a field guide for agent context
Agent skills for the work around the implementation loop: assemble the right context, pressure-test the direction, shape the output.
01 · Premise
An agent acts on the map you hand it (prompt, skills, context), not the territory of the real project. Where the map is silent, wrong, or stale is where agents fail, confidently. Most skill bundles optimize how the agent writes code. ctxops shrinks the map/territory gap at the bookends instead.
02 · The set
Assemble a tight working context pack for a task or decision.
One-page operational brief to resume or hand off a whole project.
Search persistent memory before answering "did we already do X?"
Adversarial pass on a direction before committing. Post-hoc mode for old calls.
Stakeholder progress messages from git, memory, and open threads.
Turn a finding into paste-ready issues with checkable acceptance criteria.
Turn real work into a durable shareable artifact with a publish checklist.
Interview + scan to surface a plan's unknowns before implementation starts.
Ranked noise report and agent ignore files. Exclusion is half of assembly.
Plain SKILL.md files per the Agent Skills standard: Claude Code, Codex, Gemini CLI, OpenCode, or any compatible harness. Trackers and memory tools are examples, not dependencies.
03 · Evidence
ctxops ships its own instrumentation and publishes its own numbers. From the first 26 days of instrumented maintainer use:
| Eval | On | Off | Result | What decided it |
|---|---|---|---|---|
ctx-pack | 8 | 6 | skill-on | Freeform arm asserted a false load-bearing fact; the judge verified. |
project-briefing | 8 | 7 | skill-on | Freeform arm misreported repo push state. |
recall-prior-work | 8 | 7 | skill-on | Skill arm re-verified memory against the current repo, with dates. |
ticket-scoping | 8 | 7 | skill-on | Sizes tied to driving assumptions; ACs tied to verification artifacts. |
decision-review · status-update · artifact-packager | 8 | 8 | ties | Capable-model defaults already match these rubrics. |
Every decided eval turned on the same failure: an unverified or stale load-bearing claim. That is the exact failure mode the skills' rules target. Method and caveats: evals/ab-results-2026-07.md.
04 · Feedback loop
skill_retro.py classifies each one as adopted, adapted, or unclear from your own transcripts.ab_eval.py re-runs the blind eval before releases.05 · Install
Other harnesses: symlink or copy the skills/ directories into your skill path. A Codex-style manifest ships in the repo.