Draft vs. published lifecycle for specs: mode model, branch-aware mode inference, publish transition, post-publish numbering stability, re-draft window, cross-reference integrity, status surfacing, malformed-front-matter handling, and the publish-time changelog rule. Single source of truth for the V1–V18 normative rules. Triggers on: draft mode, published, version bump, publish, re-draft, renumber FR, branch detection, status field, version field, pre-merge reminder, CHANGELOG.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Draft vs. published lifecycle for specs: mode model, branch-aware mode inference, publish transition, post-publish numbering stability, re-draft window, cross-reference integrity, status surfacing, malformed-front-matter handling, and the publish-time changelog rule. Single source of truth for the V1–V18 normative rules. Triggers on: draft mode, published, version bump, publish, re-draft, renumber FR, branch detection, status field, version field, pre-merge reminder, CHANGELOG.
Versioning Discipline — Draft vs. Published Lifecycle
This skill is loaded unconditionally by @spec-author,
@prd-drafter, and @prd-critic, and conditionally by
@context-detective (V5 branch probe only). It defines the 18
normative rules (V1–V18) that govern a spec's lifecycle from
initial draft through publish, re-draft, and re-publish.
It is the single source of truth for these rules. prd-evolution
delegates the lifecycle / mode model to this skill and retains only
post-publish update mechanics (RFC headers, ADR deprecation,
naming aliases). On any conflict between this skill and
prd-evolution, this skill wins.
When to Use This Skill
Load whenever a spec is being read, edited, scored, or status-checked.
The mode model (V1) governs every spec at every moment; the loaders
above therefore load it always. The detective loads it only for the
narrow V5 branch-probe contract.
The 18 normative rules
V1 — Mode model
Every spec is, at any moment, in exactly one of two modes:
draft — the spec is still being shaped; numbering is fluid.
published — the spec is a contract; numbering is frozen.
Mode is encoded in the spec front-matter as Status: draft or
Status: published. The presence or absence of the -draft suffix
on the Version: field MUST agree with Status:. Mismatch is a
critic-blocking error (D7 sub-rubric d7.status-version-consistency).
V2 — Initial state (creation mode)
A newly created spec starts at:
Status: draft
Version: 0.0.1-draft
The drafter MUST emit those two values verbatim on creation. The
drafter MUST NOT pick a different starter version even if the user
prompt mentions one — those mentions are recorded as an Open
Question and resolved at publish time, not at creation.
V3 — Draft invariants (no mid-draft bumps; no mid-draft changelog)
While Status: draft:
The Version: field MUST NOT change as a result of any edit. All
edits (1, 10, or 100 of them) accumulate under the same
0.0.1-draft (or whatever draft tag the spec carries — see V11
for re-draft tags).
No CHANGELOG.md entry is written for in-draft edits.
Changelog entries are written only at the publish transition
(V8 / V11). This is a hard invariant — drafts MUST NOT log changes
to a changelog file even on user request; if the user asks for a
changelog mid-draft, the orchestrator surfaces an Open Question
("changelog will be written at publish") and proceeds.
The ## Changes since vN preamble is NEVER used during any
draft window — initial OR re-draft. Drafts carry no in-spec
change-tracking artefact (prd-evolution §5). Git is the
history source during the draft phase.
V4 — Mode-signal precedence
The agent decides whether the next edit set keeps the spec in
draft or transitions it to published using exactly this
precedence, highest first:
Explicit user statement (current turn). Recognised gestures:
STATUS: draft / STATUS: published.
"this is still draft", "keep it draft", "still in draft".
"publish this", "promote to ", "ship it",
"this is final", "publish at v0.1.0".
Any user statement naming an explicit non--draft semver as
the target — interpreted as publish-intent.
When (1) fires the agent records mode_signal: explicit in
state.json and proceeds without inference.
Branch-based inference (V5). Used only when (1) is silent.
Previous state. If branch detection is unavailable AND the
user said nothing, the spec stays in whatever mode it was in.
Initial creation defaults to draft (V2).
The orchestrator MUST NOT silently override an explicit user
statement. If (1) and (2) disagree (e.g. user says "still draft"
while on main), (1) wins; the agent surfaces a one-line warning
at end of turn but does not block.
