Claude-Code-Game-Studios/CCGS Skill Testing Framework/skills/pipeline/vertical-slice.md

Skill Test Spec: /vertical-slice

Skill Summary

/vertical-slice [--review full|lean|solo] is the Pre-Production validation build: one complete [start → challenge → resolution] loop at near-production quality, answering a falsifiable question about player experience and build feasibility. It reads the concept (design/gdd/game-concept.md, or design/game-brief.md at rigor: minimal), the systems index, the architecture and the control manifest, defines the validation question and a 3–5 minute scope, asks to record the plan in production/session-state/active.md, then asks "May I create the vertical slice directory at prototypes/[concept-name]-vertical-slice/ and begin implementation?". Every slice file starts with a VERTICAL SLICE - NOT FOR PRODUCTION header in that file's own comment syntax.

After a one-question-at-a-time debrief it writes prototypes/[concept-name]-vertical-slice/REPORT.md (from .claude/docs/templates/vertical-slice-report.md) and a prototypes/index.md row, one ask naming both. Verdicts are PROCEED / PIVOT / KILL, plus NOT ASSESSED when nobody has played the loop through from scratch — ranked above PROCEED and below PIVOT and KILL, and never recorded as PROCEED. The verdict feeds /gate-check production, which FAILs a built slice with any NO among its Vertical Slice Validation items, so PROCEED is not available after such a NO. In full review mode creative-director reviews the result via CD-PLAYTEST (APPROVE / CONCERNS / REJECT, or NOT ASSESSED): APPROVE keeps the recommendation; CONCERNS go to the user (Revise / Accept / Discuss); REJECT means PROCEED cannot stand — the user chooses PIVOT or KILL and the report records the director's reason. PIVOT writes a PIVOT-NOTE.md; KILL runs a five-item soundness check and appends to prototypes/GRAVEYARD.md.

---

Static Assertions (Structural)

Checked against the SKILL.md by /skill-test spec — no fixture needed.

• Has required frontmatter fields: name, description, argument-hint, user-invocable, allowed-tools
• Has ≥2 phase headings
• Contains verdict keywords: PROCEED, PIVOT, KILL, NOT ASSESSED
• Contains "May I" language before creating the slice directory and before writing REPORT.md, PIVOT-NOTE.md and GRAVEYARD.md
• Has a next-step handoff per verdict (PROCEED → /create-epics, /create-stories, /sprint-plan, /gate-check production; PIVOT → /design-system, /architecture-decision, /vertical-slice; KILL → /brainstorm or /prototype)
• Documents CD-PLAYTEST: spawned in full, skipped with a note in lean/solo

---

Director Gate Checks

• Full mode: CD-PLAYTEST — creative-director receives the REPORT.md content, the validation question and the pillars and core fantasy (from design/gdd/game-concept.md, or the pitch and "what they feel" line from design/game-brief.md). APPROVE → the recommendation stands; CONCERNS → shown with the recommendation, and the user decides (Revise / Accept / Discuss); REJECT → a PROCEED cannot stand: the user chooses PIVOT or KILL, and REPORT.md records the director's reason (a PIVOT or KILL stands); NOT ASSESSED [missing input] → neither backs nor overturns the recommendation: the input is supplied and the gate re-run, or CD-PLAYTEST: NOT ASSESSED — [input] is recorded. Any change to REPORT.md and its prototypes/index.md row is asked first. When the slice's own verdict is NOT ASSESSED the gate is skipped: "CD-PLAYTEST skipped — the slice has not been played yet."
• Lean mode: skipped — "CD-PLAYTEST skipped — Lean mode."
• Solo mode: skipped — "CD-PLAYTEST skipped — Solo mode."

