Patterns 156–206 · 2026-07 Harvest

The 51 patterns added in the 2026-07 harvest — from superpowers, claude-plugins-official, and mattpocock/skills, including the new Skill Authoring category. Full write-ups are in progress; this page is the static, searchable index of names and problems.

Interactive version: pattern browser · Patterns 1–155: catalog-index · 中文版：模式 156–206

156. No-Pre-Judging Reviewer Dispatch

A controller dispatching a reviewer subagent can quietly bias it—pre-rating severity or saying what not to flag—defeating independent review.

Category: advanced-quality · Source: superpowers · Confidence: high

157. Rationale-as-Claim

Implementers embed self-justifications (‘left it per YAGNI’); a reviewer that accepts them lets the author grade their own work.

Category: advanced-quality · Source: superpowers · Confidence: high

158. Plan-Mandated Defect Is Still a Finding

When a plan explicitly mandates something the rubric calls a defect, agents skip the finding as ‘the plan chose this’.

Category: advanced-quality · Source: superpowers · Confidence: high

159. Diff-Scoped Review with Named-Risk Escape Hatch

Reviewers waste context re-reading whole files, yet some cross-cutting risks genuinely need looking beyond the diff.

Category: advanced-quality · Source: superpowers · Confidence: medium

160. Reviewer Does Not Re-Run the Implementer’s Tests

Reviewers re-running the full suite duplicate the implementer’s TDD evidence and burn context; run a focused test only on specific doubt.

Category: advanced-quality · Source: superpowers · Confidence: medium

161. Explicit-Model-Required Dispatch Guardrail

Dispatches that omit the model field silently inherit the session’s most expensive model, defeating cost-optimized routing.

Category: advanced-orchestration · Source: superpowers · Confidence: medium

162. File-as-Inter-Agent-Bus Handoff

Everything pasted into a dispatch and everything a subagent prints stays in the orchestrator’s context, re-read every turn.

Category: advanced-orchestration · Source: superpowers · Confidence: medium

163. Anti-Pasted-History Dispatch

Controllers pad each dispatch with prior-task summaries (‘state after Tasks 1-3’)—one real dispatch hit 42k chars, 99% history.

Category: advanced-orchestration · Source: superpowers · Confidence: medium

164. Batch Remediation: One Fixer for All Findings

Dispatching a fix subagent per finding makes each rebuild context and re-run suites; the fix wave can cost more than all tasks combined.

Category: advanced-orchestration · Source: superpowers · Confidence: medium

165. Reviewer-Gate Task Right-Sizing

No crisp criterion for task boundaries: split too fine drowns reviewers, too coarse bundles separately-approvable changes.

Category: structural-scaffolding · Source: superpowers · Confidence: high

166. Verbatim Global-Constraints Block

Project-wide requirements get scattered/paraphrased/dropped across tasks; hoist them once, verbatim, as a block all tasks inherit and reviewers use as a lens.

Category: structural-scaffolding · Source: superpowers · Confidence: high

167. Per-Task Interfaces Block (Consumes/Produces)

Isolated subagents drift on signatures (Task 3 defines clearLayers, Task 7 calls clearFullLayers) with no view of neighbors.

Category: agent-orchestration · Source: superpowers · Confidence: high

168. Match the Form to the Failure

Authors reach for prohibition lists for every problem, but the form that fixes a discipline violation backfires on wrong-shaped or omitted output.

Category: skill-authoring · Source: superpowers · Confidence: high

169. Wording Micro-Test with No-Guidance Control

Full pressure runs are slow per iteration; cheap wording-level A/B with a no-guidance control stop-condition and cross-rep variance as the binding signal.

Category: skill-authoring · Source: superpowers · Confidence: high

170. Just-In-Time Capability Offer

Offering an optional token-heavy capability upfront wastes turns on a feature that may never be needed; defer to first genuine benefit.

Category: execution-control · Source: superpowers · Confidence: medium

171. Action-Verb Tool Abstraction

A skill body hardcoding one harness’s tool names breaks on every other; author in capability verbs, bind names in a per-harness adapter.

Category: claude-code-platform · Source: superpowers · Confidence: medium

172. Harness-Conditional Reference Dispatch

An always-loaded bootstrap pays token tax every turn; carry only the universal rule + a dispatch table, self-identify harness and load one adapter file.

