Spec-Literal Execution: Stop Hidden Agent Transforms

Where this fits

You have an exact-match grader that expects every input record to appear as its own output line. Your agent reads a spec that says one line per product with quantity > 0, sees duplicate product names, decides to merge them, and fails with score 0.0. This skill installs a rule: only clauses explicitly present in the spec may be applied. Each input row maps to its own output line unless the spec says otherwise.

Why agents benefit

• Prevents unstated aggregation, deduplication, rounding, or cleanup transforms.
• Preserves input cardinality, so exact-match and diff-based graders stay meaningful.
• Removes ambiguity around phrases like per product by anchoring to the literal rows.
• Makes agent behavior auditable: every transformation can be traced to a spec clause.
• Reduces debugging time by eliminating invisible, well-intentioned edits.

Practical scenarios

Invoice line-item extraction

The spec asks for each line item from the invoice. An agent might sum two identical SKUs into one row. With spec-literal execution, each original line item remains separate, matching the expected output format.

Log event export

A log contains repeated health-check events. The agent decides to deduplicate them to clean up the output. The downstream system expects every event, including duplicates. The skill forces the agent to keep every record.

Database snapshot rows

A spec says active users, one per line. An agent may group users by role and output counts. Literal execution outputs the actual user rows, no pivot or summary.

Add it to your agent workflow

Install the skill with:

npx skills add ashutoshsinghpr7/wikiskill --skill spec-literal-execution

Then instruct your agent: Use the spec-literal-execution skill. Apply only clauses written in the spec. Do not aggregate, deduplicate, or clean unless a clause explicitly asks for it.

Compare before adopting

Before adding it to a toolchain, compare quality signals: does the skill consistently improve exact-match scores in your eval set? Check maintenance freshness, including version, license, and recent commits. Consider alternatives like a custom system prompt that says do not transform, and test whether the skill’s explicit rules outperform that prompt. Finally, evaluate workflow fit: it is most useful for graders that compare line-for-line, not for downstream systems that consume aggregated reports.

Why it is worth tracking

With 40 stars and an MIT license, this small skill is a useful reminder that many agent failures come from adding transforms the spec never requested. It is worth tracking because the pattern of literal reading before interpretation can be reused across grading, parsing, and migration tasks. When you see repeated exact-match failures caused by silent normalization, evaluate this skill against your own rubrics. For a closer look at the documented clauses, visit the OpenAgentSkill listing at https://www.openagentskill.com/skills/ashutoshsinghpr7-spec-literal-execution.