---

Test Cases

Case 1: Happy Path — slice built, played unaided, PROCEED

Fixture:

• project.yaml: engine.name: Godot, engine.language: GDScript, modes.review_mode: solo
• design/gdd/game-concept.md (core fantasy, pillars), design/gdd/systems-index.md, docs/architecture/architecture.md and docs/architecture/control-manifest.md exist; prototypes/ has no ember-trail-vertical-slice/
• In the debrief the player finished the loop unaided, reached the first meaningful action in 40 seconds, felt the core fantasy and hit no fun blocker; the user answers PROCEED

Input: /vertical-slice

Expected behavior:

1. Reads the concept, systems index, architecture and manifest; states a validation question covering player experience AND build feasibility
2. Presents a 3–5 minute scope and the build plan; the confirmation names production/session-state/active.md, and the checkpoint is written there
3. Asks "May I create the vertical slice directory at prototypes/ember-trail-vertical-slice/ and begin implementation?"; every GDScript file it writes starts with # VERTICAL SLICE - NOT FOR PRODUCTION
4. Iterates with the user until the loop is playable, then asks for the from-scratch playthrough and puts the six debrief questions one at a time
5. Drafts the report from .claude/docs/templates/vertical-slice-report.md with a day-by-day velocity log; asks "May I write this report to prototypes/ember-trail-vertical-slice/REPORT.md and add its row to prototypes/index.md?"; writes both
6. Notes "CD-PLAYTEST skipped — Solo mode."
7. PROCEED routes to /create-epics layer:foundation, /create-epics layer:core, /create-stories, /sprint-plan and /gate-check production

Assertions:

• The validation question has both a player-experience and a build-feasibility part
• Scope is confirmed, and the checkpoint ask names active.md, before any slice file is written
• The directory ask comes before any slice file; each GDScript file's header is a # comment, never //
• The debrief questions are asked one at a time, after the from-scratch playthrough
• REPORT.md and the prototypes/index.md row are written only after one ask naming both; the row records PROCEED
• The formal stage step is /gate-check production (Pre-Production → Production), not /gate-check pre-production

---

Case 2: Failure Path — a built slice with a NO cannot PROCEED

Fixture:

• As Case 1, but in the debrief the player needed the developer to explain the core action (question 1: no)
• The user answers PROCEED anyway, then picks PIVOT when asked

Input: /vertical-slice

Expected behavior:

1. The skill does not record PROCEED: /gate-check production FAILs a built slice with any NO among its Vertical Slice Validation items (.claude/skills/gate-check/references/gate-production.md)
2. It asks the user to choose PIVOT or KILL, naming question 1's answer as the reason
3. The user picks PIVOT; REPORT.md is written, after its ask, with verdict PIVOT and the NO named
4. The PIVOT route runs: the two carry-forward questions, one at a time, then "May I write this to prototypes/ember-trail-vertical-slice/PIVOT-NOTE.md?"

Assertions:

• PROCEED is not written to REPORT.md or prototypes/index.md while a validation item is NO
• The user, not the skill, chooses between PIVOT and KILL
• The report names the failed validation item
• PIVOT-NOTE.md is written only after its ask, with what worked, what failed and what the next slice should prove

---

Case 3: Mode Variant — CD-PLAYTEST in full, skipped in lean and solo

Fixture:

• A slice has been debriefed with a PROCEED recommendation and REPORT.md written
• Run A: modes.review_mode: full; CD-PLAYTEST returns APPROVE
• Run B: modes.review_mode: lean
• Run C: modes.review_mode: full, input /vertical-slice --review solo

Expected behavior:

1. Run A: creative-director is spawned for CD-PLAYTEST after REPORT.md is written, with the REPORT.md content, the validation question and the pillars and core fantasy; APPROVE leaves PROCEED standing and REPORT.md unchanged
2. Run B: no spawn; "CD-PLAYTEST skipped — Lean mode."
3. Run C: --review solo overrides the resolved full; no spawn; "CD-PLAYTEST skipped — Solo mode."

