Claude Code Design

Calibration: Tier 3 (High-effort recommended) — run on Opus 5 or Sonnet 5 (both default to high on Claude API and Claude Code, the recommended starting point). Step up to xhigh for long-horizon or particularly demanding runs. Quality degrades at low effort and on legacy models (Sonnet 4.6, Opus 4.8 fallback-graceful). See repository README for model compatibility.

Design Claude Code prompts and environments for production delivery projects. This Skill produces paste-ready artifacts — CLAUDE.md drafts, agent topology recommendations, scope-authorization frameworks, halt-and-escalate triggers, Skills/Hooks/MCP plans, and chat-to-Code handoff specifications. The methodology is grounded in tested patterns from a production CC deployment 2026-05-04 (27/27 ship, 18-WI evolution arc), Anthropic primary documentation, and named practitioner consensus (rosmur, obra/superpowers, alexop.dev, Shrivu Shankar, Marc Nuri).

This Skill operates in both chat-based design conversations (CP) and Claude Code (CC). In CP, it focuses on DESIGN, EVOLVE, RESEARCH, and TEMPLATE modes — design conversations, brainstorming, scaffolding, cross-project synthesis. In CC, the same modes work alongside REMEDIATE — the closed-loop mode that consumes a HYGIENE_REPORT.md produced by rootnode-repo-hygiene and produces (then executes, after explicit user acceptance) an EXECUTION_PLAN.md against the actual repo.

The Skill produces design artifacts in all modes. REMEDIATE mode is the only mode that also executes — and only its Phase 2, gated by an explicit user acceptance step that follows Phase 1 plan generation.

Activation

Auto-invoke on these trigger phrases. Column two names the mode the phrase routes to; confirm mode in one line at the start of the response.

Trigger phrase	Mode
"design CC for X"	DESIGN
"build CC environment"	DESIGN
"design CLAUDE.md"	DESIGN
"build a CC prompt" / "design a CC prompt for X"	DESIGN
"write a session prompt"	DESIGN
"give me a CLAUDE.md skeleton"	TEMPLATE
"we hit X friction in CC" / "the agent keeps doing Y"	EVOLVE
"should we adopt Y for CC" / "is Z worth using"	RESEARCH
"remediate the hygiene findings" / "close the loop on the report"	REMEDIATE

Delegation architecture

See references/cc-delegation-patterns.md for the role table, delegation briefs, Builder-to-Refuter loop, cap mechanisms, and the orchestrator/worker model. See references/cc-methodology-patterns.md for the verification topology and the agent-warranted test.

Important

Source discipline is non-negotiable. Every substantive claim about Claude Code behavior, MCP capability, or agent design pattern carries an inline source tag — [Anthropic docs], [<project> §X], [practitioner: name + artifact], or [speculation]. Tier 5 community sources are signal-only, never authoritative. See references/source-grading-and-tagging.md for the full hierarchy.

Generalizable vs. project-specific tagging. Every pattern drawn from a production CC deployment or other working examples is tagged inline as [generalizable], [project-specific], or [generalizable structure, project-specific content] with a one-line basis for the tag. Default to "may be project-specific until proven otherwise." See references/source-grading-and-tagging.md.

Tool/agent gap discipline. Never recommend a tool, MCP server, agent role, Skill, or hook without identifying the specific operational gap it fills. "Adding X for completeness" is anti-pattern. The question is always: what concrete failure mode does this address, and is that failure mode actually present in the user's situation?

The 7-layer placement rule comes first. Before specifying CLAUDE.md content, place each piece of content in the correct mechanism (CLAUDE.md / rules / Skills / subagents / hooks / MCP / settings). Misplaced content is the single most common cause of agent unreliability. See references/cc-environment-design-patterns.md.

Halt-and-escalate triggers are mandatory. Every DESIGN MODE deployment plan and every EVOLVE MODE update specifies explicit halt-and-escalate triggers. Autonomous CC iteration without halt triggers is the failure mode that produces the worst outcomes.

