Prd Create

/prd-create — PRD, Ontology and AERS Authoring Interview

Purpose: Elicit a universe of discourse and a requirements set from a human, seeded from the codebase when one exists, and write a per-PRD folder: the human-readable PRD.md, the machine-facing AERS.md, the ONTOLOGY.md that defines the domain, and the derived UBIQUITOUS_LANGUAGE.md. This is the interactive front of the requirements pipeline; /prd-validate gates what it produces.

Two rubrics own the standards this skill applies. Cite them; never restate them:

• _internal/ontology-readiness — elicitation categories, ### Item states (settled / deferred / unknown), Rules 1–4 and the mandatory core, the ## ONTOLOGY.md format, and the Automated ontology check.
• _internal/aers-readiness — Required Sections, Entry Modes, Interaction Rules, Prioritize Ambiguity by Risk, the Automated readiness check, and the Readiness Assessment template.

When to Use

• An idea, a /goal statement, or a /thesis exists but no requirements artifact does
• A new feature must extend an existing ONTOLOGY.md
• A brownfield repo needs a PRD anchored to what the code already says
• You want the domain modelled before implementation planning starts

When NOT to Use

• A requirements artifact already exists → /prd-validate (it hardens; this skill drafts)
• Only the goal is unclear → /goal first
• Scope is wobbly and there is no thesis → /thesis first (optional)
• Structural refactor with no user-visible behaviour change → /modernize
• You want tickets → /issue-slices, after this
• You want to build → /execute-prd, after this

Arguments

• <description> — plain-language ask (optional; if omitted and no --from, ask).
• --mode greenfield|feature|refresh|rewrite — override auto-detection. Detection is otherwise confirmed with one question; this flag skips that question.
• --extend <path-to-ONTOLOGY.md> — load an existing ontology and interview only the delta. Implies --mode feature unless --mode says otherwise. If omitted in feature or rewrite mode and exactly one docs/prds/*/ONTOLOGY.md exists, propose it; if several exist, ask which.
• --from <file> — seed from a /goal statement, a /thesis artifact, or free notes; classified by content (see Input Modes).
• --out <dir> — output folder. Default docs/prds/<slug>/, slug = kebab-case of the goal title, confirmed in Step 0. Never overwrite an existing folder without showing the diff first.
• --no-scan — skip the Step 1 codebase seed. Refused in refresh (there is no baseline without a scan) and in rewrite unless --extend is given. Recorded in the ONTOLOGY.md header as seeded-from-code: no (--no-scan) and in AERS Repo Starting State.
• --dry-run — run Steps 0–1 only: print the detected mode with its evidence, the resolved output paths, and the seed proposal. Write nothing, ask no interview questions, invoke no sibling skill.
• --full-spec — accepted for one release as the blank-start batched alias absorbed from /prd-validate. Batching applies to Step 4 only.

Input Modes

Blank start

Ask the _internal/aers-readiness Entry Modes prompt verbatim:

"Tell me what you want to achieve in plain language. You do not need to format it yet — I will help turn it into a structured, implementation-ready artifact."

Then run Steps 2–4 one question at a time.

Blank start (batched) — under --full-spec only, Step 4's requirements interview may batch questions within one rubric category. Step 3's ontology interview is never batched: the mandatory core is per entity, and each answer changes the next question.

From /goal or /thesis output (--from)

Classify the file by content, not by filename:

Existing notes or brownfield repo

Project shape is detected in the cwd via _internal/modernization-rubric §1. The codebase is a source, not a substitute for the human: Step 1 proposes, Step 3 confirms.

Workflow

Step 0: Preflight

1. Interactive check. This is an interview. In a non-interactive context (CI, autonomous run, scheduled agent), refuse to start with the same shape as /prd-validate's non-interactive failure mode: say so, and point at _internal/aers-readiness and _internal/ontology-readiness for a deterministic score of an existing artifact instead. Never start a partial interview and fill the rest in.

2. Resolve mode — first match wins; then confirm with one question whose default is the detected mode. --mode skips the question.

   Detect	Evidence
   greenfield	no manifest (_internal/modernization-rubric §1) in cwd and no --extend
   rewrite	project shape detected AND the ask contains rewrite verbs (rewrite, replace, port to, migrate from X to Y, re-platform, v2 from scratch)
   refresh	project shape detected AND the ask contains modernize triggers (modernize, refresh, bring up to date, upgrade the stack)
   feature	project shape detected (brownfield default), or --extend given

   Tie-break refresh versus rewrite with one question: "May the public API or stack change? No → refresh. Yes → rewrite." If refresh turns out to carry no user-visible behaviour change, say so and offer /modernize — link it, do not invoke it.

3. Locate the ontology for --extend. Read its header. If it has no uod: field, halt: an ontology without a UoD boundary cannot license a deferral, so the delta cannot be classified. Say that, and point at /prd-validate.