Assertions:

• Run A spawns CD-PLAYTEST only after REPORT.md exists, with the three inputs
• Run A's APPROVE keeps the recommendation and triggers no REPORT.md update ask
• Runs B and C spawn no director and print their skip notes
• No other director gate is spawned in any run

---

Case 4: Edge Case — the slice is built but nobody has played it

Fixture:

• The loop became demonstrable, but the user cannot play it through from scratch this session
• modes.review_mode: full

Input: /vertical-slice

Expected behavior:

1. The Phase 5 playthrough does not happen, so the verdict is NOT ASSESSED, with the reason
2. REPORT.md, after its ask, records the velocity log and NOT ASSESSED; the prototypes/index.md row records NOT ASSESSED
3. CD-PLAYTEST is not spawned: "CD-PLAYTEST skipped — the slice has not been played yet."
4. Next steps: finish the playthrough and debrief from the active.md checkpoint, then update REPORT.md after asking; the slice is not taken to /gate-check production as if it had passed

Assertions:

• The verdict is NOT ASSESSED — never PROCEED — and its reason is stated
• NOT ASSESSED is ranked above PROCEED and below PIVOT and KILL
• CD-PLAYTEST is not spawned for an unplayed slice, and the skip is announced
• PROCEED's next steps are not offered

---

Case 5: Director Gate — CD-PLAYTEST returns REJECT, CONCERNS or NOT ASSESSED

Fixture:

• A slice has been debriefed with a PROCEED recommendation and REPORT.md written; modes.review_mode: full; design/gdd/game-concept.md holds the pillars
• Run A: CD-PLAYTEST returns REJECT — the core fantasy is not present. Asked PIVOT or KILL, the user picks KILL; no playtest session showed an emotional high point, and this is the third slice attempt on the concept
• Run B: CD-PLAYTEST returns CONCERNS — the resolution beat undercuts a pillar; the user picks Accept with noted concerns
• Run C: CD-PLAYTEST returns NOT ASSESSED — the pillars were missing from the context it received

Input: /vertical-slice

Expected behavior:

1. Run A: the REJECT is read as the gate's word, not as a slice verdict. PROCEED cannot stand: the skill asks the user to choose PIVOT or KILL; the user picks KILL; after "May I update prototypes/[concept-name]-vertical-slice/REPORT.md and its prototypes/index.md row?" both record KILL and REPORT.md records the director's reason. Phase 8's KILL route runs: two checklist items apply, so the KILL is sound, and it asks "May I append this to prototypes/GRAVEYARD.md?"
2. Run B: the concerns are shown beside PROCEED with Revise the recommendation / Accept with noted concerns / Discuss further; on Accept, PROCEED stands and, after the update ask, REPORT.md records the concerns
3. Run C: the missing input is named; the gate is re-run with the pillars, or CD-PLAYTEST: NOT ASSESSED — [input] is recorded after the update ask; PROCEED is neither confirmed nor overturned by it

Assertions:

• The director's answer is read as APPROVE / CONCERNS / REJECT / NOT ASSESSED — never as a PROCEED / PIVOT / KILL of its own
• Run A: PROCEED is not kept after the REJECT; the user chooses PIVOT or KILL, and the director's reason is in REPORT.md
• Run A: the GRAVEYARD.md entry is appended only after its ask, with a specific kill reason
• Run B: the user decides; the recommendation changes only if the user revises it
• Run C: NOT ASSESSED is not treated as APPROVE
• Every REPORT.md / prototypes/index.md update follows an ask naming both

---

Protocol Compliance

• Reads the concept (or the brief at minimal) before defining the validation question
• Confirms scope before building; asks before recording the plan in active.md and before creating prototypes/[concept-name]-vertical-slice/
• Keeps slice files under prototypes/, each with the VERTICAL SLICE - NOT FOR PRODUCTION header in its own comment syntax
• Asks before writing REPORT.md (with its prototypes/index.md row), PIVOT-NOTE.md, GRAVEYARD.md and any post-gate report update
• Never records PROCEED after a NO validation item, and never for a slice nobody has played
• Applies the review mode before CD-PLAYTEST and prints the skip note when skipped
• Ends with the per-verdict next steps

---

Coverage Notes

• The multi-turn build loop, the day-3 sunk-cost stop and the day-by-day velocity log need a live run to verify.
• Reading a prior PIVOT-NOTE.md on a re-run to frame the new validation question is not tested.
• The rigor: minimal path (the brief's pitch and "what they feel" line in place of pillars) is not given its own fixture.
• A KILL whose soundness checklist gives 0–1 items (one targeted PIVOT suggested instead) is not tested.
• The networked-game limit (a local slice cannot validate network feel) is advisory text, not tested.