Output complete files. Per the user's working preference: produce complete files (CLAUDE.md, agent specs, design briefs), never diffs or partial sections — except in EVOLVE MODE where the existing file is in context and a section-level delta is the deliverable.

---

How to Use This Skill

Step 1 — Determine the mode

Five modes. Mode is inferred from the request signal; confirm in one line at the start of the response, then proceed. Some modes have natural surfaces (CP vs CC) but all 5 work in both — surface emphasis is a tendency, not a constraint.

Mode	Trigger signals	Output	Natural surface
DESIGN	"design CC for X", "build CC environment for Y", "scaffold a Claude Code repo", new project scaffolding	Complete deployment plan: CLAUDE.md draft + initial prompt + agent topology + scope rules + halt triggers + handoff point + runtime tooling recommendations	CP for greenfield (no repo yet); CC for in-repo new work (new module, new service inside existing repo)
EVOLVE	"we hit X friction in [project]", "the agent keeps doing Y", "this pattern emerged across sessions", "we need to add Z to existing CLAUDE.md"	Section-level deltas to existing artifacts: CLAUDE.md updates as inserts/replacements, new agent specs, new scope rules, change-log entry template, regression-sweep specification	Both — speculative friction in CP; in-repo friction in CC where the deltas can be applied directly
RESEARCH	"should we adopt X", "is Y worth using", "what's the landscape for Z", "what tool fills the gap of W"	Structured assessment: gap definition, candidates with source class, fit analysis, decisive recommendation, speculative notes	CP-leaning (web search heavy; polluting a delivery repo's CC session is anti-pattern)
TEMPLATE	"give me a CLAUDE.md skeleton for X-class projects", "build a reusable prompt template", "produce an agent role template I can reuse"	Complete template with placeholders in {BRACKETED_CAPS} + usage notes + filing recommendation	CP-leaning (cross-project synthesis is wrong context inside one delivery's CC)
REMEDIATE	"remediate the hygiene findings", "fix the audit issues", "close the loop on the report", HYGIENE_REPORT.md exists in repo and user invokes this Skill without naming a different mode	Phase 1: EXECUTION_PLAN.md written to repo (steps + validation per step + summary + addressed-findings list). User reviews and explicitly accepts via one of three approval forms (blanket / fragmented / conditional). Phase 2: Skill executes the accepted plan step-by-step, validates each step, halts on failure	CC-native (must read HYGIENE_REPORT.md from repo; Phase 2 writes to repo). CP-only invocation requires pasted findings and produces plan without executing

Step 2 — Check for or generate the design brief

The Skill operates from a structured design brief for the delivery project. The brief captures: mission, current state, tech surface, governance state, runtime tooling state, authority constraints, verification surface, and known friction. It lives at {project_code}_design_brief.md in the delivery project's KFs (or filesystem if working at the repo level).

On first invocation in a new project: conduct the 5-question interview (mission / current state / tech surface / authority constraints / verification surface), serialize to brief format, offer as a downloadable artifact, and recommend adding to the delivery project's KFs. See references/cc-methodology-patterns.md §"The design brief workflow" for the full interview script and brief format.

On subsequent invocations: read the existing brief as input. If the brief is stale (front-matter last_updated > 90 days, or current state field is wrong), surface this and offer to refresh.

If brief generation is over-investment for the request: small focused asks (one CLAUDE.md section, one prompt rewrite, one agent spec review) can proceed without a brief. Use judgment — the brief exists to ground multi-faceted design work, not to bottleneck simple tasks.

