structify

Structify

Turn a recurring agentic failure into a deterministic, tested skill.

When to Use

An agent (verify, implement, any) fails at a mechanical pipeline step — not because the code is wrong, but because the infrastructure, skill, or config that the agent depends on is missing or broken.

Signals:

• Permission denied on a tool the agent needs
• Agent improvises a multi-step pipeline with no skill to guide it
• Same failure across multiple issues (different code, same infra)
• Output quality is garbage (tiny images, corrupted uploads, truncated data)
• Agent retries the same broken step in a loop
• "It worked on that other issue but not this one"

Not for:

• Code bugs in the target repo (that's implement/rework)
• Orchestrator crashes (that's standard investigating)
• One-off flaky failures (that's retry)

---

The 6 Steps

1. DETECT — name the failure pattern

State the failure clearly. What operation failed? What did the agent try to do? What was the actual output vs expected output?

FAILURE: Verify agent uploaded 138x300 pixel JPEG thumbnails to Linear.
EXPECTED: Readable screenshots at 800px+ resolution.
PATTERN: Agent couldn't pass retina base64 through context, panic-resized.

2. COMPARE — find a working reference

Find an instance where the same operation succeeded, even on a different issue or in a different context. The diff between working and failing reveals the gap.

Where to look:

• Linear comments on recent issues (search for the operation type)
• Agent logs (~/.autosymph/logs/ or equivalent)
• Git history for when this used to work
• Other agents/projects that do the same thing

What to compare:

• Tool calls in ndjson logs (what was called, in what order, with what params)
• File sizes and dimensions of artifacts
• Permission denials vs approvals
• Time spent on the operation

If no working reference exists, the comparison is "what a human would do" vs "what the agent tried."

3. ROOT CAUSE — identify the specific gap