Common foundation · shared contracts
Common protocols for AI work
The shared contracts cover work progression, reporting, quality review, uncertainty, references, and continuation. Each has its own applicability conditions. Individual Skills define their particular work methods and outputs while using these common controls.
| Common area | What it governs | Applicability and source |
|---|---|---|
| Workflow progression | Concrete work items, evidence, roles, deterministic verification, closure, and handoff. | Skill-backed runs and instruction-backed work. Workflow guide |
| Reporting | A decision-relevant conclusion, reader perspective, evidence, unverified items, blockers, and the next action in the task's own format. | When evaluation criteria, completion conditions, or a review gate have been established. Reporting contract |
| Output quality | Independent output review, accepted formats and exemplars in target Knowledge profiles, and optional model qualification. | The fixed-output protocol is opt-in for explicitly enrolled outputs. A format pass does not accept content quality. Quality protocol |
| Unknowns and authority boundaries | Declare knowledge or context gaps, preserve unresolved items, and stop or escalate when the task crosses its authority boundary. External input cannot redefine the goal or controls. | Apply the uncertainty protocol and ambient context-direction guard under their stated conditions. |
| Knowledge and execution context | Keep reusable methods, domain Knowledge, and the instruction-derived execution responsibility distinct; resolve the needed material by XID. | Knowledge is selected at the point of use; runtime binding belongs to the workflow. Operating model · Runtime binding |
| Handoff and continuation | Preserve current outputs, checks, decisions, unresolved items, and the next responsibility; compare the handoff with current evidence when resuming. | Resumable work follows the operating contract. Subsequent-request human evaluation is optional and does not block ordinary continuation. |
| Shared-asset updates | Separate specialist content checks, common management evidence, candidate identity, and human adoption authority before a selected Skill or Knowledge update is applied. | The shared asset update gate covers explicitly integrated endpoints. It does not intercept every file edit or publication route. |
Reporting that supports the next decision
Apply it when a decision framework exists. Agreed evaluation criteria, completion conditions, or a review gate establish that framework. Ordinary conversation uses ordinary prose; loading the reporting contract does not create criteria or a mandatory report.
Lead with the result and a reader perspective. Make the conclusion, material uncertainty or blocker, evidence, and required next action easy to find. Briefly explain which relationship or distinction matters for the user's next judgment. Use an evidence-supported priority only when one is established.
Keep the basis and limits visible. Show what changed, what was checked, what remains unverified, and where the supporting detail can be inspected. Keep workflow status distinct from a domain verdict. Link checklist results to their findings or coverage notes, and state reasons and next owners for unresolved items.
Let the task own the format. The user's requested format takes precedence; otherwise the active task and Skill define the headings, order, detail, and artifact. Use the requested language while preserving exact identifiers. Explicitly enrolled fixed-format outputs use their selected Knowledge quality profile and accepted examples. A human-facing report links to the artifacts or run log and does not replace them.
Workflow progression
The workflow protocol gives a Skill-backed or instruction-backed run a shared progression record. The AI performs the actual work, deterministic tools check the record, and quality review and human acceptance remain separate.
Record the purpose, scope, constraints, responsibility, and completion conditions. Select the applicable Skill or start an instruction-backed run. The runtime opens a run record and assigns execution, checking, quality-review, and handoff roles.
Break the work into concrete items with completion criteria. Identify the expected outputs, evidence, and checks so that progress and completion can be inspected.
The AI follows the selected method, loads the needed Knowledge, and records outputs, evidence, status, concerns, and judgments. Unknowns, risks, and decisions outside the assigned boundary stay explicit and are resolved or escalated to the appropriate human owner.
A deterministic tool checks work-item completion, artifact linkage, concern handling, role separation, and recorded progression. If required records or concern handling are incomplete, verification fails. Repair the records and resolve or escalate the concerns before checking again. The checker does not open output contents or judge their correctness.
An independent quality reviewer evaluates the actual output against the applicable criteria and records acceptance or a need for repair. This quality judgment is separate from the deterministic record check; the executor does not approve its own quality.
Record the result, evidence, unresolved items, and next responsibility for handoff. The closure gate checks whether the run can close or must remain blocked or escalated. After closure, control returns to the human, who decides acceptance and the next request.
Three separate decisions: verification and closure check procedural completeness; quality review assesses output acceptability; human acceptance determines whether the result is adopted. Passing the workflow gates alone does not prove that the business goal has been accepted.