Step 3 — Apply mode-specific reasoning

Read the relevant references before producing output. Each mode draws from a specific subset:

• DESIGN MODE consults: cc-methodology-patterns, cc-environment-design-patterns, cc-skills-and-hooks-composition, chat-to-code-handoff-patterns. Optional: cc-prompt-design-patterns (if the deliverable includes a CC initial prompt or session prompt); cc-delegation-patterns (if the deployment plan involves subagents, model tiering, parallel work, or delegation caps).
• EVOLVE MODE consults: cc-anti-patterns (diagnose the friction), cc-methodology-patterns (find the relevant pattern), cc-environment-design-patterns or cc-skills-and-hooks-composition (apply the fix at the correct layer); cc-delegation-patterns (when friction involves delegation, context bloat, or subagent behavior).
• RESEARCH MODE consults: source-grading-and-tagging (apply the source authority hierarchy), cc-anti-patterns (verify the candidate doesn't introduce known anti-patterns).
• TEMPLATE MODE consults: cc-methodology-patterns (the structural pattern being templated) and the relevant CC pattern reference for the artifact class.
• REMEDIATE MODE consults: remediate-mode-execution (the full Phase 1 + Phase 2 protocol, three approval forms, plan format, validation grammar, conflict-resolution rules), cc-anti-patterns (the per-pattern fix recipes — when consuming a finding tagged with a canonical reference like §4.2, look up the structural fix here), cc-environment-design-patterns (the 7-layer placement framework — used to validate that proposed fixes land in the correct mechanism). Schema: schema/execution-plan.schema.json defines the EXECUTION_PLAN.md structure.

Step 4 — Apply output standards

• File naming. CLAUDE.md drafts for delivery projects use {delivery_project_code}_CLAUDE.md (the destination project's prefix). Agent role specs use {delivery_project_code}_agent_{name}.md. Design briefs use {project_code}_design_brief.md and live in the delivery project's KFs. Templates produced for cross-project reuse use {project_code}_template_{descriptor}.md or shared_template_{descriptor}.md.
• Markdown for CLAUDE.md drafts. CLAUDE.md is consumed by Claude Code, which expects Markdown. Use ## and ### headers, fenced code blocks, tables for matrices. Do not use XML tags inside CLAUDE.md drafts — those are for system prompts, which CLAUDE.md is not.
• Halt triggers and scope authorization. Every DESIGN MODE plan and every EVOLVE MODE update explicitly includes both. Ultracode stays off by default and is opted into per task (see references/cc-delegation-patterns.md §6 for the rationale). Non-negotiable.
• Assert the END state, never the PRE state. Non-negotiable. See references/cc-prompt-design-patterns.md § Assert the END state for the rule, the LF-normalize corollary, and the observed instances. [generalizable]
• Delegation briefs are complete or absent. Any role the plan names carries a model, an effort level, a tool set, an isolation mode, and a return contract including a numeric length cap. Verification clauses name external artifacts; never instruct an agent to re-check its own output.
• Token budgets. When the deployment plan involves sub-agents or context management, include explicit token budgets that sum to the model's context window. A context plan without numbers is not a context plan.
• Source-tagged claims. Every substantive technical claim carries an inline source tag.
• Output length. When the deliverable is a design document, CLAUDE.md draft, or session prompt, target ≤200 lines per artifact. Multi-file designs may exceed this per-file if each file is self-contained.
• CC-prompt-specific output discipline. When the deliverable is a CC initial prompt, session prompt, or autonomous prompt, see references/cc-prompt-design-patterns.md § Output discipline for CC prompts for additional output standards (shell-agnostic syntax, pre-flight Skill enumeration, continuation-phrase ambiguity gate, forward-state-aware authoring, competing hypotheses and instrument fit for diagnostic prompts).

Step 5 — Surface the chat→Code handoff explicitly

Every DESIGN MODE output identifies the chat→Code handoff point: at what milestone does design work end and Claude Code execution begin? Build the round-trip into the design — also specify what conditions trigger a return to chat (pattern emerges across 2+ sessions, new tool enters consideration, friction with existing CLAUDE.md surfaces, reusable artifact extraction needed).

See references/chat-to-code-handoff-patterns.md for the readiness signals, artifact bundle composition, and round-trip pattern.

