Task Design

• Find directions
• Select one Task
• Write the Task Spec
• Review the Task and World Skill with the human
• Build a family

Find directions

Use repository capabilities, supplied traces, user priorities, production failures, existing eval coverage, and project World knowledge. Start with real work and an observable result. Do not create a Task only because an unusual fixture is available.

Propose one grounded Task in the first user-facing design response after inspection. Use any concise structure that shows the work and capability, non-trivial condition, important failure, Environment and access, independent Verifier evidence, difference from existing Tasks, source evidence, and open decisions. Offer alternatives only when a real user choice changes the design.

In the same response, show relevant knowledge already in the project World Skill and the reusable additions or corrections suggested by the Task. If the skill does not exist, show the small initial content that will support this Task and future Tasks. Frequency in traces can support relevance but does not prove business priority. A repository feature can support feasibility but does not prove the Task is valuable.

Select one Task

A strong Task:

• represents useful work;
• requires the selected capability;
• is solvable from visible information or normal discovery;
• contains a realistic reason the agent can fail;
• produces independently observable evidence;
• accepts materially equivalent results;
• checks harmful side effects;
• avoids hidden trivia and answer leakage; and
• differs from existing Tasks by condition, state, evidence, or failure mode.

Reject a Task that mainly tests prompt decoding, brittle formatting, infrastructure luck, impossible knowledge, or a preferred internal tool sequence.

Use normal request language. State the goal, necessary context, expected output or effect, and real limits. Do not prescribe steps unless the procedure is the capability.

Write the Task Spec

Copy assets/task/Task.md.template into the runnable Task directory beside task.toml. Keep it outside the agent-visible image and workspace. Create or update .agents/skills/<project>-world/SKILL.md in the same design loop by following World knowledge.

Purpose and evidence

Name the work, capability, value, source, and difference from existing Tasks. Use paths, symbols, focused tests, trace IDs, policies, or human decisions for important claims.

Agent input

Write the exact initial instruction and any later turns or triggers. State context passed outside the instruction. The generated instruction.md must match either the approved input or the reviewed Draft input used under the documented no-pause exception.

Relevant agent conditions

Include only Harness behavior this Task depends on: tool interfaces, session persistence, time, retries, or reconstruction differences. Do not repeat a complete repository map in every Task.

Environment

Describe initial state, relationships, visible and hidden information, dependencies, access, identity, permissions, time, network, setup, readiness, reset, and material production differences. Cite project World knowledge used by the design.

Verification

For each required or prohibited result, state independent evidence, the exact objective check or bounded semantic question, and its pass condition. List accepted alternatives, the complete pass rule, and invalid-run conditions.

Fairness and leakage

Explain why the Task is solvable, how the agent can find required information, which shortcuts might work, and how hidden truth stays inaccessible. Name a realistic wrong result and a collateral change that must fail.

Review the Task and World Skill with the human

Show the full Task Spec and World Skill changes before building by default. Explain what already existed, what the current Task adds or corrects, and what stays only in Task.md. Let the human revise both. Ask the human to review:

• whether the work and failure matter;
• whether the request is realistic;
• whether the Environment contains enough and only appropriate information;
• whether live, frozen, and simulated choices are acceptable;
• whether the evidence measures the intended outcome;
• whether valid alternatives are allowed;
• whether prohibited effects are complete; and
• whether cost, access, privacy, and safety are acceptable.

Mark Status: Approved only after explicit approval. If the user requested a full build without another pause, continue with a Draft and label it honestly. Treat World Skill changes as accepted only after human review.

After implementation, ask for evidence review. Show representative trajectories, final state, Verifier evidence, defects found, revisions made, results, and remaining limits. A reviewed markdown design does not prove a good runnable Task.

Build a family

Prove one Task before creating a family. Then vary one meaningful condition:

• longer or conflicting history;
• similar identities;
• missing or stale information;
• permission or policy boundary;
• required clarification;
• multiple valid outcomes;
• delayed or concurrent state;
• partial failure or retry; or
• greater collateral-change risk.

Keep the underlying work stable when comparing difficulty. Avoid shallow instruction rewrites. Update project World knowledge when a new condition reveals a reusable rule, builder, evidence source, or known failure.