Swain Design

Spec Management

Before proceeding with any state-changing operation, check for an active session:

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)" && bash "$REPO_ROOT/.agents/bin/swain-session-check.sh" 2>/dev/null

If the JSON output has "status" other than "active", inform the operator: "No active session — start one with /swain-init?" Proceed if they dismiss.

This skill defines the canonical artifact types, phases, and hierarchy. Detailed definitions and templates live in references/ (relative to this skill's directory). If the host repo has an AGENTS.md, keep its artifact sections in sync with the skill's reference data.

Artifact type definitions

Each artifact type has a definition file (lifecycle phases, conventions, folder structure) and a template (frontmatter fields, document skeleton). Read the definition for the artifact type you are creating or transitioning.

Choosing the right artifact type

When the user's request doesn't name a specific type, infer it from their intent:

Initiative vs Epic — the key distinction:

• Initiative: a direction with multiple deliverables. "Harden security" is an initiative — it spans scanning, gates, policies. The operator steers it.
• Epic: a deliverable with multiple specs. "Build the scanning tool" is an epic — it has clear completion criteria. Agents execute it.
• Rule of thumb: if the work needs 2+ epics to describe, it's an Initiative. If it needs 2+ specs, it's an Epic. If it's one spec, just create the spec.

Spec under Initiative (small work path) — bugs, minor enhancements, and chores that relate to an Initiative's direction but don't warrant an Epic can attach directly to the Initiative via parent-initiative. If small work clusters, suggest promoting it to an Epic.

Updating artifact metadata

When the operator asks to update a field on an existing artifact (e.g., "set VISION-001 priority to high", "re-parent EPIC-017 under INITIATIVE-001"):

1. Read the artifact's definition file to confirm the field name and valid values
2. Edit the frontmatter field directly (e.g., priority-weight: high)
3. Update the last-updated date
4. Run bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/chart.sh" build to refresh the graph cache
5. Commit the change

Common updates:

• priority-weight on Visions, Initiatives, Epics, and Specs — accepts high, medium, or low. Cascades: Vision → Initiative (can override) → Epic (can override) → Spec (can override). Affects downstream recommendation scoring and sibling sort order in swain chart.
• parent-initiative on Epics and Specs — re-parents them under an Initiative. A Spec can have parent-epic OR parent-initiative, never both.
• parent-vision on Initiatives — attaches to a Vision.

When the operator says "priority" or "weight" in the context of a Vision or Initiative, they mean the priority-weight frontmatter field.

Creating artifacts

Error handling

When an operation fails (missing parent, number collision, script error, etc.), consult references/troubleshooting.md for the recovery procedure. Do not improvise workarounds — the troubleshooting guide covers the known failure modes.

Complexity tier detection (SPEC-045)

Before running the full authoring ceremony, classify the artifact into a complexity tier:

Low complexity (fast-path eligible):

• SPEC with type: bug or type: fix and no parent-epic and no downstream depends-on links
• SPIKE with no parent-epic
• Any artifact where the user uses language like "quick", "simple", "trivial", or "fast"

Medium/High complexity (full ceremony):

• Feature SPECs (type: feature)
• Any SPEC or SPIKE with a parent-epic
• EPICs, INITIATIVEs, Visions, Journeys, ADRs — always full ceremony
• Any artifact where the user describes significant architectural decisions

When fast-path applies, output: [fast-path] Skipped: specwatch scan, scope check, index update

Workflow

1. Determine the next available number for the prefix by running:

   bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/next-artifact-id.sh" <PREFIX>
   

   This scans ALL local branches and the working tree to prevent ID collisions across worktree sessions (SPEC-193). If the script is unavailable, fall back to scanning docs/<type>/ on the current HEAD — but note this risks collisions in worktree workflows.

2. For VISION artifacts: Before drafting, ask the user whether this is a competitive product or a personal product. The answer determines which template sections to include and shapes the entire downstream decomposition. See the vision definition for details on each product type. 2a. For DESIGN artifacts: First, ask which domain this design covers: interaction (UI/UX — screens, flows, states), data (data architecture — entities, schemas, flows, invariants), or system (system contracts — API boundaries, behavioral guarantees, integration interfaces). Default to interaction if unclear. Then prompt for Design Intent content — Context (one sentence anchoring the design to its purpose), Goals (what experience or guarantee we're trying to create), Constraints (reviewable boundaries), and Non-goals (what we explicitly decided not to do). This section is write-once: it is set at creation and not updated as the mutable sections evolve. Use the domain-specific template sections from the DESIGN template.