Category: claude-code-platform · Source: superpowers · Confidence: medium

173. Capability-Absent Anti-Fabrication Fallback

On a harness lacking a capability, the agent hallucinates the missing tool call; forbid fabrication, route to a named documented degradation.

Category: claude-code-platform · Source: superpowers · Confidence: medium

174. Reachability-Scoped Delta

Migration reports drown in generic breaking changes that never apply here; scope to the intersection of general knowledge and what this tree actually calls.

Category: advanced-quality · Source: claude-plugins-official · Confidence: high

175. Tool-Ran Tri-State Provenance

Agents claim coverage from a tool merely installed but never run; track present / runnable-here / actually-ran and never fold findings unless it ran.

Category: advanced-quality · Source: claude-plugins-official · Confidence: high

176. Differential Baseline-Equivalence Oracle

Proving a version uplift changed nothing is hard; run the same suite on both runtimes, diff per-test—goal is ‘no behavior changed’, not ‘all green’.

Category: advanced-workflow · Source: claude-plugins-official · Confidence: high

177. Capability Smoke-Test Over Presence-Check

‘command -v tool’ proves a binary exists, not that it can process this codebase; exercise it on one real input (syntax-only compile).

Category: advanced-quality · Source: claude-plugins-official · Confidence: high

178. Derivation-Order Staleness Check

A downstream artifact silently goes stale when an upstream one it derives from is regenerated; compare mtimes along the derivation DAG.

Category: advanced-workflow · Source: claude-plugins-official · Confidence: high

179. Self-Describing Artifact with Embedded State

A living status page must report only what changed since last time, across machines; embed a machine-readable state block the next run diffs against.

Category: advanced-io-domain · Source: claude-plugins-official · Confidence: high

180. Dependency-Encoding Numbering Scheme

Drawing a dependency diagram is heavy and drifts; encode order in a two-part item number (X blocked-on-prior-stage, Y parallel-within-stage).

Category: structural-scaffolding · Source: claude-plugins-official · Confidence: medium

181. Cross-Service Credential-Routing Detection

A credential read is malicious only when routed to a different service than it belongs to; judge ownership by name/storage, not claimed use.

Category: advanced-safety · Source: claude-plugins-official · Confidence: high

182. Whole-Payload Review Scope (Ship != Load)

Review anchored to the loaded surface misses code that ships to disk but isn’t auto-loaded (a git install clones the whole repo, incl. .claude/).

Category: advanced-safety · Source: claude-plugins-official · Confidence: medium

183. Untrusted-Content-as-Markup Escaping

Fetched text embedded into generated HTML is untrusted markup too; a PR title with </script> can break out—entity-encode wherever it lands.

Category: advanced-safety · Source: claude-plugins-official · Confidence: medium

184. Transitive Taint of Derived Artifacts

An intermediate artifact generated from untrusted input inherits its taint; follow its structure but never execute imperatives embedded in it.

Category: advanced-safety · Source: claude-plugins-official · Confidence: medium

185. No-Real-Secret-in-Generated-Fixture

A credential literal in source can leak into a committed fixture/config default; use same-shape fakes + env placeholders, read real secrets from env only.

Category: advanced-safety · Source: claude-plugins-official · Confidence: medium

186. In-Loop Self-Correction Before Surfacing

Post-hoc review after the user sees output forces a round-trip; a stop-hook fast reviewer feeds high-severity findings back for in-turn repair.

Category: advanced-quality · Source: claude-plugins-official · Confidence: medium

187. Inline-Justification Suppression of Findings

An automated reviewer re-flags the same known-safe line every run; let a human waive it via a co-located inline comment the reviewer honors.

Category: advanced-quality · Source: claude-plugins-official · Confidence: medium

188. Immutable-Slug Identity with Rename Redirect Map

A registry entry’s name is an immutable slug—renaming breaks installs; edit displayName for labels, add a renames map for unavoidable renames.

Category: advanced-safety · Source: claude-plugins-official · Confidence: medium

189. Leading Word (Leitwort)

Anchor a whole region of agent behavior in the fewest tokens by reusing a compact pretrained concept (lesson, fog of war, tracer bullets) as a behavioral hook.