Step 6 — Recommend runtime tooling only if it fills a specific gap

The rootnode runtime Skills (handoff-trigger-check, critic-gate, mode-router, profile-builder) are external tooling. Runtime Skills are thin governance over thick workflow systems — they compose with existing deployment tooling rather than replacing it. When a deployment plan involves autonomous Claude Code execution, evaluate whether one or more of these Skills fills a specific operational gap. If yes, recommend with one-line rationale per Skill and note the deployment target (CP-side runs in the design Project itself; CC-side deploys into the delivery project). If no, do not mention them — the Skill's methodology stays decoupled from any specific tooling.

For scheduled and automated CC deployments, consult the Routines section in references/cc-methodology-patterns.md.

---

Reference Files

Read the relevant files based on the mode (Step 3 above). Each file is standalone-readable; load only what the current task needs.

File	When to read
references/cc-methodology-patterns.md	Always for DESIGN, EVOLVE, TEMPLATE modes. Contains the project-agnostic distillation of the methodology — authority matrix pattern, scope-authorization framework, change_log discipline, additive evolution, agent-warranted test, design brief workflow.
references/cc-prompt-design-patterns.md	When the deliverable includes a CC initial prompt, session prompt, autonomous prompt, or subagent delegation prompt. Patterns organized by use case.
references/cc-delegation-patterns.md	When the deployment plan involves subagents, model tiering, parallel work, or delegation caps. Role table (model / effort / tools / isolation / return), the eight-field delegation brief, the Builder-to-Refuter loop, orchestrator context hygiene, and the cap mechanism table.
references/cc-environment-design-patterns.md	Always for DESIGN mode. Reference for EVOLVE and REMEDIATE when placement decisions are involved. Contains the 7-layer decomposition framework — what content goes in CLAUDE.md vs .claude/rules/ vs Skills vs subagents vs hooks vs MCP vs settings — with selection criteria for each layer.
references/cc-skills-and-hooks-composition.md	When the deployment plan involves Skills, hooks, or both. Skills frontmatter reference, auto-activation patterns, hooks-vs-skills decision tree, the verification "iron law" pattern.
references/cc-anti-patterns.md	Always for REMEDIATE mode (per-pattern fix recipes). Reference for EVOLVE when diagnosing friction. Reference for DESIGN as a "things to avoid creating" checklist. The 17-pattern catalog using canonical numbering from root_AGENT_ANTI_PATTERNS.md and root_CC_ENVIRONMENT_GUIDE.md §10, with structural signature, cause, and fix per pattern. Note: anti-pattern scanning is rootnode-repo-hygiene's territory; this Skill consumes the catalog as a fix-recipe library, not as an audit checklist.
references/chat-to-code-handoff-patterns.md	Always for DESIGN mode. When designing how chat work transitions to CC execution. Handoff readiness signals, artifact bundle composition, the round-trip pattern.
references/source-grading-and-tagging.md	Always for RESEARCH mode. Reference whenever a claim needs source backing. The 5-tier source authority hierarchy and the generalizable-vs-project-specific tagging discipline.
references/remediate-mode-execution.md	Always for REMEDIATE mode. Full protocol for Phase 1 (plan generation) + Phase 2 (execution): three approval forms (blanket / fragmented / conditional), critic-gate composition (required / optional), action types, validation grammar, conflict-resolution rules, halt semantics, HYGIENE_REPORT.md input format expectations, EXECUTION_PLAN.md output format.
references/troubleshooting.md	When the Skill output isn't landing or the user signals a specific problem with how the Skill is being invoked or what it produced.

The brief schema lives at schema/cc-design-brief.schema.json. The execution-plan schema lives at schema/execution-plan.schema.json.

---

Examples

Example 1: DESIGN MODE — new CC deployment for a delivery project

Input: "I'm starting a Claude Code project for a Python CLI tool that ingests CSV files and produces a normalized SQLite DB. Help me design CLAUDE.md and the initial agent topology."

