html-to-editable-pptx

Safely execute a delegated repository task by proving the git environment, protecting dirty worktrees, running required gates, choosing commit/push or patch export, and halting instead of publishing when verification fails.

Use when building or debugging format converters, parsers, or export pipelines (e.g. HTML→PPTX, DOM→IR, doc→doc): three cross-checking techniques — probe-first classification, dual-source extraction, declaration-level validation — because no single representation of the data can be trusted.

Build converter/parser fixture matrices by probing third-party trees before handlers, adding regression fixtures before feature fixtures, declaring expected behavior and primary gates, isolating outputs, and making renderer checks opt-in.

Create, critique, and validate Mermaid architecture diagrams

Record decisions reached through interview as a non-authoritative repository document that preserves consensus across sessions and across agents. Use whenever interview outcomes need to outlive the conversation in a form that future readers — both human users and other agents — can act on without re-deriving the reasoning.

Run a structured interview where one decision is taken per turn, automatic inferences from prior consensus are surfaced before manual questions, and depth on each decision is preferred over interview round-trips. Use whenever a user needs to make a sequence of related decisions whose record will outlive the conversation.

Process an external evaluation report into prioritized work groups with clear dependencies, separating procedural fixes from substantive interview questions and keeping each group small enough for a single self-contained delegation. Use whenever an external agent or reviewer has produced a list of findings that the project needs to act on without re-litigating each finding.

Diagnose and recover when the same delegated instruction halts repeatedly, by treating environment assumptions as falsifiable and revising the instruction rather than the receiving agent. Use whenever two or more attempts at the same task have failed in a way that suggests the writing agent's model of the executing environment is wrong, not that the executing agent is misbehaving.

Close a session that has accumulated too much context by producing two artifacts — a one-time session summary that captures current state and immediate next actions, and a permanent methodology Skill that captures patterns for re-use. Use whenever a session reaches a logical milestone and a new session needs to resume work without re-reading the full conversation.

Write a self-contained instruction sheet that hands a mechanical task to another agent or session with explicit environment requirements, explicit input partitions, explicit failure modes, and explicit lifetime. Use whenever the writing agent cannot execute the work itself due to environment, authority, or role separation, and the receiving agent will execute without further interview.