4. Resolve slug and --out. May share the mode confirmation message when both are defaults. If the folder exists, show the diff before writing.

5. Extended-thinking gate (as /goal Step 2). Before asking anything, reason privately about which rubric categories the inputs already answer, which entities are actually in this release, and where the expensive mistakes are. Produce a per-entity question budget: the 4 mandatory-core categories, plus at most 2 discretionary questions before offering deferral.

Under --dry-run, print the resolved mode with its evidence and the resolved output paths here. In brownfield modes continue to Step 1 and stop after the seed proposal; in greenfield stop here. Either way write nothing and ask nothing.

Step 1: Codebase seed (brownfield only; skipped by --no-scan where permitted)

Compose existing scanners. Never invent a new one.

• /audit-existing → ## Existing State feeds AERS Repo Starting State, Tooling Assumptions and Execution Preflight; ## Duplicated Or Divergent Contracts is the strongest homonym signal; ## Missing Or Partial feeds PRD Current State; ## Test And Verification Gaps and ## Planning Implications feed the AERS Verification Matrix.
• Domain-entity discovery, the ubiquitous-language codebase-scan idea applied to the ontology categories: read references/codebase-seed.md for the full code-signal → ontology-category table (models and exported domain types → entities; PK and unique indexes → reference schemes; FKs and relations → fact types with cardinality; status enums plus transition code → lifecycles with a totality check; NOT NULL and check constraints → alethic candidates; validators and guards → deontic candidates; timestamps, history tables and soft delete → temporality; the same name in two packages → homonym).
• refresh only: _internal/modernization-rubric §1 shape detection (language, type, size class, test signal, patterns) feeds the PRD Current State header and AERS Repo Starting State. Sample-read per its §3 and stop at the 30% context guardrail. Do not run /modernize: it produces a refactor plan, not a PRD.

Emit a seed proposal and print it. Every seeded item enters ONTOLOGY.md as unknown with a code:<file:line> source until a human confirms it in Step 3. Nothing seeded is ever written settled without a human answer — a scan reports what the code does, which is not evidence of what the domain requires.

Under --dry-run, the seed proposal is the last thing printed: stop here, write no files, and do not enter Step 2.

Step 2: UoD boundary and thesis anchoring

Use the thesis (from --from, or stated in the ask) as the UoD boundary test. Otherwise ask for one sentence. If the user cannot give one and scope is wobbly, offer /thesis — link it, do not run it inline.

Propose the boundary as two lists — representable this release, and not representable this release — and distinguish UoD from Scope in the rubric's own words: Scope bounds the work, the UoD bounds representable truth. Ask one confirm-or-amend question.

Brownfield: seeded entities outside this release's ask are proposed as out of the UoD. That is what licenses deferring their fact types.

rewrite only — agree the permitted-change list here, alongside the UoD boundary, with one question:

"Which ontology items may this rewrite revise? Everything not on this list is preserved as-is."

Record the answer as a confirmed closed decision. Step 3 gates every revision against this list, and Step 5 records it as the PRD's What May Change. Agreeing it in Step 2 is what makes the Step 3 gate checkable — the PRD does not exist until Step 5.

Step 3: Ontology interview

Ask in this order: entities and reference schemes → homonyms and synonyms → fact types → constraints → modality → lifecycle totality → temporality. Per entity, the mandatory core first.

• One question at a time. Propose a default, explain why the question matters, challenge ambiguity — "both, probably" is a homonym, not an answer. "You choose" → propose a default and ask for confirmation.
• Explore before asking (/grill-me): if the Step 1 seed already answers a question, show the code:<file:line> evidence and ask only for confirmation.
• Mandatory core cannot be deferred. Refuse with the Rule 2 rationale from _internal/ontology-readiness. The only two exits are settled, or removing the entity from this release's UoD.
• A deferral requires a re-entry condition. Without one it is unknown, and scores per _internal/ontology-readiness § Item states. Say so at the time.
• One elementary predicate per fact type; bold the entity names; split compound statements. Every fact type gets a constraint or the explicit [unconstrained] marker — blank is an omission. Constraint cells use the unique:, mandatory: and value domain: prefixes.
• Every "shall" or "must" is classified alethic or deontic before it is written.
• Say the cost of deferral out loud: only settled rows are consumed downstream by /prd-acceptance and /test-plan.

Under --extend: interview only new entities plus deferred items this feature now touches — exactly the delta declared by scope: and extends:.

