Spec-driven development with existing task packets
Map Spec Kit and optional OpenSpec authoring to one packet authority, observable acceptance, and preserved evidence.
Цю сторінку ще не перекладено. Ви читаєте англійський оригінал.
Keep one authority for the task
Spec-driven development starts by describing the intended behavior and how a reviewer will recognize it. In this foundation, GitHub Spec Kit is the selected method to bring that discipline into existing task packets. OpenSpec is an optional small-change authoring pilot. Neither becomes a second scheduler, permission system, or mandatory approval chain. A generated plan does not widen the authorized paths or turn a deferred integration into an active one.
The local packet contract already records an identifier, task class, level, skills, dated inputs, outcome, acceptance, stop conditions, budget, allowed paths, and forbidden work. Retain that record as the authority. Use specification documents to explain behavior and implementation choices in enough detail to make the packet reviewable. This guide proposes a mapping, not a report that either CLI was installed or run.
Map artifacts to questions a reviewer asks
| Specification purpose | Existing packet destination | Reviewer question |
|---|---|---|
| Problem and intended behavior | Outcome and dated inputs | What changes for whom, and why? |
| Assumptions and boundaries | Forbidden work and stop conditions | What must remain outside this phase? |
| Implementation plan | Task class, skills, and exact allowed paths | Can this worker perform this work here? |
| Work breakdown | Bounded steps inside this phase | Is each step necessary for the outcome? |
| Scenarios and checks | Acceptance plus evidence artifacts | What result proves or disproves the behavior? |
| Clarifications and unresolved decisions | Explicit open questions | Which uncertainty blocks which action? |
Keep the packet identifier on every supporting document. If an authoring tool creates additional files, give those files an approved location before generation. In a shared worktree, file existence is not ownership: preserve other writers' modifications, capture a status baseline, and use non-overlapping edit paths. Review discrepancies between the specification and packet before work proceeds.
Worked example: a catalog search change
Suppose a fictional local catalog viewer should distinguish a missing schema from an empty one. The behavior specification says: when a source card supplies only a field count, display “schema not supplied.” Do not generate parameter names, and do not expose a Run button. When a real schema is present in a later authorized fixture, preserve its provenance and validate it under a separately defined rule. These are proposed behaviors for an exercise, not claims about the present product.
The packet outcome becomes “a reviewer can see that numeric field counts do not establish callable parameters.” Inputs name a small synthetic fixture and its revision. Allowed work is limited to that fixture and an isolated renderer draft. Forbidden work includes credentials, live connectors, package installation, and existing catalog records. A stop condition names unexpected changes by another writer. Acceptance includes one card with a count, one with no count, and one malformed count. Each must produce an explicit, understandable result without suggesting executable readiness.
Plan evidence before implementation
Write expected results beside the cases before changing the draft. Record the exact command or manual procedure needed to inspect them. Save an observed result, timestamp, and artifact reference after running a check. “Tests planned,” “tests run,” and “tests passed” describe different states. A screenshot can show presentation; it cannot by itself prove database integrity, account permissions, or absence of side effects.
If a check fails, preserve the result and identify the violated criterion. If a needed permission is missing, continue only with independent authorized work. Do not broaden scope merely to make a demonstration look complete. At handoff, explain which behavior is implemented, which is documented, and which remains untested.
Choose the authoring method conservatively
Use Spec Kit's specification, planning, and task decomposition approach where it improves the existing packet. Treat concrete commands and generated layouts as version-dependent details to verify before adoption. Use OpenSpec optionally for a small change description and review its telemetry policy before any pilot. The supplied review establishes CLI/file workflows for these projects; transcript assertions of REST or official MCP interfaces are unsupported. Do not publish those assertions as integrations.
Exercise and acceptance
Take the fictional catalog change and create one packet plus a short behavior specification. Include the three cases, a phase boundary, source dates, an exact file allowlist, and an unresolved-question list. Ask a peer to introduce a contradictory requirement: “generate missing parameter names automatically.” Acceptance: you can show why that request conflicts with the outcome and evidence limits, preserve the original packet, and record the proposed change for review. You should finish with one authoritative contract and traceable supporting artifacts, not competing task queues.
Continue with the platform architecture guide for infrastructure boundaries and the integration directory for readiness terms. The learning roadmap provides prerequisite exercises. Existing knowledge practice and adoption paths pages offer native navigation; their current rendering was not verified here.
Sources and limits
| Evidence | Observed | What it supports |
|---|---|---|
| GitHub Spec Kit, supplied primary-source review | 2026-10-30 | Selected MIT CLI/file method |
| OpenSpec, supplied primary-source review | 2026-10-30 | Optional MIT authoring pilot; telemetry needs policy |
| Local prompt contract and task packet | 2026-10-30 | Existing authority fields and one-phase scope |
This lane made no live documentation requests or CLI trials. An authenticated client connection is not verified. The mapping and exercise are original recommendations, not vendor compatibility certification. Generated files, actual command behavior, telemetry, and fit with a particular team's review process still require bounded evaluation.
Джерела
Нижче наведено джерела та дати спостережень. Статус перевірки зазначено для кожного матеріалу.
🔗 Зв'язки статті
Відкрити повний граф знань ↗Зв’язки опублікованих статей. Виберіть вузол, щоб відкрити статтю біля заголовка.
Зв’язки статей · клавіатура / текст
Наступні кроки
Пов'язані матеріали
Data platform process and provenance across products
Plan traceable records across Tvijo, Hotido, FinEdg, and BibaMoney using synthetic contracts, explicit rights, and reviewable evidence.
AI and process foundationAn independent AI and automation learning roadmap
Build task framing, source checking, evaluation, and automation skills through portable exercises using synthetic data.
💬 Коментарі