I ran the robot factory by hand first
Agent operations · Workflow discovery
What you will leave with
Build a batch intake and clarification protocol before automating code changes.
What you will build
A top-half protocol that produces confirmed tickets; existing delivery controls own the bottom half.
- 01Intake batchPreserve raw source, order, and media.
- 02FindingsOne observable issue or product judgment per record.
- 03Uncertainty queueConfidence and questions are explicit fields.
- 04ConfirmationOne bounded human round resolves intent.
- 05Ticket groupsCode ownership and dependencies shape dispatch.
Build it in sequence
Capture without fixing
Walk the entire journey and record everything in the fastest natural medium. Preserve the raw source so every structured finding can be traced back when interpretation is disputed.
Normalize into an intake schema
Separate observation, proposed interpretation, confidence, affected surface, and source. Do not turn a vague note directly into implementation language.
interface Finding {
id: string
source: { artifact: string; region: string }
surface: string
observation: string
interpretation?: string
confidence: number
status: 'confirmed' | 'needs-human-confirmation'
question?: string
}Batch uncertainty once
Route low-confidence, contradictory, or direction-setting findings into one clarification packet. Ask the smallest decision that unlocks each item and keep confirmed work moving.
Group by ownership and dependency
Form tickets around cohesive code surfaces and independently verifiable outcomes, not equal item counts. Queue shared-file or migration work behind the lane that owns the dependency.
Hand off with acceptance evidence
Each ticket carries confirmed intent, exclusions, owned paths, and deployed proof. The downstream worker must not reinterpret the original handwriting.
Definition of done
- Every finding links to raw source.
- Observation and interpretation are separate.
- Low confidence cannot create a ticket without confirmation.
- Clarifications are batched rather than interrupting per item.
- Tickets map to clean ownership and acceptance evidence.
What breaks and what it means
The human receives too many questions.
Improve extraction and ask only decisions that change intent; batch the remainder.
Tickets still collide in shared files.
Group by dependency and ownership, not product page or finding count alone.
Questions, corrections, and useful disagreement
Sign in with GitHub to join the conversation. Comments are public. Article views are anonymous and counted once per browser, per article, each day.