Classify every change as an addition (appends freely) or a revision (one of the five kinds in Rule 4). Revision handling is per mode:

• feature — a revision halts. List the stale downstream artifacts (PRD, AERS, data models, tests, code) and write nothing.
• refresh — same rule as feature: a revision halts with the same stale artifact list. Add to the halt message that a change of this class means the ask is really a rewrite, and offer re-running in --mode rewrite. Refresh has no permitted-change list of its own.
• rewrite — a revision is allowed only when the item appears in the permitted-change list agreed in Step 2, and is confirmed there as a closed decision. Log it as revision in the Extension Log, citing that decision. Anything outside the Step 2 list halts. Do not consult the PRD's What May Change: Step 5 writes that section from the Step 2 agreement, so it does not exist yet when this gate runs.
• greenfield — no baseline exists; every change is an addition.

Step 4: Requirements interview (aers-readiness risk order)

Under --full-spec, print once before asking anything: --full-spec is a one-release alias; it will be removed.

Second extended-thinking gate — the /prd-validate Step 2.5 questions: which gaps cause the most expensive mistake, which small-looking ambiguities hide a load-bearing decision, what breaks first if implementation started today. Then ask in Prioritize Ambiguity by Risk order, semantic and structural gaps in one list.

• Stack, runtime and tooling come from the Step 1 audit. Never ask what the audit already answered.
• Draft Functional Requirements as FR-n, with bold ontology terms and a modality tag on each rule.
• Closure check: every noun and verb in the FRs resolves to an ontology term. Unresolved terms go back to Step 3 as additions, not into the PRD as prose.
• Acceptance criteria are - [ ] checkboxes carrying the load-bearing ontology constraints (a uniqueness constraint becomes a duplicate-insert criterion, a mandatory role a null-rejection criterion, a total lifecycle an exhaustive transition criterion) so /prd-acceptance and /test-plan can find them.

Batching is permitted here, within one rubric category, only under --full-spec.

Step 5: Write PRD.md

Write <out>/PRD.md from references/prd-template.md. Sections: header (title, date, status, owner, Mode, thesis, links to the three siblings); Summary; Problem and Outcome; Thesis and UoD Boundary; Users and Actors; Scope (in / out / later); Current State → Target State (brownfield only; delta table | Aspect | Current | Target | Change class |); Functional Requirements; Non-functional Requirements; Acceptance Criteria; Closed Decisions and Open Decisions (product-level only — engineering decisions live in the AERS); Risks and Assumptions; Non-goals; What May Change (rewrite only).

In rewrite mode, What May Change is not elicited here: it records the permitted-change list agreed in Step 2, together with the preserved list that is its complement. Every Extension Log revision written in Step 3 must appear in it.

The PRD carries no Public API, Data Models, Verification Matrix, Repo Starting State, Tooling Assumptions, Execution Preflight or Readiness Assessment. PRD content flows into AERS.md in Step 7, never the reverse.

Step 6: Write ONTOLOGY.md

Write <out>/ONTOLOGY.md in the ## ONTOLOGY.md format owned by _internal/ontology-readiness. Do not reproduce that format here; read it there and follow it exactly, including the header fields (mode, extends, scope, uod, seeded-from-code, thesis, status) and the status summary line.

Mandatory-core rows can only be settled. The Extension Log is created empty in greenfield and appended in every other mode; it is append-only.

Step 7: Translate to AERS.md and score

This skill writes <out>/AERS.md itself, by applying the _internal/aers-readiness transformation to the PRD:

• Map PRD sections onto the Required Sections, one of which is Domain Ontology — a pointer to the sibling ONTOLOGY.md, never an inline copy.
• Generate the sections the PRD does not carry: Public API or Public Interface; Data Models (citing ONTOLOGY.md as the source and naming which entity each structure represents); Verification Matrix; Repo Starting State; Tooling Assumptions; Execution Preflight; Definition of Done.
• Apply both automated checks: the Automated readiness check for the structural score, and the Automated ontology check for the Ontology: line and its capped composite contribution.
• Write the Readiness Assessment in the rubric's template, with both verdict lines, the structural score, the ontology contribution, the composite, the blocking gaps and the recommended follow-ups.

Recommend /prd-validate only when the verdict is not Ready, or when unknown rows remain in the ontology. A Ready artifact with no unknown rows does not need the hardening interview.

Step 8: Derived glossary

Invoke /ubiquitous-language --from-ontology <out>/ONTOLOGY.md, which writes <out>/UBIQUITOUS_LANGUAGE.md. The glossary is a view of the ontology: never hand-write it, never hand-edit it, and never let it disagree with the ontology.

Step 9: Report and handoff

Print this to the console. It is the run report, not the AERS Readiness Assessment section — that one is written into AERS.md in Step 7, in the _internal/aers-readiness template, and is not restated here.