Actions:

1. Confirm mode: "DESIGN MODE — building a new Claude Code deployment plan."
2. Check for a design brief; none exists for this project. Conduct the 5-question interview (mission, current state, tech surface, authority constraints, verification surface).
3. Serialize the brief and offer it as a downloadable artifact. Recommend adding to the project's KFs.
4. Apply the 7-layer placement rule (cc-environment-design-patterns) — for a single-script CLI, most concerns collapse to CLAUDE.md + 1-2 hooks; no Skills or subagents warranted yet.
5. Apply the agent-warranted test (cc-methodology-patterns §2) — single-loop iteration; recommend single-agent loop, no sub-agents.
6. Draft the CLAUDE.md with the 5 required sections (R1-R5) and skip W1-W3 (warrant tests fail at day 1).
7. Specify scope authorization (in-scope / in-scope-with-notification / out-of-scope) and halt-and-escalate triggers.
8. Identify the chat→Code handoff point: "Once CLAUDE.md is in place, hand off to CC for the first iteration."
9. Evaluate runtime tooling — handoff-gate is overkill for a single-script project; recommend deferring.

Output: Design brief artifact + CLAUDE.md draft (Markdown, paste-ready, < 200 lines) + initial CC prompt + scope authorization clauses + halt triggers + handoff point + a "what to revisit when the project grows" note.

Example 2: EVOLVE MODE — friction in an existing CC project

Input: "In the RT project, the agent keeps re-reading the same files and forgetting what it learned in the previous session. It's also editing files in the migrations directory that it shouldn't touch."

Actions:

1. Confirm mode: "EVOLVE MODE — incremental update to the RT CC project."
2. Read the existing RT design brief (or ask if not in context).
3. Diagnose: two distinct frictions. (a) Cross-session memory — likely missing CLAUDE.md notes about durable file artifacts (cc-anti-patterns §4.1 transcript dump if there's also no change_log; or §4.14 stale CLAUDE.md if the relevant info exists but is wrong). (b) Migrations directory edit — §4.4 enforcement-as-preference (rule was probably in CLAUDE.md as advice, not backed by a hook).
4. Design fixes — additive, narrow detection per cc-methodology §4. (a) Add a change_log discipline section to CLAUDE.md + introduce a session-handoff Skill or rule pattern. (b) Add a PreToolUse hook blocking writes to migrations/** — prompt-level rule moves to enforcement layer.
5. Specify regression sweep: re-run the 3 most recent deployment scenarios after the changes; verify no false-positive blocks from the hook.

Output: Section-level deltas to RT_CLAUDE.md (insert + replacement) + new hook config in .claude/settings.json (paste-ready) + change-log entry template + regression-sweep specification.

Example 3: REMEDIATE MODE — closing the loop on hygiene findings

Input: User invokes the Skill in CC. HYGIENE_REPORT.md exists in the repo, produced by rootnode-repo-hygiene in a prior session. User says: "remediate the hygiene findings."

Actions (Phase 1 — plan generation):

1. Confirm mode: "REMEDIATE MODE — Phase 1 (plan generation). I'll read HYGIENE_REPORT.md and produce EXECUTION_PLAN.md for your review."
2. Read HYGIENE_REPORT.md from repo root. Parse findings. Filter to in-scope: Cat 11–14 structural findings + 7-layer leak findings. Cat 1–10 direct-cleanup findings are out of scope for REMEDIATE — those route to repo-hygiene Phase 2; surface this routing in the response and skip them.
3. For each in-scope finding, look up the structural fix in cc-anti-patterns.md (e.g., a finding tagged §2.1 bloated CLAUDE.md → trim to <200 lines + extract procedures to Skills + extract path-rules to .claude/rules/).
4. Group findings by file affected; resolve conflicts where two findings touch the same file (e.g., both §2.1 and §4.14 touch CLAUDE.md — sequence: trim first, then date-stamp).
5. Validate proposed fixes against the 7-layer placement framework (cc-environment-design-patterns.md §1) — does each fix land in the correct mechanism?
6. Build EXECUTION_PLAN.md per the schema (schema/execution-plan.schema.json): plan metadata, addressed-findings list (using addresses_finding: F-X.Y references), ordered steps with action/target/payload/risk/validation per step, pre-flight + post-flight validation.
7. Write EXECUTION_PLAN.md to repo root.
8. Surface plan summary: "Plan addresses 7 findings, modifies 4 files, creates 2 new files, adds 1 hook. Review EXECUTION_PLAN.md and respond with one of three approval forms — blanket ('execute'), fragmented ('execute steps 1, 2, 4'), or conditional ('execute medium-and-low risk; halt on high-risk') — to proceed to Phase 2, or describe changes you want."
9. Halt and wait for explicit user acceptance. No execution in Phase 1.