3. Read the artifact's definition file and template from the lookup table above.

4. Create the artifact in the correct phase subdirectory. Create the phase directory with mkdir -p if it doesn't exist yet. See the definition file for the exact directory structure.

5. Populate frontmatter with the required fields for the type (see the template). Set authored-by to the model identity resolved from the system prompt (e.g., GLM-5.1). If subagents were involved in creating the artifact, list all models with roles — e.g., GLM-5.1 (supervisor), Kimi-K2.5 (subagent). If the model identity is unavailable, use AI Assistant as a fallback.

6. Initialize the lifecycle table with the appropriate phase and current date, using this rule:

   • User-requested → Active: if the user explicitly asked for this artifact (e.g., "new SPIKE about X", "write a spec for Y"), create it directly in Active. The user has already decided they want this work — Proposed adds no value.
   • Agent-suggested → Proposed: if the agent creates the artifact on its own initiative (e.g., suggesting a SPIKE while the user asked for an EPIC, decomposing a Vision into child Epics), create it in Proposed. The user hasn't explicitly committed — Proposed signals "here's what I recommend, please confirm."
   • Fully developed in-session → later phase: an artifact may be created directly in a later phase if it was fully developed during the conversation (see Phase skipping). 6.5. Hyperlink bare artifact ID references in body text — after writing the artifact body, scan all text below the closing --- frontmatter fence for bare artifact ID references matching the pattern (SPEC|EPIC|INITIATIVE|VISION|SPIKE|ADR|PERSONA|RUNBOOK|DESIGN|JOURNEY|TRAIN|CHORE)-[0-9]+. For each bare ID that is:
   • not already inside a markdown link ([...](...)), and
   • not inside a code fence (``` block) or inline code (`backtick`),

   resolve it with:

   bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/resolve-artifact-link.sh" <ARTIFACT-ID> <SOURCE-FILE>
   

   Replace the bare ID with [ARTIFACT-ID](relative-path). If the script returns a non-zero exit code or empty output (artifact not found), leave the bare ID as-is — do not fail the operation. Frontmatter values must remain as plain IDs (YAML compatibility); only body text gets hyperlinks.

7. Validate parent references exist (e.g., the Epic referenced by a new Agent Spec must already exist). 7.5. Same-type overlap check — (standing-track types only: DESIGN, Persona, Runbook) scan docs/<type>/Active/ for existing Active artifacts of the same type. Flag overlap if:

   • The new artifact's linked-artifacts references another artifact of the same type — this is a direct supersession signal.
   • The new artifact's scoping section (Interaction Surface for DESIGNs, Trigger for Runbooks, Role for Personas) describes a surface that overlaps with or subsumes an existing Active artifact's scope. If overlap is detected, ask the operator: "This overlaps with <EXISTING-ID> (<title>). Does the new artifact supersede it?" If yes, transition the existing artifact to Superseded (set superseded-by, update status, move to Superseded/ directory, add lifecycle entry) as part of the same operation.

8. ADR compliance check — run bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/adr-check.sh" <artifact-path>. Review any findings with the user before proceeding. 8a. Alignment check — (skip for fast-path tier) run bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/chart.sh" scope <artifact-id> and assess per references/alignment-checking.md. Report blocking findings (MISALIGNED); note advisory ones (SCOPE_LEAK, GOAL_DRIFT) without gating the operation. When displaying scope chains or ancestry to the operator, use artifact-context.sh for each node to show plain-language names alongside IDs. Fall back to bare IDs if unavailable. 8b. Unanchored check — after validating parent references, check if the new artifact has a path to a Vision via parent edges. If not, warn: ⚠ No Vision ancestry — this artifact will appear as Unanchored in swain chart. Offer to attach to an existing Initiative or Epic. Do not block creation.

9. Post-operation scan — (skip for fast-path tier) run bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/specwatch.sh" scan. This now also runs design-check.sh as part of the scan pipeline. Fix any stale references or design drift findings before committing.

10. Index refresh step — (skip for fast-path tier; batch refresh at session end via rebuild-index.sh) update list-<type>.md (see Index maintenance).