V5 — Branch-based inference
When the user has not stated mode in the current turn, the
orchestrator delegates a branch-detection probe to
@context-detective (which has the research / shell-execution
authority in this pack). The probe runs:
git rev-parse --abbrev-ref HEAD
falling back to git symbolic-ref --short HEAD. Result categories:
Feature branch — anything that is notmain, master,
trunk, default. Inference: stays draft. No prompt.
Trunk branch — main, master, trunk, or default.
Inference: draft is suspect. Trigger V6 only if the spec
currently carries Status: draft.
Detached HEAD / no git / git error. Treat as inference
unavailable. Apply V4 step 3. Add a one-line note: "Could not
detect branch; assumed current status was preserved. State
explicitly with STATUS: draft or STATUS: published if needed."
The detective's branch-probe response is cached on state.json
(branch_name, branch_kind, branch_inference_at) for the
session; it is re-checked at session-start of each follow-up turn.
The probe is capped at one invocation per session; re-checks
read the cache.
V6 — Main-branch-still-draft prompt (verbatim)
Triggered by V5 when branch_kind == trunk AND
spec.status == draft AND the user did not state mode in the
current turn. The orchestrator parks at
phase: awaiting-mode-decision and presents this verbatim message:
The spec at <output_path> is currently Status: draft
(Version: <current>), but the working branch is <branch_name>
(looks like a trunk/release branch). Drafts on trunk are unusual —
please tell me how to proceed:
PUBLISH <proposed-version> — drop the -draft suffix,
freeze numbering, write a CHANGELOG entry. (Default proposed
version: 0.1.0 from 0.0.1-draft; otherwise the smallest bump
matching your changes — see V9.)
KEEP-DRAFT — acknowledge the risk and continue editing in
draft on this branch. I will repeat this prompt on every future
turn that detects the same condition.
ABORT — make no further edits this turn.
Reply with one of: PUBLISH <version>, KEEP-DRAFT, or ABORT.
Reply handling:
PUBLISH <ver> → V8 publish transition with that version.
KEEP-DRAFT → record state.json:keep_draft_acknowledged: true
for THIS turn only (re-prompt next turn).
ABORT → no draft edit this turn; orchestrator returns control.
Any other reply → re-prompt with the same template (no third
interpretation; mirrors Stop A C4 disambiguation).
V6.1 — Initial-publish version choice (OQ-1 resolution)
When the user asks to publish a spec currently at 0.0.1-draft
(initial publish), the orchestrator MUST ASK the user to pick the
target version rather than silently default. The Stop A bump-line
shows BOTH options:
Initial publish from 0.0.1-draft. Pick a target:
PUBLISH 0.1.0 — early-development semver (default,
recommended for evolving APIs).
If the user simply types PUBLISH with no version, treat as
disambiguation and re-prompt. Never silently auto-bump.
V7 — Pre-merge reminder (draft cadence)
While the spec is in draft, the orchestrator's session-summary
output at the end of every edit turn MUST include a
pre-merge-reminder fenced block:
pre-merge-reminder
status: draft
current_version: <version>
reminder: "Before merging this branch to main/master, manually
publish the spec (drop -draft, set explicit semver). Reply
`PUBLISH <version>` to do it now."
Cadence rule: the reminder fires on every edit turn that left
the spec in draft. It is suppressed on read-only turns (no spec
mutation). The orchestrator gates on a mutated_this_turn flag
in state. At session-start (first turn) the orchestrator MUST also
restate Status and Version (V14).
V8 — Publish transition
"Publishing" is a single, explicit, atomic operation. It is
triggered ONLY by:
An explicit user gesture per V4(1) that conveys publish intent
(e.g. PUBLISH <ver>, "publish this", "promote to 0.1.0").
A PUBLISH <ver> reply to the V6 main-branch prompt.
Publishing is NEVER triggered by branch inference alone, by elapsed
time, or by edit count.
Mechanics performed atomically by the drafter:
Strip the -draft suffix from Version:.
Set the explicit semver:
When transitioning from 0.0.1-draft, V6.1 applies: the user
was asked and named 0.1.0 or 1.0.0 (or another explicit
value). The drafter takes that verbatim.
