This repository contains one engineer's date-stamped, sanitized working notes on coding-agent instruction design and verification. It collects a failure-pattern taxonomy, a trace self-audit template, a regression-corpus method, evolution notes, and sanitized configuration and skill samples.
The material is n=1, subjectively graded, and model/version sensitive. Treat it as a set of artifacts to inspect and adapt, not as a framework, benchmark, policy, comprehensive methodology, best-practice or production-readiness claim, or scientifically validated evidence. Take what's useful, verify it in your own environment, and discard the rest.
I kept finding that a plausible instruction change had no durable test in my own workflow. So I started using this loop: observe the failure, classify the trace, decide whether anything should change, route any warranted fix to the right instruction layer, add a regression case where a stable check exists, and revisit it after model or harness drift.
The example configurations are supporting artifacts, not the main claim. The focus is the inspectable method — including cases where it invalidates its own apparent successes.
verification-kit/— instruments for classifying trace failures, auditing a conversation, maintaining private regression cases, plus a dated account of how the method changed.harness/— sanitized cross-surface instruction samples, operating notes, and reusable skills for Claude Code, claude.ai, Codex, ChatGPT, Cursor, and Snowflake CoCo. Its own index links each artifact.
- Audit a conversation with the trace self-audit template.
- Inspect recurring failure patterns in the taxonomy.
- Turn a confirmed failure into a private regression check with the corpus method.
- Adapt instruction or skill samples from the harness.
Dated changes, and the reasoning behind the conventions, are in CHANGELOG.md.
This is a sanitized personal project. It contains verification instruments, operating notes, and example configurations — not raw transcripts, live configuration, or private regression-corpus entries.
These are personal working notes, not organizational guidance. They do not represent any employer and do not replace applicable policy or mandatory requirements. If anything here conflicts with the rules in your environment, follow those rules.
This project builds on work done at Booking.com. With approval from Booking.com, it was released as open source, for which the author would like to express his gratitude.
Licensed CC BY 4.0 — use, adapt, and share, including
commercially, as long as you credit this repository and link back to it.
Crediting the repo is the attribution asked for here; no personal name is
needed. Provided as-is, with no warranty. The LICENSE file is the
unmodified Creative Commons legal code; attribution belongs here, not in
it. Linked third-party works remain under their own terms; this license
covers this repository's notes and samples, not the works they cite.