Actions (Phase 2 — execution, only after explicit user acceptance):

1. Confirm mode: "REMEDIATE MODE — Phase 2 (execution). Walking the plan step-by-step."
2. Run pre-flight validation. Halt if any fails.
3. For each step in plan order (filtered by user's acceptance form — blanket walks all; fragmented walks named step IDs; conditional walks steps matching the risk predicate): a. (If critic_gate_threshold: required and rootnode-critic-gate installed) Submit step plan to critic-gate; APPROVE → continue; REQUEST_CHANGES → adjust + re-submit, cap 3 cycles; REJECT → halt. b. Apply the action (edit / create / delete / modify / run). c. Run the step's validation. d. If validation fails, halt and report the failure with current state. Do not auto-rollback. e. If validation passes, log step completion and continue.
4. Run post-flight validation.
5. Append a CHANGELOG entry summarizing the remediation cycle.
6. Report: steps approved / total, steps executed / approved, steps skipped (with reason), critic-gate review summary (if applicable), files modified, findings closed, any halts.

Output: Phase 1 produces EXECUTION_PLAN.md. Phase 2 produces actual repo changes + CHANGELOG entry + execution report. Both phases are gated by explicit user acceptance between them.

---

When to Use This Skill

Use this Skill when:

• The user is planning a new Claude Code deployment for a delivery project (DESIGN mode)
• An existing Claude Code project has surfaced friction or needs a new pattern added (EVOLVE mode)
• The user is evaluating a tool, MCP server, Skill, hook, or pattern for inclusion in a CC deployment (RESEARCH mode)
• The user wants a reusable artifact for cross-project use — CLAUDE.md skeleton, agent role template, scope-authorization clause template (TEMPLATE mode)
• The user has finished designing a deliverable in chat and is preparing the chat→Code handoff bundle
• A HYGIENE_REPORT.md exists in the repo and the user wants to close the loop on it (REMEDIATE mode)

REMEDIATE routing

REMEDIATE handles the structural and placement-decision portions of a hygiene report. Direct-cleanup findings have a different route. Use this routing rule when a HYGIENE_REPORT.md is in play:

Finding type	Owner	Action
Cat 1–10 (permissions, hooks, parent-vestige, stale code, Skills hygiene, etc.)	rootnode-repo-hygiene Phase 2	Direct cleanup. Run repo-hygiene Phase 2 against the [APPROVED] markers in the report. Not REMEDIATE territory.
Cat 11–14 (structural, including process-abstraction candidates)	rootnode-cc-design REMEDIATE	Structural fix recipes from cc-anti-patterns.md; conflict resolution; 7-layer placement validation.
7-layer leak findings	rootnode-cc-design REMEDIATE	Placement-decision required; route to REMEDIATE for layer reassignment.

Phase 1 surfaces this routing explicitly: when REMEDIATE is invoked on a report containing Cat 1–10 findings, the response names them as out-of-scope and recommends the repo-hygiene Phase 2 path before walking the in-scope findings.

Do NOT use this Skill when:

• The user wants to scan or audit an existing CC environment for hygiene issues / anti-patterns (use rootnode-repo-hygiene if available — that Skill produces the HYGIENE_REPORT.md this Skill consumes)
• The user wants to evaluate or score a chat prompt (use rootnode-prompt-validation if available)
• The user wants to compile a chat prompt or scaffold a chat Project (use rootnode-prompt-compilation if available)
• The user wants to audit a chat-based Claude Project's Custom Instructions or knowledge files (use rootnode-project-audit if available — different surface from CC)
• The user is doing non-CC delivery work that doesn't involve Claude Code at all

---

Troubleshooting

Skill produces generic CLAUDE.md not tailored to the project: The design brief was missing or thin. Generate the brief first via the 5-question interview before producing the deployment plan. The brief is the grounding artifact — without it, output drifts toward generic patterns.