User-named version overrides verbatim. Validation: SemVer 2.0
syntax, MUST NOT contain -draft or any pre-release identifier.
Invalid input → re-prompt with V6 template.
Set Status: published.
Freeze numbering (V9 takes effect from this point).
4a. Re-materialise pending deletions. Walk
state.json:pending_published_id_deletions (queued during the
draft window per prd-evolution §3b). For each entry, insert
a stub heading bearing the deleted ID in the published artefact
body with a [Deprecated in <publishing-version>] (or
[Deprecated in <publishing-version>, superseded by <successor>]
if successor is set) marker and a one-line pointer body. The
stub honours V9's cross-reference guarantee at the published
boundary. Clear the field once all entries are materialised.
The draft body itself never carried these markers — they exist
only in the published artefact.
Write a CHANGELOG.md entry (V17). For initial publish (kind
publish-initial) the entry is an aggregate summary by
default ("12 FRs, 8 NFRs, 34 ACs added"); for re-draft publish
(kind publish-redraft) the entry uses enumerated
### Added / ### Changed / ### Removed per V17. (OQ-5
resolution: aggregate at initial-publish; enumerated thereafter.)
Set Last Updated: <today> in Document Information.
Emit version-bump-json with kind: publish-initial (or
publish-redraft for V11) so the critic can score it.
V9 — Post-publish numbering stability
Once Status: published:
All requirement IDs are immutable. Renumbering is forbidden.
This re-asserts prd-evolution §3, scoped to published specs.
New items get the next available ID after the highest current
ID for that kind (e.g. if FR-29 is highest, the next added FR is
FR-30 — even if FR-12 was deprecated and a "natural" slot exists).
Deletions are forbidden in the published artefact. Items
that should disappear are marked in place in the published
artefact with one of:
[Deprecated in vX.Y]
[Deprecated in vX.Y, superseded by FR-NN]
[Removed in vX.Y] — only after at least one MINOR cycle of
deprecation.
During a draft window, items MAY be deleted from the working
spec body per prd-evolution §3a / §3b. Prior-published IDs
that are deleted in-window are recorded in
state.json:pending_published_id_deletions and re-materialised
as [Deprecated in vX.Y] / [Removed in vX.Y] markers in
the published artefact only at the publish transition (V8 step
4a — added by prd-evolution §3b). The draft body itself never
carries the marker.
Cross-references to deprecated/removed IDs MUST still resolve
in the published artefact (the stub heading remains there;
only the body is a brief pointer). Cross-references in the
draft body are scrubbed at deletion time per V12.
V10 — Post-publish version bumps
When editing a published spec the drafter classifies the change
set against these triggers:
Change type
Bump
Typo, formatting-only, wording that does not change interpretation
PATCH (0.1.0 → 0.1.1)
New FR / NFR / AC / R / OQ; new gated section because an axis newly fires; clarifying rewording of an AC for testability
Pre-1.0.0 nuance (OQ-3 resolution). SemVer 2.0 §4 says any
breaking change pre-1.0.0 may be released as a MINOR bump. The
drafter records this classification (pre_1_0_0_warning: true in
version-bump-json). The orchestrator surfaces a Stop-A
"consider bumping to 1.0.0" prompt — it does not auto-bump
silently. The user explicitly opts in.
User override: the user MAY name an explicit target version in
their prompt or at Stop A; that wins. The drafter records the
override and the auto-classified bump in version-bump-json with
override: true|false.
V11 — Re-draft cycle
When the user issues an edit instruction against a Status: published
spec, the drafter enters a re-draft window:
Compute the auto-classified bump (V10) → <next> (e.g. 0.1.1).
Set the spec's working version to <next>-draft (e.g.
0.1.1-draft).
Set Status: draft.
The session is now in re-draft. V3 invariants apply (in-window
edits do NOT bump the working version further; no mid-window
changelog writes; no in-spec change-tracking artefacts per
prd-evolution §5).
Pre-merge reminder (V7) fires again every edit turn.
Publishing this re-draft uses V8 with kind: publish-redraft.
The -draft suffix drops; changelog entries are written under
[<next>] - YYYY-MM-DD using enumerated Keep-a-Changelog
categories per V17.