PRD folder: docs/prds/<slug>/
  PRD.md · AERS.md · ONTOLOGY.md · UBIQUITOUS_LANGUAGE.md
Mode: <mode>   Seeded from code: yes @ <sha> | no (--no-scan) | n/a (greenfield)
Readiness: Ready | Partially ready | Not ready   (structural <n>, ontology +<c>, composite <n>)
Ontology: Ready / Partial / Absent   (settled <n> · deferred <n> · unknown <n>; mandatory core complete | INCOMPLETE)
Blocking gaps:
- ...
Next step:
  /prd-validate docs/prds/<slug>/AERS.md   (if not Ready, or unknown rows remain)
  /execute-prd docs/prds/<slug>/AERS.md    (if Ready)
  /issue-slices                            (tickets first)

Per-mode differences

Rules

• Apply the Interaction Rules of both rubrics: one question at a time, multiple choice with a recommended default, explain why it matters, challenge ambiguity, "you choose" gets a proposal and a confirmation.
• Item states are settled, deferred (with a re-entry condition) and unknown, defined once in _internal/ontology-readiness § Item states. Use those words; do not redefine them.
• The mandatory core is Rule 2's four categories. It is never deferred, and a mandatory-core row that is not settled carries the mandatory-core cap defined in _internal/ontology-readiness Rule 2.
• Question budget per entity: 4 mandatory-core, then at most 2 discretionary before offering deferral. Deferral is free; silence is not.
• Every seeded row keeps its code:<file:line> source until a human settles it.
• Ontology describes the world; Data Models describe the representation. Keep them separate.
• Halt conditions surface and stop: --extend target without a uod: header, a revision in feature or refresh mode, a revision outside the Step 2 permitted-change list in rewrite mode.
• Score with the rubrics' automated checks. Do not invent a variant.

CRITICAL: Do Not

• Do NOT fabricate a fact, entity, reference scheme, constraint, modality or temporality to close a gap. unknown is an honest row; an invented one is not.
• Do NOT run from a non-interactive context, and do NOT let another skill auto-invoke this one. /kickoff and /execute-prd never auto-invoke it.
• Do NOT defer a mandatory core item, and do NOT accept any deferral without a re-entry condition — record it as unknown and say so.
• Do NOT write a settled row for anything a human has not answered, including every row the code scan proposed.
• Do NOT hand-write or hand-edit UBIQUITOUS_LANGUAGE.md; it is derived in Step 8.
• Do NOT write a revised ontology when Step 3 classifies a change as a revision — halt, list the stale artifacts, and let the human decide.
• Do NOT overwrite an existing --out folder without showing the diff first.
• Do NOT drift into /execute-prd planning or /modernize refactor work; this skill writes requirements, not plans.
• Do NOT batch questions outside --full-spec, and never batch Step 3.
• Do NOT report Readiness: Ready while a mandatory-core row is missing or a high-risk ambiguity category is unresolved.

Contract

• Inputs: <description> and/or --from <file> (goal, thesis or notes); the repo in cwd; optionally an existing ONTOLOGY.md via --extend <path>. Flags: --mode greenfield|feature|refresh|rewrite, --out <dir>, --no-scan, --dry-run, --full-spec. Embeds _internal/ontology-readiness and _internal/aers-readiness; composes /audit-existing and _internal/modernization-rubric; reads references/codebase-seed.md and references/prd-template.md.
• Preconditions: a human operator is at the keyboard — this is an interview, never a gate and never auto-invoked. Write access to --out (default docs/prds/<slug>/). For --extend, the named ontology exists and its header carries uod:.
• Outputs: the folder docs/prds/<slug>/ containing PRD.md, ONTOLOGY.md, AERS.md and the derived UBIQUITOUS_LANGUAGE.md, plus the Step 9 report with the Readiness: and Ontology: lines and a next step. Under --dry-run: the detected mode with evidence, the resolved paths and the seed proposal, and no files.
• Postconditions: every ontology row carries settled, deferred with a re-entry condition, or unknown; the mandatory core is complete or the report says INCOMPLETE; the Extension Log is append-only; pre-existing settled rows are unchanged; the glossary is derived, not authored.
• Failure modes: non-interactive context → refuse and point at the two rubrics for a deterministic score. --extend target with no uod: header → halt. revision classified in feature or refresh mode → halt with the stale-artifact list, write nothing (refresh additionally offers --mode rewrite). revision in rewrite mode outside the Step 2 permitted-change list → halt. --no-scan in refresh, or in rewrite without --extend → refuse and explain there is no baseline. Existing --out folder → show the diff and ask before overwriting. User asks to skip the mandatory core → refuse with Rule 2's rationale and offer to drop the entity from this release's UoD instead. User wants an existing artifact hardened rather than drafted → hand off to /prd-validate.

Copilot integration

This skill is the durable entrypoint for Copilot hosts that load agent skills. The matching prompt file is an optional VS Code shortcut. Read repository instructions before edits; use the tools actually exposed by the host, including connected GitHub access when available. A prompt or skill does not itself grant additional permissions.