Superpowers integration

When superpowers is installed, the following chains are mandatory — invoke the skills, do not skip them or do the work inline:

1. Before creating Vision, Initiative, or Persona artifacts: Invoke the brainstorming skill for Socratic exploration. Pass the artifact context (goals, audience, constraints). Capture brainstorming output into swain's artifact format with proper frontmatter and lifecycle table.

2. When new feature work begins (brainstorming → artifacts → implementation): Brainstorming explores the idea and produces a design doc. Brainstorming's terminal state is invoking swain-design to create formal artifacts (epic + specs). Then, for each SPEC coming up for implementation, invoke writing-plans per-spec. After writing-plans saves a plan file, invoke swain-do for plan ingestion. The chain is: brainstorming → swain-design → per-spec writing-plans → swain-do.

2b. When an existing SPEC comes up for implementation (no brainstorming needed): Invoke writing-plans with the SPEC's acceptance criteria and scope. After writing-plans saves a plan file, invoke swain-do for plan ingestion.

3. For Testing → Implemented transitions: Invoke requesting-code-review for spec compliance and code quality review (if the review skills are available).

Detection: ls .agents/skills/brainstorming/SKILL.md .claude/skills/brainstorming/SKILL.md 2>/dev/null — if at least one path exists, superpowers is available. Cache the result for the session.

Read references/superpowers-integration.md for thin SPEC format and full routing details. All integration is optional — swain functions fully without superpowers.

Phase transitions

Phases are waypoints, not mandatory gates — artifacts may skip forward. Read references/phase-transitions.md for phase skipping rules, the transition workflow (validate → move → commit → hash stamp), verification/review gates, and completion rules.

Supersession specwatch-ignore maintenance

Whenever ANY artifact transitions to Superseded — whether via the phase transition workflow (step 5a in phase-transitions.md) or during artifact creation (same-type overlap detection) — append glob patterns to .agents/specwatch-ignore for the intentional backward references that the supersession creates. This prevents specwatch from flagging provenance links as warnings.

1. Create .agents/specwatch-ignore if it doesn't exist.
2. Append patterns for: (a) the superseded artifact path, (b) the superseding artifact path, (c) any ADR created as part of the same operation that references the superseded artifact.
3. Each entry gets a comment: # <OLD-ID> superseded by <NEW-ID> (<YYYY-MM-DD>).
4. Deduplicate: skip patterns that already exist in the file.

# INITIATIVE-001 superseded by INITIATIVE-013 (2026-03-19)
docs/initiative/Superseded/(INITIATIVE-001)*
docs/initiative/Active/(INITIATIVE-013)*

This step runs before the back-reference update (step 5b) and specwatch scan (step 9 in the creation workflow, step 8 in phase-transitions.md) so the scan output is clean.

Back-reference update on supersession

After specwatch-ignore maintenance (step 5a), update all non-terminal artifacts that reference the superseded artifact in frontmatter (linked-artifacts, depends-on-artifacts, addresses). See step 5b in phase-transitions.md for the full procedure. Key points:

• Check alignment before updating — supersession often changes scope. Read the referencing artifact's context and compare against the successor. If the relationship doesn't hold for the successor, flag it for the operator instead of silently repointing.
• Dedup — if the successor is already in the list, remove the old entry instead of adding a duplicate.
• Commit message provenance — record what changed (e.g., "update EPIC-031 linked-artifacts: INITIATIVE-001 → INITIATIVE-013") so git history preserves the original reference.
• Provenance links from the superseding artifact itself go to specwatch-ignore, not rewritten.

DESIGN lifecycle hooks

These hooks apply to DESIGN artifacts during phase transitions:

On DESIGN creation:

• Validate all sourcecode-refs paths exist at HEAD (if any are populated). Warn on broken paths before completing creation.

On Proposed → Active transition:

• Run design-check.sh on the DESIGN — all refs must be CURRENT.
• If any are STALE or BROKEN, warn the operator before completing the transition. Do not silently proceed with stale refs.

On Active → Superseded transition:

• The new (superseding) DESIGN should inherit sourcecode-refs from the old DESIGN with fresh pins via --repin.

Decision protection hooks

These hooks are agent-level behavioral guidance — they are not enforced by scripts but by the agent following this skill file.