Skill recommends rootnode runtime tooling for a project that doesn't need it: The tool/agent gap discipline was skipped. For each runtime Skill, name the specific operational gap it fills in this project. If you can't, don't recommend it. See SKILL.md Step 6.

Skill outputs are too long / context-heavy: TEMPLATE mode is a useful pressure release — extract the reusable structural pattern into a {project_code}_template_* or shared_template_* artifact and reference it from delivery-project briefs. Avoid re-deriving the same pattern across multiple project deployments.

Skill produces "vibe" recommendations without source backing: Source discipline was skipped. Every substantive claim about CC behavior, MCP capability, or pattern needs an inline source tag. See references/source-grading-and-tagging.md. If a tag is [speculation], that's acceptable as long as it's marked.

User asks for a quick answer and the brief workflow feels heavy: Skip the brief for narrow asks (one section, one prompt, one review). The brief grounds multi-faceted design work, not single-shot questions. Use judgment.

A CC prompt halted because its stated starting state was wrong: The prompt asserted a pre-state ("expect N lines", "the file currently contains X") that went stale between authoring and execution — commonly because the replacement file was already staged, or another pass touched the tree. The guard was correct; the assertion was not. Fix at the prompt, not at the guard: make the starting condition report-only and bind the halt to a post-action digest. See Step 4, Assert the END state, never the PRE state. Resolve the specific stuck run by giving CC the expected sha256sum and instructing it to skip the already-satisfied step on an exact match.

The orchestrator session fills up even though the work was delegated: Delegation isolates the agent's working set, not its report. An agent that returns a full file, a full test log, or a full diff moves the flood one layer out and then back in. Fix at the brief: set a return-length cap, require locations and findings rather than contents, and route oversized results to a scratch file the next agent reads. See references/cc-delegation-patterns.md.

User asks the Skill to "audit my CC project" — the Skill claims this is rootnode-repo-hygiene's territory but the user disagrees: Different verbs. rootnode-repo-hygiene scans for hygiene issues against the 15-anti-pattern catalog and produces findings. This Skill consumes those findings (REMEDIATE mode) or designs new artifacts from scratch (DESIGN/EVOLVE/TEMPLATE). If the user wants a "review" of a draft CLAUDE.md during design work (not against a deployed environment), that fits EVOLVE mode framed as "review and propose deltas." If the user wants the deployed-environment scan, route to rootnode-repo-hygiene.

REMEDIATE mode invoked but no HYGIENE_REPORT.md exists in the repo: The Skill cannot proceed. Two options: (a) ask the user to run rootnode-repo-hygiene first to produce the report, then re-invoke REMEDIATE; (b) accept findings pasted directly into the conversation, but recommend the file-based workflow for repeatability. Plan generation without source findings is anti-pattern — the loop only closes when there's a defined input to remediate against.

REMEDIATE Phase 2 step validation fails: Halt. Do not auto-rollback. Report which step failed, which validation, and the current repo state (what was completed before the halt). The user decides whether to manually revert, fix forward, or re-plan. Safety-via-halt is the v2 discipline; rollback semantics are out of scope.

Two REMEDIATE conflict resolutions seem reasonable for the same finding cluster: Phase 1 still produces a plan, but the conflicting steps are marked requires_user_choice: true and Phase 2 will halt at those steps for explicit choice. Don't silently choose one — the conflict surfaced because the proper sequencing isn't unambiguous, and that's a user decision, not a Skill decision.

EVOLVE mode keeps re-deriving the same fix: This is a signal the fix should be promoted to a TEMPLATE. After the second invocation of the same pattern, extract it as a template and reference from future EVOLVE outputs.