# Field notes

Terms from agent-driven software delivery, each recorded with what Nicholas Hopper did with it in practice.

- [Agent skill](https://nicholashopper.com/notes/agent-skill.md): An agent skill is a packaged set of instructions, usually a SKILL.md file, that an AI coding agent loads to carry out a specific kind of task. Nicholas Hopper writes skills as the operating procedures of his systems and tests them like code.
- [Agent slop](https://nicholashopper.com/notes/agent-slop.md): Agent slop is the low-value output that AI coding agents leave in a repository: duplicated code, redundant documents and commits too large to review. Nicholas Hopper measured it in a real repository and found that it is structural and historical, and invisible to a linter.
- [Block instead of guess](https://nicholashopper.com/notes/block-instead-of-guess.md): Blocking is an agent's explicit, recorded stop when it cannot proceed without a person. Nicholas Hopper gives his agents a block command and tells them to use it for sign-ins, quotas and judgment calls instead of guessing.
- [Burn-down list](https://nicholashopper.com/notes/burn-down-list.md): A burn-down list is a file of known violations of a rule that is allowed to shrink but never to grow. Nicholas Hopper uses one to enforce a new lint rule in continuous integration from the first day, without first fixing every existing violation.
- [Diff-scoped gate](https://nicholashopper.com/notes/diff-scoped-gate.md): A diff-scoped gate is a set of pre-merge checks chosen according to which files a change touches. Nicholas Hopper built one in which a script, not the agent, decides the scope, and in which the scope can be widened by judgment but never narrowed.
- [Fail closed](https://nicholashopper.com/notes/fail-closed.md): Failing closed means that when a component cannot do its job, the system refuses to proceed. Nicholas Hopper applies it at every boundary with a language model, so that a model outage or an unparseable response produces a refusal and never a silent pass.
- [File handoff](https://nicholashopper.com/notes/file-handoff.md): A file handoff is a transfer of work between two AI agents in which the only thing passed is a file that both sides can validate. Nicholas Hopper uses file handoffs so that agents on different harnesses can share a workflow without sharing a conversation.
- [Git worktree delivery](https://nicholashopper.com/notes/git-worktree-delivery.md): Git worktree delivery is the process of landing work from several parallel git worktrees, each used by a different AI agent session, onto one main branch. Nicholas Hopper wrote a delivery procedure that proves no agent's commit was lost.
- [Golden set](https://nicholashopper.com/notes/golden-set.md): A golden set is a fixed collection of inputs with known correct outputs, used to score a model or tool before and after a change. Nicholas Hopper builds a golden set per tool and gates prompt and model changes on the scores.
- [Green](https://nicholashopper.com/notes/green.md): Green is the state of a check or pipeline that reports success. Nicholas Hopper defines in writing what does not count as green, because a pipeline run by AI coding agents can report success without having tested anything.
- [Human approval gate](https://nicholashopper.com/notes/human-approval-gate.md): A human approval gate is a point in an agent workflow where the agent must stop and a person must act before the work continues. Nicholas Hopper places the gate at the one irreversible action in a workflow and lets agents do everything before it unattended.
- [Reciprocal Rank Fusion](https://nicholashopper.com/notes/reciprocal-rank-fusion.md): Reciprocal Rank Fusion (RRF) is a method for combining several ranked result lists into one by scoring each result on its position in each list. Nicholas Hopper used it in Memector to fuse up to five search channels, and measured where it fails.
- [Run directory as memory](https://nicholashopper.com/notes/run-directory-as-memory.md): A run directory is a folder that holds every input, output and decision of one execution of an agent workflow. Nicholas Hopper treats the run directory as the workflow's only memory, in place of chat history or the state of any external tool.

---

Canonical page: https://nicholashopper.com/notes  
Last updated: 2026-11-17  
Index of all pages: https://nicholashopper.com/llms.txt
