nicsuzor/dogfood

Delegated Instruction Testing Guidelines

Test whether a set of instructions produces correct, complete, and verified outcomes when executed by a contextless agent. Do not perform the work yourself.

Core Directives

Phase 0: Verification & Data Landscape Mapping

Before writing instructions or propagating subagent results:

• Verify verdicts: Use /verify by citation to evaluate the subagent's actual output (freshness, completeness, limitations) against the original brief. Do not blindly accept or relay its self-reported status.
• Sample data sources: Open and read sample files directly. Do not assume data formats or presence (e.g. verify if a file contains input vs. output).
• Map data channels: Understand how data flows between main agents, subagents, and hooks/logs. Verify the exact delivery mechanism (e.g. tool result vs. system message).
• Glob safety: Avoid globbing large directories (10K+ files) with commands like ls *.md (which fail silently). Use targeted list/find queries (find or ls | head).

Phase 1: Research and Draft

• Write self-contained instructions detailing objectives, exact data paths, sampling parameters, expected output format, and saving locations.
• Work directly in the target skill file for mature instructions. Avoid leaving stray scratch files in the repo.
• Enforce /craft author mode review to check for shallow-execution vulnerabilities before delegating.

Phase 2: Commission Execution

• Scale incrementally: Start with a small batch size (e.g. N=2 tasks) to verify the pipeline before scaling.
• Isolate context: Launch the subagent with only the instruction file as context. Do not provide verbal coaching.
• Execution management: Let the agent run. Do not abort/restart for scope changes; send a redirect message to the running agent instead. Interrupt only for active harm.

Phase 3: Analyze Friction and Iterate

• Analyze transcripts to classify friction (e.g. missing paths, ambiguous criteria, shallow analysis).
• Update the instructions in-place to address root causes, avoiding over-fitting to the specific test instance.
• Run at least 2 verification trials per condition when verifying that an edit closed a gap.

Phase 4: Independent Quality Review

• Commission a separate review agent (e.g. /strategic-review or /verify) to evaluate depth, accuracy, and fitness of the subagent's deliverables.
• Enforce qualitative assessment by agents rather than relying on deterministic script checks.
• For decompose-mode instruction tests, use references/decomposition-eval.md (epistemics rubric + worked gold-standard pair).

Phase 5: Codify & Land

• Promote tested instructions to canonical skills or commands.
• Verify deliverables actually reached their target destinations (e.g., reviews posted, commits pushed, task tracker updated).
• Always leave a loose thread: File follow-up tasks for any remaining friction items, promotion work, or subsequent phases before exiting.

Output Expectations

• Respond with structured, concise summaries of dogfooding outcomes, listing specific instruction defects found, edits made, and verification run verdicts.