CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

docs-grounding-verifier

Use this skill to verify CLAIM-LEVEL grounding of a documentation page (or set of pages) against the source code. Activate when you have specific pages to check for factual accuracy -- not when sweeping a whole corpus (use docs-corpus-audit for that) and not when triaging a PR diff (use docs-sync for that). Trigger nouns: "is this doc accurate", "verify the page against the code", "fact-check this section", "any claims that drifted from source", "fact-checking", "grounding audit", "drift hunt", "claim verification". Returns per-claim verdicts (GROUNDED | PARTIAL | CONTRADICTED | UNSUPPORTED) with file:line evidence citations. Catches paragraph-level inaccuracies that page-level audit averages over -- e.g. a paragraph with 5 claims where 4 are grounded and 1 is fabricated. Does NOT modify files (returns advisory only); does NOT re-architect the docs; does NOT triage PRs.

Quality

91%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

81%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A well-engineered orchestration skill: a clearly sequenced pipeline with explicit validation gates, feedback loops, concrete JSON contracts, and real bundled scripts. Weaknesses are minor — some repetition of the deterministic-retriever rationale, an underspecified invocation contract for the retrieval script, and three eval files referenced in the ship gate that are missing from the bundle.

Suggestions

Create the referenced evals/ files (trigger-evals.json, content-evals.json, run-evals.sh) or remove them from the 'Bundled assets' list and ship gate — three referenced paths do not exist in the bundle.

State the invocation contract for scripts/retrieve-evidence.sh (stdin format vs. argument) as concretely as the 'scripts/extract-claims.py <page>' invocation is stated.

State the deterministic-retriever rationale once (in the Architecture section) and cut its repeats in Phase 3 and 'Failure modes guarded against'.

DimensionReasoningScore

Conciseness

Mostly efficient, skill-specific content with no teaching of concepts Claude already knows, but the deterministic-retriever rationale is repeated three times ("DETERMINISTIC (grep + AST hints), not LLM", "Stage 2 is deterministic; judge sees only real evidence", "evidence retrieval is DETERMINISTIC") and the 'Failure modes guarded against' section restates earlier points. Minor trims possible, so anchor 4 ('minor instances of over-explanation that could be trimmed') fits better than the every-token-earns-its-place anchor 5, and there is no padded concept explanation that would drop it to 3.

4 / 5

Actionability

Concrete and mostly executable: exact script invocation ("scripts/extract-claims.py <page>"), precise per-stage JSON I/O shapes, a numeric ship gate ("grounding_rate >= 0.9"), and a monotonic-increase stop rule. Minor gaps keep it at anchor 4 — the invocation contract for retrieval ("pipe to scripts/retrieve-evidence.sh") and judge dispatch ("Send the prompt + the tuple") is underspecified relative to the extract stage, though the scripts' --help flags partially compensate. Anchor 3's pseudocode/missing-key-details pattern does not apply.

4 / 5

Workflow Clarity

Six clearly sequenced phases (SCOPE through ALIGNMENT LOOP) with explicit validation and feedback loops: re-run the pipeline after fixes, "The grounding_rate must MONOTONICALLY INCREASE between iterations or the loop has diverged -- stop and escalate to the operator", and UNSUPPORTED claims get a retrieval-retry path before being treated as CONTRADICTED. This matches anchor 5's feedback loops with error recovery; the batch-operation cap at 3 does not apply because validation checkpoints are present, and anchor 4's 'most checkpoints' understates how explicit they are.

5 / 5

Progressive Disclosure

Good one-level-deep structure with a clearly labeled 'Bundled assets' section, and scripts/extract-claims.py, scripts/retrieve-evidence.sh, scripts/verify-page.sh, and assets/judge-prompt.md all exist in the bundle. However evals/trigger-evals.json, evals/content-evals.json, and evals/run-evals.sh are referenced in both the ship gate and the bundled-assets list but are absent from the bundle (as are the ../../personas/*.persona.md links), so navigation is not 'easy' as anchor 5 requires — anchor 4's 'minor organization gaps' is the best fit. Not anchor 3: the structure itself is well-signaled and nothing that belongs in separate files is inlined.

4 / 5

Total

17

/

20

Passed

Description

100%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is exemplary: concrete capabilities and outputs in third person, an explicit and natural trigger list, clear what/when coverage augmented with negative triggers, and active disambiguation from sibling skills. It runs long, but every clause carries discriminating information rather than padding.

DimensionReasoningScore

Specificity

Multiple specific concrete actions with comprehensive coverage: "verify CLAIM-LEVEL grounding of a documentation page (or set of pages) against the source code", "Returns per-claim verdicts (GROUNDED | PARTIAL | CONTRADICTED | UNSUPPORTED) with file:line evidence citations", "Catches paragraph-level inaccuracies that page-level audit averages over". All in third person ('Returns', 'Catches', 'Does NOT modify'), matching the anchor-5 example's breadth; anchor 4's 'minor gaps in coverage' does not apply since inputs, outputs, granularity, and side-effect boundaries are all stated.

5 / 5

Completeness

Explicitly answers both what ("verify CLAIM-LEVEL grounding... Returns per-claim verdicts... with file:line evidence citations") and when ("Activate when you have specific pages to check for factual accuracy" plus concrete trigger nouns), and adds negative triggers ("not when sweeping a whole corpus... not when triaging a PR diff") that make the 'when' even more explicit than the anchor-5 example. No 'Use when' clause is missing, so the cap at 3 does not apply, and anchor 4's 'could be more explicit' clearly understates it.

5 / 5

Trigger Term Quality

Explicitly enumerates natural phrases a user would actually say: "is this doc accurate", "verify the page against the code", "fact-check this section", "any claims that drifted from source", "fact-checking", "grounding audit", "drift hunt", "claim verification". That is comprehensive synonym coverage meeting the anchor-5 bar; anchor 4 ('a few natural terms missing') is beaten by both the count and the naturalness of the phrases.

5 / 5

Distinctiveness Conflict Risk

Clear niche (claim-level grounding of specific pages) with explicit disambiguation from the two nearest siblings: "not when sweeping a whole corpus (use docs-corpus-audit for that) and not when triaging a PR diff (use docs-sync for that)". Minimal conflict risk as anchor 5 requires; anchor 4's 'minor overlap risk with closely related skills' is already preempted by the named routing to those siblings.

5 / 5

Total

20

/

20

Passed

Validation

93%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

relative_links

Relative link issues: 2 suspicious

Warning

Total

15

/

16

Passed

Repository

microsoft/apm

Path

.apm/skills/docs-grounding-verifier/SKILL.md

Reviewed

Table of Contents

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.