Prior-published IDs inside a re-draft window (OQ-6 resolution,
revised). Prior-published IDs MAY be deleted from the working
draft body per prd-evolution §3a / §3b. They MUST NOT be
renumbered — V9 immutability over the ID itself is permanent.
If a prior-published item is removed in-window, the drafter
queues a publish-time re-materialisation in
state.json:pending_published_id_deletions so V9's
published-artefact cross-reference guarantee is honoured at
publish (§3b). Newly added items inside the re-draft window MAY
be renumbered among themselves until publish — they have not
yet been published, so their IDs are not yet a contract.
Practically: if the user adds FR-30, then later inserts a more-
logical FR before it, the new pair MAY be reordered as FR-30 and
FR-31 either way, provided that no published ID is renumbered.
Cross-references inside the re-draft window MUST still update
atomically (V12).
This rule applies in all modes, but bites hardest during draft
renumbering. It is a hard invariant, not a guideline:
Every renumber, insert, or delete operation MUST update all
internal cross-references to the affected IDs in the same edit.
A spec MUST NOT, at the end of any drafter turn, contain a
cross-reference to an ID that does not exist or whose meaning has
shifted.
Eligible references the drafter MUST scan and update:
"depends on FR-N" / "see FR-N" / "supersedes FR-N" prose.
Acceptance Criteria sub-IDs (AC-<FR>.<n>) where <FR> shifted.
The drafter MUST emit a cross-ref-audit-json fenced block on
every turn that touches IDs, enumerating every reference it
touched, with old → new ID mapping. The critic scores cross-ref
integrity as a sub-rubric of D6 (d6.cross-ref-integrity) at
severity: blocker — a single broken reference fails the dimension.
V13 — Eligible items for draft renumbering
During an active draft window (initial or re-draft, subject to V11's
"only newly-added items in re-draft"), the following item types MAY
be renumbered:
Item kind
ID pattern
Notes
Functional Requirement
FR-NN
Insert/delete/reorder allowed.
Non-Functional Requirement
NFR-NN
Same.
Acceptance Criterion
AC-<FR>.<n>
Renumbers if parent FR renumbers; per-FR sub-numbering also fluid.
Risk
R-NN
Same.
Open Question
OQ-NN
Same.
Test Scenario
TS-NN
Same.
Section ordering is also fluid in initial draft; in re-draft only
newly added sections may be moved (existing ones are part of
the published contract).
(The catalogue does not currently declare G-NN glossary or
UC-NN use-case IDs; if a future version of prd-template adds
them, this list extends accordingly. F22 is filed as future work.)
V14 — Status surfacing
At session-start (first orchestrator turn that loads an existing
spec) the orchestrator MUST emit a spec-status line up-front:
After every mutation turn the orchestrator restates the same block
in session-summary. Published specs surface the same block sans
pre-merge-reminder.
V15 — Front-matter location and malformed handling
The version metadata lives in the spec's ## Document Information
section (already present per current template), with two normative
fields:
- **Status**: draft | published
- **Version**: <semver> # carries -draft suffix iff status == draft
Parser rules:
Both fields MUST be present in every spec.
If Status is missing or unrecognised → treat as
Status: draft AND surface a CONCERN at end of turn ("Status
field missing; treated as draft. Confirm with STATUS: draft or
STATUS: published.").
If Version is missing → treat as 0.0.1-draft AND surface a
CONCERN.
If Version is present but malformed (not SemVer 2.0, or has
-draft while Status: published, or omits -draft while
Status: draft) → block the turn with a structured error.
Do NOT auto-correct; ask the user to fix the metadata.
Old-style Status: [Draft | In Review | Approved | Archived]
template-comment values are tolerated on first read and
rewritten by the drafter to the canonical two-value enum on next
edit, with a one-line note in session-summary.
V16 — Multi-spec independence
Each spec's Status and Version are independent. There is no
shared cross-spec version manifest. When the user names multiple
specs in one prompt, each is processed with its own status/version
state derived from its own front-matter.
V17 — Publish-time changelog
At every publish (V8 step 5) the drafter writes a CHANGELOG.md
entry sibling to the spec, following the existing
prd-evolution §6 Keep-a-Changelog format. New rules:
No mid-draft changelog writes (OQ-5). Drafts NEVER log to
the changelog file; the file is created or appended only at the
publish transition.
