$npx -y skills add golbin/agent-skills --skill prdCreate or update an evidence-backed living Product Requirements Document (PRD) and, when useful, linked context and phase execution files under tasks/. Use for feature planning, implementation scoping, autonomous agent work, or checked-off PRDs that must be validated, reviewed, a
| 1 | # Evidence-Backed Living PRD Generator |
| 2 | |
| 3 | Create compact PRD file(s) that are grounded in the existing system, explicit |
| 4 | enough for a junior developer or fresh AI agent to execute, and structured as a |
| 5 | living source of truth. Work is checked off in the PRD, later phases are revised |
| 6 | as implementation reveals new information, and every phase ends with structured |
| 7 | validation and review. |
| 8 | |
| 9 | Do not implement the feature unless the user explicitly asks. This skill creates |
| 10 | or updates PRD, context, and phase files only. |
| 11 | |
| 12 | ## Core Rules |
| 13 | |
| 14 | - Always create or update markdown files under `tasks/`; do not only answer in |
| 15 | chat. |
| 16 | - Start from evidence. Inspect relevant code, tests, docs, configs, APIs, and |
| 17 | official external references before designing the plan when such context is |
| 18 | available. |
| 19 | - Use **phases** as the execution structure. Use scenarios or user stories only |
| 20 | for product context. |
| 21 | - Keep the master PRD compact. Split large plans into linked context and phase |
| 22 | files. |
| 23 | - Use markdown checkboxes (`- [ ]`, `- [x]`) for discovery, implementation, |
| 24 | validation, reviews, and completion. |
| 25 | - Make every requirement and checklist item concrete, verifiable, and small |
| 26 | enough to become one focused execution task. |
| 27 | - Prefer autonomous progress. Ask only for material unknowns; otherwise make the |
| 28 | best reasonable decision, record the assumption, and continue. |
| 29 | - Preserve completed checkboxes and meaningful user edits when updating existing |
| 30 | PRDs. |
| 31 | - Use the real current date for `Last Updated` and change log entries. |
| 32 | |
| 33 | ## Discovery Policy |
| 34 | |
| 35 | A good PRD starts with a fast but serious discovery pass. Do not invent the plan |
| 36 | from the prompt alone when relevant project context is available. |
| 37 | |
| 38 | Before writing or materially updating a PRD: |
| 39 | |
| 40 | 1. Read related PRDs, tasks, issues, README files, architecture notes, API docs, |
| 41 | product notes, and implementation notes. |
| 42 | 2. Inspect affected code surfaces: routes, screens, components, services, |
| 43 | models, schemas, migrations, jobs, permissions, config, tests, fixtures, |
| 44 | observability, and existing patterns. |
| 45 | 3. Identify current behavior, extension points, naming conventions, data flows, |
| 46 | integration boundaries, dependency/framework versions, and likely blast |
| 47 | radius. |
| 48 | 4. Review available validation paths: static checks, unit tests, integration |
| 49 | tests, API-level E2E, browser/UI checks, app simulator checks, screenshots, |
| 50 | manual smoke checks, and observability checks. |
| 51 | 5. Use official external docs or web sources when third-party APIs, framework |
| 52 | behavior, platform limits, recent rules, or current product behavior affect |
| 53 | the design. |
| 54 | 6. Record what was reviewed, what could not be reviewed, material constraints, |
| 55 | risks, unknowns, and design implications. |
| 56 | 7. Ask the user only if material ambiguity remains after this discovery pass. |
| 57 | |
| 58 | For small PRDs, include a compact `Discovery Summary` in the master PRD. For |
| 59 | larger plans, create: |
| 60 | |
| 61 | - `tasks/prd-[feature-name]/context.md` |
| 62 | |
| 63 | The context file is a source map, not a code dump. Include reviewed inputs, |
| 64 | current system summary, validation surface, existing patterns, constraints, |
| 65 | risks, unknowns, and design implications. |
| 66 | |
| 67 | Every phase must also include a **Phase Discovery Gate** that tells the executor |
| 68 | which files, docs, commands, APIs, tools, and assumptions to re-check before |
| 69 | editing code. |
| 70 | |
| 71 | ## Question Policy |
| 72 | |
| 73 | Ask at most 3 questions, and only when the answer materially changes scope, |
| 74 | architecture, user-visible behavior, data impact, compliance, security, cost, or |
| 75 | an irreversible decision. |
| 76 | |
| 77 | Ask after discovery whenever possible. Do not ask about preferences the agent can |
| 78 | reasonably decide. Use numbered questions with lettered options, for example |
| 79 | `1B, 2D, 3A`. |
| 80 | |
| 81 | ## File Mode Selection |
| 82 | |
| 83 | Choose the layout automatically unless the user explicitly requests one. |
| 84 | |
| 85 | ### Single-File Mode |
| 86 | |
| 87 | Use one file when the feature remains easy to scan: |
| 88 | |
| 89 | - 1-3 short phases |
| 90 | - one primary domain or workflow |
| 91 | - no long phase checklist |
| 92 | - discovery notes fit comfortably in the master PRD |
| 93 | |
| 94 | Path: `tasks/prd-[feature-name].md` |
| 95 | |
| 96 | ### Split-File Mode |
| 97 | |
| 98 | Use split-file mode by default when any condition is true: |
| 99 | |
| 100 | - the plan has **4 or more phases** |
| 101 | - any phase has more than about 10-12 implementation items |
| 102 | - discovery notes would distract from the master PRD |
| 103 | - multiple domains are involved, such as data model, backend, frontend, |
| 104 | migration, permissions, async jobs, observability, rollout, or billing |
| 105 | - the feature is likely to change while being implemented |
| 106 | - the master PRD would become long enough that developers may stop rea |