Category: skill-authoring · Source: mattpocock · Confidence: high

190. Invocation Economics (Context vs Cognitive Load)

Decide model-invoked vs user-invoked by trade-off: a description sits in context every turn (context load); user-invoked spends the human’s memory (cognitive load).

Category: skill-authoring · Source: mattpocock · Confidence: high

191. Sequence Cut

Visible downstream steps tug attention toward ‘being done’, so the agent rushes the current step; hide them across a real context boundary.

Category: skill-authoring · Source: mattpocock · Confidence: high

192. No-Op Test (Behavior-vs-Default Pruning)

Skills accumulate lines the model already obeys by default; hunt no-ops sentence by sentence—settle by running the skill, not by debate.

Category: skill-authoring · Source: mattpocock · Confidence: medium

193. Two-Axis Completion Criterion (Checkable + Exhaustive)

Vague ‘done’ lets the agent declare victory early or do shallow work; make it checkable (guards rushing) and exhaustive (guards thoroughness).

Category: skill-authoring · Source: mattpocock · Confidence: medium

194. Controlled Vocabulary with Banned Synonyms

Term drift (component/service/API/boundary) dissolves the shared language a skill depends on; pin exact terms with a per-term ‘Avoid’ ban list.

Category: skill-authoring · Source: mattpocock · Confidence: medium

195. Rejected-Framings Inoculation

The agent adopts a plausible-but-wrong mental model of a concept, corrupting everything downstream; pre-empt and refute competing definitions.

Category: skill-authoring · Source: mattpocock · Confidence: medium

196. Grounding Ledger

Long-form generation loses the reader when a passage leans on a concept not yet established; keep a running ‘grounded’ set and only pick reachable beats.

Category: advanced-io-domain · Source: mattpocock · Confidence: high

197. Fog of War (Frontier Planning)

Planning a large effort where downstream decisions can’t yet be phrased; keep a deliberately incomplete map, graduate fog to tickets once statable.

Category: advanced-workflow · Source: mattpocock · Confidence: high

198. Two-Axis Orthogonal Review (No Cross-Axis Reranking)

One review dimension masks a failure in another (clean code that builds the wrong thing); run axes separately and never rerank across them.

Category: advanced-quality · Source: mattpocock · Confidence: high

199. Push-Right Checkpoint (Ask Once, Late, as a Brief)

Involving the human too early or dumping raw output wastes their attention; defer the checkpoint, ask once with a decision-ready brief, never a draft.

Category: execution-control · Source: mattpocock · Confidence: high

200. Explore / Exploit as Separate Skills

Agents converge on structure too early, collapsing the space of what could be produced; make divergence and convergence different invocations.

Category: advanced-workflow · Source: mattpocock · Confidence: medium

201. Absence-of-Seam as a Finding

When you can’t write a clean regression test, the temptation is a bad one (false confidence); report the gap as a first-class architectural finding.

Category: advanced-quality · Source: mattpocock · Confidence: medium

202. Feedback-Loop-Before-Hypothesis Gate

Agents jump to theories before they can even detect the bug; require one already-run red-capable command before any hypothesis phase.

Category: advanced-workflow · Source: mattpocock · Confidence: medium

203. HITL Typing — Agent Never Simulates the Human

In a human-in-the-loop step, an agent will fabricate the human’s answers to keep moving; a HITL ticket only resolves through the live exchange.

Category: safety-and-trust · Source: mattpocock · Confidence: medium

204. Static-Trace Verification When Execution Is Unsafe

Some artifacts can’t be safely run (open browsers, block on input, irreversible) yet need verification; trace values statically instead of executing.

Category: advanced-quality · Source: mattpocock · Confidence: medium

205. Smart Zone — Fork Before Reasoning Degrades

Reasoning quality falls off well before the hard context limit; fork/handoff at the reasoning-quality cliff, not at the token cap.

Category: claude-code-platform · Source: mattpocock · Confidence: medium

206. ADR-Worthiness Triple Gate

Decision logs bloat with trivia; record a decision only when all three hold—hard to reverse, surprising without context, and a real trade-off.

Category: knowledge-and-context · Source: mattpocock · Confidence: medium