Initial publish format (kind publish-initial): a single
aggregate-summary line under ### Added, e.g.
### Added: 12 FRs, 8 NFRs, 34 ACs, 4 risks, 3 open questions defined. This is the default per OQ-5; users may override at
Stop A by replying PUBLISH 0.1.0 ENUMERATE to force per-item
enumeration.
Re-draft publish (kind publish-redraft): every diff between
the prior published version and the re-drafted-and-now-published
version MUST appear in exactly one Keep-a-Changelog category
(### Added / ### Changed / ### Deprecated / ### Removed
/ ### Fixed / ### Security). This is enumerated, not
aggregated.
V18 — Quality-checklist additions for the critic
Add to the critic's checklist (D6/D7 sub-rubrics):
Status and Version are present and mutually consistent
(both indicate draft, or both indicate published).
If Status: draft, no CHANGELOG.md entry was written this
turn.
If Status: published, no in-place renumbering of IDs that
were present in the immediately prior published version.
Every renumber/insert/delete in cross-ref-audit-json has a
corresponding update to all references.
On publish, CHANGELOG.md entry exists and accounts for
every diff (or, for publish-initial, contains the aggregate
summary).
If Status: draft, the spec body MUST NOT contain a
## Changes since, ## Changelog, ## Revision History,
or ## What's Changed heading, and MUST NOT contain inline
[Changed in vX.Y] markers or prose narrating what changed
since the prior version (per prd-evolution §5;
d7.draft-no-change-tracking).
On publish, every entry queued in
state.json:pending_published_id_deletions was
re-materialised as a [Deprecated in <ver>] /
[Removed in <ver>] stub in the published artefact (V9
published-contract preserved; prd-evolution §3b step 5).
On the V6 prompt path, the user replied with PUBLISH ...,
KEEP-DRAFT, or ABORT — never silently transitioned.
Drafter renumbering algorithm (pseudo-code)
on_insert_or_delete(item, position):
affected_ids = compute_displaced_ids(item.kind, position)
rename_map = build_rename_map(affected_ids) # old -> new
apply_rename_to_headings(rename_map)
for each ref_pattern in [inline, anchored, prose, ac_sub, changes_preamble]:
scan spec body and apply rename_map left-to-right (single pass)
validate_no_orphan_refs() # every FR-/NFR-/etc. reference must resolve
emit cross-ref-audit-json with rename_map and updated reference list
If the orchestrator's task prompt declares an edit that would
renumber a published ID (i.e. an ID from last_publish_version),
the drafter refuses with a structured error referencing V9.
Branch-probe contract (used by @context-detective)
When the orchestrator's task prompt to the detective declares
probe: branch-only, the detective runs the V5 git commands and
emits only a branch-probe-json fenced block (no discovery, no
context pack). Schema:
Trunk names: main, master, trunk, default. Anything else
(other than HEAD / detached) is feature. On git error,
branch_kind: "unknown", populate git_error with the stderr
fragment.
State-schema fields touched by this skill
These fields are added to prd-template/references/state-schema.md
(single source of truth):
Field
Purpose
spec_status
mirror of front-matter (draft | published).
spec_version
SemVer; carries -draft when draft.
branch_name
result of V5 probe; nullable.
branch_kind
trunk | feature | detached | unknown.
branch_inference_at
ISO8601; cached probe time.
mode_signal
explicit | branch | preserved.
keep_draft_acknowledged
reset every turn.
awaiting_mode_decision
true while V6 prompt is parked.
last_publish_version
helps re-draft compute <next>-draft.
mutated_this_turn
gates V7 reminder cadence.
pending_published_id_deletions
array of prior-published IDs deleted from the working draft during a re-draft window; re-materialised as deprecation markers in the published artefact at V8 step 4a (prd-evolution §3b). Cleared at publish.
Cross-references
prd-evolution/SKILL.md §3 (now scoped to Status: published).
prd-template/references/specification-template.md (Status /
Version field shape per V15).
prd-template/references/state-schema.md (canonical
state.json schema; the fields above are listed there).
