Skip to main content

Task Contract (L0)

TaskContract is Layer 0 metadata: it describes what kind of task the model is doing before the nine main sections (Role, Goal, Rules, …). It calibrates vocabulary, tone, output shape, and evidential rules in one place.

Import​

from mycontext import TaskContract, Context, Guidance, Directive
# or
from mycontext.foundation import TaskContract

Fields​

All fields are optional. Only populated fields appear in the assembled L0 table.

FieldTypeDescription
domainstr | NoneSubject area and scenario (anchors vocabulary).
audiencestr | NoneWho reads the output (formality, assumed knowledge).
genrestr | NoneOutput genre — e.g. internal brief, blog post, tutorial, api reference. Should align with how you want the model to answer (markdown vs JSON).
groundingstr | NoneEvidential standard — e.g. trace claims to a provided packet; use a literal fallback token when evidence is missing.
metaphorstr | NoneWhether analogical language is allowed.

Helpers​

  • has_content() — True if any field is set.
  • to_dict() — only non-null fields as a plain dict.
  • TaskContract.from_dict(d) — build from a dict (e.g. legacy metadata["l0"]).

With Context and research flow​

When research_flow=True, a populated task_contract is rendered as L0 — TASK CONTRACT at the very top of the assembled prompt (before Role), as a small markdown table.

from mycontext import Context, Guidance, Directive, TaskContract

ctx = Context(
    task_contract=TaskContract(
        domain="Software engineering — service observability",
        audience="Staff engineers",
        genre="internal brief",
        grounding="Every claim must trace to the MATERIAL PACKET. If absent, write [NOT IN PACKET].",
    ),
    guidance=Guidance(
        role="You are a senior technical writer",
        goal="Produce an evidence-based decision brief",
        rules=["Use ## headings as specified in the user message"],
    ),
    directive=Directive(content="Follow the user’s structure and tables."),
    research_flow=True,
)

print(ctx.assemble()[:500])

Backward compatibility​

If you do not set task_contract but put an L0 dict on metadata, the assembler still works:

Context(
    metadata={
        "l0": {
            "domain": "...",
            "audience": "...",
            "genre": "internal brief",
            "grounding": "...",
        }
    },
    research_flow=True,
    ...
)

Prefer task_contract=TaskContract(...) for type safety and clarity.

With PromptArchitect​

Pass the same contract (and ideally the real user_message) into improve() or build() so the rewriter can align output_contract with the user’s requested format:

from mycontext import TaskContract
from mycontext.intelligence import PromptArchitect

tc = TaskContract(
    genre="internal brief",
    audience="Engineering leadership",
    grounding="Use [NOT IN PACKET] when the source lacks data.",
)

arch = PromptArchitect(model="gpt-4o-mini")
result = arch.improve(
    "You are a writer. Draft an internal brief from user materials.",
    user_message=user_message_text,
    task_contract=tc,
)

See PromptArchitect for full API details.

See also​