ai-accuracy-guardrails-and-grounding-design

Designs absolute guardrails and grounds AI outputs against a named source of truth to counter sycophancy and hallucination, weighed against "tokonomics" — the cost, latency, and token price of every a

SKILL.md

GitHub

AI Accuracy Guardrails & Grounding Design

Purpose

Separates two things that get conflated under "the AI is wrong": accuracy (is the underlying claim actually true) and precision (is the response consistent and repeatable). An AI can be highly precise and confidently, consistently wrong — the specific failure mode this skill guards against is sycophancy: a model tuned to please the user, agreeing with or elaborating on an incorrect premise instead of correcting it. This skill designs the two mechanisms that catch that failure — absolute guardrails and grounding against an external source of truth — and weighs their cost.

Anchored in research

  • The accuracy/precision distinction and the sycophancy risk are drawn from the source material's synthesis of 2026 AI product-design practice (unattributed to a single named individual in a way that could be independently verified) — treat as an established industry concern rather than one person's original claim; sycophancy as a systemic LLM risk is broadly documented across the AI safety and product-design literature.
  • The Whoop fitness-bot example (absolute topical guardrails — a fitness AI must refuse and redirect away from unrelated topics regardless of how the user tries to provoke it) and the IMDb example (grounding a movie cast question against IMDb as the specific source of truth, rather than the model's own trained knowledge) are used in the source material as illustrative cases. Neither has been independently verified against the named companies' own published documentation — use them as illustrations of the pattern (absolute topical guardrails; grounding against a named, authoritative source), not as confirmed citations of those companies' actual systems.
  • "Tokonomics" (the source material's term for the cost/latency/token trade-off of guardrail complexity) is used here as a practical design heuristic, not a citation-requiring claim — the underlying trade-off (longer system instructions cost more tokens and add latency) is a straightforward, verifiable property of how LLMs are billed and run.

Method

  1. Separate every quality complaint into an accuracy problem or a precision problem before designing a fix — they need different guardrails. A model that gives different plausible-sounding answers to the same question each time is a precision problem (see ../ai-behavioral-specification-design/SKILL.md's temperature guidance). A model that confidently states something false is an accuracy problem — that's this skill's territory.
  2. Name the sycophancy risk explicitly wherever the AI interacts with a confident or leading user statement. Write a rule into the guardrail spec that the AI checks a claim against ground truth (step 4) before agreeing with or elaborating on it, especially when the user states something as fact. An AI that only ever validates the user's framing is optimizing for approval, not correctness.
  3. Design absolute guardrails for anything genuinely out of scope or unsafe — a hard "never" rule, not a soft preference, for topics the product must categorically decline regardless of how persistently or cleverly a user tries to provoke a response (see ../ai-behavioral-specification-design/SKILL.md step 4 for HOW it declines — the redirect language). List these explicitly; don't assume the model will infer scope boundaries reliably on its own.
  4. Name a specific source of truth for every category of factual claim the AI makes, and require the AI to check against it rather than answer from its own trained knowledge alone. "The system should only report what's actually true according to [named authoritative source]" is a testable design rule; "the system should be accurate" is not. If no authoritative source exists for a claim category, that's a signal the AI shouldn't be making confident claims in that category at all — flag it for a lower-confidence framing or an escalation (see ../hitl-maturity-and-confidence-routing/SKILL.md).
  5. Weigh every added guardrail instruction against its "tokonomics" — longer, more elaborate system instructions cost more in tokens, add latency, and (past a point) can start to confuse the model about priority among instructions. The design goal is maximum safety with minimum instruction footprint, not maximum instruction length. If a guardrail can be achieved with a shorter, sharper rule instead of an exhaustive list of edge cases, prefer the shorter rule and test it against the edge cases rather than writing all of them into the prompt.
  6. Test the guardrail set against adversarial framing, not just straightforward requests — a user rephrasing a disallowed request as a hypothetical, a role-play, or a "just curious" aside is a normal real-world pattern, not an edge case to skip.
  7. Produce a structured output: the accuracy/precision split for the product's main claim categories, the named source-of-truth per category, the absolute-guardrail list, and a short note on the token/latency cost of the guardrail set as designed.

What this skill does NOT do

  • Doesn't design the AI's tone or sentiment — see ../ai-behavioral-specification-design/SKILL.md.
  • Doesn't implement the retrieval/grounding pipeline technically (e.g. a RAG architecture against the named source of truth) — that's an engineering task; this skill specifies WHICH source and WHAT must be checked against it.
  • Doesn't guarantee zero hallucination — grounding and guardrails reduce the risk materially, they don't eliminate it; a genuinely high-stakes claim category still needs the escalation path from ../hitl-maturity-and-confidence-routing/SKILL.md, not guardrails alone.
  • Doesn't perform the regulatory/ethics risk check — see ../../../ai-strategy-and-governance/skills/responsible-ai-and-governance-check/SKILL.md.

Refinement notes

Areas to keep deepening with real practice:

  • your own worked examples of a sycophancy failure caught in review, and the guardrail rule that would have prevented it
  • a concrete "source of truth" mapping template per claim category (into ../../references/)
  • real before/after token-cost comparisons for a guardrail set that was successfully shortened without losing safety coverage

Once this section is filled in and validated in practice, update the maturity field in skills_index.json to draft, validated, or canonical (see ../../../meta/maturity_levels.md). Don't add new fields to the frontmatter — name and description are the only ones allowed (see ../../../meta/frontmatter_schema.md).

Continue from here

  • Before this in this pack: ../ai-behavioral-specification-design/SKILL.md — the tone/sentiment layer this skill's guardrails sit underneath.
  • Next in this pack: ../hitl-override-metrics-and-feedback-audit/SKILL.md — measures whether the guardrails designed here are actually holding up in production.
  • Related skill in another pack: ../../../ai-strategy-and-governance/skills/responsible-ai-and-governance-check/SKILL.md — the regulatory/ethics layer, complementary to this skill's product- design layer.
  • This pack's shared guardrails: ../../CLAUDE.md

References

  • ../../references/hitl-partnership-heuristics-research.md — full sourcing and grounding-strength notes for this pack
  • ../../references/ — the pack's shared background material
  • ../../CLAUDE.md — the pack's shared guardrails