SPEC Implementation transition: When a SPEC transitions to Implementation and has a linked DESIGN (via either side's linked-artifacts or artifact-refs):

• Surface the DESIGN's Design Intent section (Goals, Constraints, Non-goals) for alignment awareness. Present this to the operator so implementation stays within design boundaries.

SPEC completion: When a SPEC completes and its implementation changed files tracked by a DESIGN's sourcecode-refs:

• Cross-reference changed files (from the SPEC's commits) against active DESIGNs' sourcecode-refs.
• If overlap is found: nudge the operator to update the DESIGN and re-pin via design-check.sh --repin.

Alignment cascading: When an Epic has artifact-refs with rel: [aligned] pointing to a DESIGN:

• When child SPECs are created or modified, check scope against the DESIGN's Constraints and Non-goals.
• Traversal path: SPEC → parent EPIC → artifact-refs with rel: [aligned] → DESIGN → Design Intent.
• Only surface violations — silent pass for aligned SPECs.

Design-to-code drift: When a DESIGN's mutable sections are modified but sourcecode-refs blobs haven't changed:

• Surface: "DESIGN-NNN evolved but tracked code hasn't caught up." Nudge the operator to reconcile.

Drift resolution (SPEC-307)

On SPEC create, edit, or phase transition — if the SPEC has a parent (EPIC, Initiative, or Vision) — run drift resolution. Read both the SPEC and the parent. If the SPEC's scope, acceptance criteria, or constraints drift from the parent's goals or success criteria, apply a fix: either re-align the child (edit the SPEC) or update the parent.

Fix direction uses signals and content judgment: count prior drift decisions against the parent (more means parent is likely stale), plus assess which direction produces the better outcome. Apply the fix, then present the result for operator review: accept, modify, or revert. All outcomes are recorded as drift decisions via swain-session-state.sh record-decision.

README reconciliation nudge

When transitioning a Vision, Design, Journey, or Persona to a new phase, emit a soft signal if the transition changes the project's public-facing claims:

README.md may need updating to reflect this change.

This is informational, not blocking — the operator can dismiss it.

Trigger conditions — nudge when:

• A Vision transitions to Active (new direction) or Abandoned (dropped direction)
• A Design transitions to Active (new interaction model, data architecture, or system contract) or Superseded (replaced)
• A Journey transitions to Active (new user flow) or Abandoned (deprecated path)
• A Persona transitions to Active (new audience) or Abandoned (dropped audience)

Brainstorming context — when the brainstorming skill runs for a project that has a README but no artifacts (or a thin artifact tree — fewer than 3 Active Visions, Designs, Journeys, or Personas combined), it should use the README as the starting context for Socratic exploration. The README's claims, audience, and described behavior seed the brainstorming conversation instead of starting from scratch.

Detection for brainstorming context:

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
ACTIVE_COUNT=$(find "$REPO_ROOT/docs" -path "*/Active/*" -name "*.md" 2>/dev/null | grep -cE "(VISION|DESIGN|JOURNEY|PERSONA)" || echo "0")
HAS_README=$( [ -f "$REPO_ROOT/README.md" ] && echo "yes" || echo "no" )

If HAS_README=yes and ACTIVE_COUNT < 3, pass README content to brainstorming as primary context.

Trove integration

During research phase transitions (Spike Proposed -> Active, ADR Proposed -> Active, Vision/Epic creation), check for existing troves and offer to link or create one. Read references/trove-integration.md for the full hook, trove scanning, and back-link maintenance procedures.

Execution tracking handoff

When implementation begins on a SPEC, invoke swain-do. Read references/execution-tracking-handoff.md for the four-tier tracking model, swain-do: required frontmatter field, intent triggers, and coordination artifact decomposition.

GitHub Issues integration

SPECs link to GitHub Issues via the source-issue frontmatter field. During phase transitions on linked SPECs, post comments or close the issue. Read references/github-issues-integration.md for promotion workflow, transition hooks, and backend abstraction.

Status overview

For project-wide status, progress, or "what's next?" queries, defer to the swain-session skill (it aggregates swain chart + tk + git + GitHub issues). For artifact-specific graph queries, use bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/chart.sh" — see references/specgraph-guide.md. The default output is a vision-rooted hierarchy tree; lenses (ready, recommend, debt, unanchored, etc.) filter and annotate the tree for different decision contexts.

