow — product(), technical(), trace(), project(), validate(), and analyze() — has its own comprehensive section: Requirements-as-Code →. New projects should start there.

The mycontext.rac package helps you author and score agent requirements. It drafts the sections of a requirements.yaml using the open-source cognitive patterns and scores candidate outputs with the evaluation stack. The emitted requirements.yaml is then handed to your own spec compile / CI / human review for enforcement.

Authoring and scoring only — by design

mycontext does not ship a requirements compiler, a CI gate executor, or a human-in-the-loop / budget runtime. It authors and scores; your stack enforces.

How sections map to patterns

Each requirements.yaml section is drafted by the patterns best suited to it:

Section	Drafting patterns
task_taxonomy	question_analyzer, intent_recognizer, stakeholder_mapper
rubrics	code_reviewer, comparative_analyzer, socratic_questioner
action_risk_matrix	risk_assessor, impact_assessor
pre_mortem	future_scenario_planner, conflict_resolver

Draft a requirements.yaml (offline)

By default, authoring is fully offline — each section emits the prompt you can run anywhere:

from mycontext.rac import draft_requirements, to_yaml

doc = draft_requirements(
    "Build a support copilot that drafts replies to billing tickets."
)
print(to_yaml(doc))

The emitted document carries an explicit anti-goal note in its metadata:

version: 1
task: Build a support copilot that drafts replies to billing tickets.
metadata:
  generated_by: mycontext-ai 0.12.0
  mode: authoring
  note: >-
    Authored and scored by mycontext (authoring + scoring only). Enforcement ...
    is the responsibility of your own stack.
task_taxonomy:
  templates: [question_analyzer, intent_recognizer, stakeholder_mapper]
  drafts:
    - template: question_analyzer
      prompt: "ROLE: ..."

Fill the drafts with an LLM (your key)

Pass execute=True to run each section's pattern and fill in the drafted content using your own API key (still no mycontext server, still your cost only):

doc = draft_requirements(task, provider="openai", execute=True)

Score candidate outputs

Score a candidate rubric answer or agent output against the context that produced it. Offline and deterministic by default:

from mycontext.rac import score_output

result = score_output(
    "Identify the root causes of the outage and propose fixes.",
    "Root cause: a misconfigured timeout. Fix: add validation. Step 1: ...",
)
print(result["overall"])      # e.g. 0.78
print(result["dimensions"])   # per-dimension scores

API

Symbol	Purpose
RequirementsAuthor(provider, execute)	Configurable author; .draft(task) returns the requirements dict
draft_requirements(task, provider, execute)	One-shot drafting helper
to_yaml(requirements)	Serialize a drafted structure to YAML
score_output(context_prompt, output, mode)	Score an output against its context