Fixture Skill: Make Text Cleanup Rules Visible

Where this fits

You are transforming raw Russian text into clean rows for analysis. The agent knows that "является" should become a dash and filler "ну" should disappear. Those are not grammar rules — they are editorial choices. The fixture skill pins them into a snapshot and makes silent edits visible, so every transformation is auditable before it enters a dataset.

Why agents benefit

• Separates "attribution" from "rule": the Sources section is explicitly not a rule, preventing agents from over-applying metadata as transformations.
• Snapshot semantics: fixture acts as an activation trigger; when it changes silently, the snapshot diff shows the exact loss, so data corruption is caught early.
• Minimal parsing: the bare table word|replacement format is directly loadable by scripts and LLMs without a custom schema.
• Interpretable structure: Contents anchors (A, B) give the agent stable points to reference when deciding whether a replacement is safe.
• Reduces prompt bloat: cleanup rules live outside the system prompt and can be versioned, diffed, and reviewed independently.

Practical scenarios

Normalizing field notes

An agent processes interview logs that contain conversational particles. By applying "ну" → remove, it keeps the snapshot in the output. A human reviewer can later distinguish intended filler removal from accidental meaning loss.

Rebuilding sentence flow

When cleaning prose, "является" is replaced with a dash or a restructured phrase. The fixture tells the agent to treat this as an editorial action, not a language rule, and the snapshot records each site so generated text can be audited.

Detecting silent rule drift

A teammate edits the fixture table from "является" to a more aggressive rewrite. Because the fixture is a snapshot trigger, the next run surfaces the diff before the change contaminates the full dataset.

Add it to your agent workflow

Install the skill:

npx skills add talkstream/ru-text --skill fixture

After installation, load the fixture as a snapshot before any text-normalization step:

const { fixture } = await loadSkill("fixture");
const cleaned = text.replace(/является/g, "—").replace(/\bну\b/g, "");
snapshot.record({ cleaned, fixtureVersion: fixture.version });

This keeps the replacement pair visible in the output and makes any future silent edit obvious.

Compare before adopting

Before adding fixture, compare it with a simple JSON dictionary or regex list. Ask: How often does the maintainer update the table? Are replacements for your domain or generic editorial preferences? Does your workflow need diffable snapshots, or do you only need static substitutions? Fixture wins when accountability matters; for one-off fixes, a plain mapping may be simpler and easier to audit.

Why it is worth tracking

The ru-text repository has 224 stars, and its fixture skill models a pattern worth watching: letting agents hold editorial choices in a snapshot instead of burying them in prompt logic. That is the exact place where quiet data loss happens. Track commit frequency and issue responses; revisit this skill when you begin cleaning Russian text at scale and need to prove transformations are reversible and reviewable. A current snapshot of the skill and related comparisons is on OpenAgentSkill: https://www.openagentskill.com/skills/talkstream-fixture