Auditing artifacts

When the user requests an audit, read references/auditing.md for the full two-phase procedure (pre-scan + parallel audit agents including ADR compliance). Include an unanchored check pass: run bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/chart.sh" unanchored and report any artifacts without Vision ancestry as domain-level findings alongside alignment and ADR compliance results.

Implementation plans

Implementation plans bridge declarative specs and execution tracking. When implementation begins, read references/implementation-plans.md for TDD methodology, superpowers integration, plan workflow, and fallback procedures.

Scenario modeling

Scenario modeling is point-in-time impact analysis via git branching. The operator surfaces a question; the agent does everything. The operator never touches git.

The goal is to produce real edits on disk so git diff can surface the impact — changed artifacts, shifted roadmaps, and code references. The operator's only role is to review the impact and say "accept" or "reject."

This is a pure LLM procedure within swain-design. No separate skill, no scripts, no new CLI flags.

Trigger phrases

Activate scenario modeling when the operator says any of:

• "what if ADR-X were active"
• "what if we'd chosen Y instead of Z"
• "compare the tree under different ADR assumptions"
• "scenario: "
• "show me the impact of switching from A to B"

Procedure

1. Confirm intent. Briefly confirm the counterfactual assumption with the operator: "You want to see the artifact tree if ADR-046 were Active instead of Superseded?"

2. Branch.

   git checkout -b scenario/<kebab-case-name>
   

   Derive the name from the question (e.g., scenario/adr-046-active, scenario/pre-helm-refactor).

3. Identify the edit set. Walk the supersession graph from the named ADR or decision. For each artifact the flip touches, trace:

   • Artifacts that superseded it (also need flips).
   • Designs paired with the ADR (e.g., ADR-048 ↔ DESIGN-033).
   • Initiative and epic roadmaps that reference the flipped artifacts in linked-artifacts, depends-on-artifacts, or addresses arrays.

   Use swain chart output and the graph cache to resolve references.

4. Apply edits. Edit frontmatter on the scenario branch:

   • Flip status on the primary ADRs and their paired designs.
   • Update back-references in initiative/epic frontmatter arrays (linked-artifacts, etc.).
   • Do not edit prose bodies — frontmatter fields only.
   • Do not create or delete artifact files.

5. Validate. Run swain chart on the branch. Confirm the tree resolves without broken references. Present the alternative tree to the operator.

6. Surface impact. Run and summarize:

   git diff trunk...HEAD --stat
   git diff trunk...HEAD -- docs/ | grep -oP 'sourcecode-refs:\K[^\n]+'
   

   Present to the operator: number of artifacts changed, which initiative/epic roadmaps shifted, source code paths referenced.

7. Offer actions:

   • Accept: merge to trunk and commit.
   • Compare side by side: add a trunk worktree, run swain chart in both, present the diff.
   • Reject: delete the branch.

Operator feedback loops

If the operator says "also flip DESIGN-034" or "don't touch INITIATIVE-005," adjust the edit set and re-validate. Iteration is conversational — the operator steers, the agent acts.

Cleanup

The agent handles both outcomes:

• Accepted: git checkout trunk && git merge scenario/<name>, commit, push.
• Rejected: git branch -D scenario/<name>.

Example

Operator: "what if ADR-046 had stood?"

1. Branch: git checkout -b scenario/adr-046-active
2. Edit set: flip ADR-046 → Active, ADR-048 → Superseded, DESIGN-032 → Active, DESIGN-033 → Superseded. Update INITIATIVE-018 roadmap to swap DESIGN-033 for DESIGN-032.
3. Validate: swain chart — passes.
4. Impact: 4 artifacts changed, 1 initiative roadmap shifted, 2 source code paths referenced.

Reference material

Consult these files when a workflow step references them:

• Artifact relationships: references/relationship-model.md — ER diagram of type hierarchy and cross-references
• Lifecycle table format: references/lifecycle-format.md — commit hash stamping convention
• Index maintenance: references/index-maintenance.md — list-<type>.md refresh rules
• Tooling: Scripts live in scripts/ (relative to this skill's directory). See references/specwatch-guide.md, references/specgraph-guide.md, references/adr-check-guide.md for details.

Session bookmark

After state-changing operations, update the bookmark: bash "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents/bin/swain-bookmark.sh" "<action> <artifact-ids>" --files <paths>