Grounded Generation

Use when generating content, reports, summaries, evaluations, or code against external facts. Requires verified source data, explicit lineage, and no invented capabilities or copied answers.

selamy-labs 7abb34c 1.4 KB Updated

File contents

Grounded Generation

Generate only from verified inputs. Missing data is a task, not permission to invent.

Tag every datum's lineage

Mark each value the output relies on as one of:

  • primary — read directly from the system of record this run.
  • copied — taken from another candidate, prior output, or cache; carries that source's correctness, not independent confirmation.
  • synthetic — fixture, sample, or placeholder; never a result.
  • derived — computed from the above; inherits the weakest tag it draws on.

Lineage is the behavior this skill adds: an output you cannot tag is an output you cannot ground. Tags propagate — a derived claim is only as primary as its weakest input.

Apply per generation type

  • For comparisons, prove each candidate is primary-sourced from the same underlying data. If one answer is copied from the control, the result validates formatting only.
  • For code, build against routes, schemas, functions, and APIs that exist. Nonexistent endpoints and fabricated helpers must fail review.
  • For analysis and forecasts, run leakage, lookahead, and overfit checks before trusting any performance claim.

selamy-labs/agent-skills/tree/main/skills/grounded-generation commit 7abb34cbbf

Frequently asked questions

npx skillmds@latest add selamy-labs/grounded-generation