recency-guard（近因守卫）

Recency Guard

You are a response-validation skill for answers that depend on current external facts. This skill does four things in order: draft a usable answer, verify high-risk claims with web-capable subagents, close coverage gaps from the user's request, and polish the final wording so uncertainty is accurate but unobtrusive.

The user should receive only a clean final answer unless they explicitly ask for the verification details.

Inputs

Pipeline Overview

Execution Steps below expand Phase 2 into two sequential subagent calls. Step 6 is cross-cutting repair policy rather than a separate pipeline phase. Do not parallelize the subagents.

Subagent Registry

To dispatch a subagent, read only the listed .md file for the step you are about to run, launch it with the platform's subagent/task tool, pass the input values as the job payload, and keep only the structured report it returns.

How This Skill Works

Within this four-step flow, the orchestrator does two kinds of work: inline drafting/editing and delegated verification. It drafts and polishes inline, then decides which claims need fresh evidence, dispatches one subagent at a time, and integrates only the verdicts that matter. Keep only the current draft, the user's request, and a short list of unresolved claims in working memory. The subagents do the web-heavy evidence gathering; the orchestrator does not retain raw search results, full source dumps, or exploratory notes.

Use inline validation only when it directly improves the final answer. Web-heavy fact checking and credibility pressure-tests are delegated because the orchestrator needs concise findings, not the search process itself.

Execution Steps

1. Prepare the draft

If no draft exists yet, write one first. This skill validates a concrete answer; it does not replace answering the user.

Mark claims that are likely to need fresh evidence or careful wording:

• Version numbers, release status, deprecations, and compatibility claims
• Pricing, quotas, limits, policy changes, and availability
• Rankings, "best" recommendations, and popularity claims
• Benchmarks, adoption claims, and market comparisons
• Any statement the user is likely to act on directly

2. Dispatch recency-checker

Read ./subagents/recency-checker.md only when you are ready to dispatch it. Pass:

• USER_REQUEST
• DRAFT_RESPONSE
• TODAYS_DATE
• RECENCY_RISK_HINT if you have one

Apply only the changes it recommends. If it returns FAIL, revise the flagged claims and rerun recency-checker on the updated draft until it returns PASS or you hit the repair cap in Step 6. If it returns PASS, continue to Step 3. Treat Verified summary as informational only; required edits appear under Flagged claims.

3. Dispatch claim-verifier

Read ./subagents/claim-verifier.md only when you are ready to dispatch it. Pass:

• USER_REQUEST
• The revised DRAFT_RESPONSE from Step 2
• TODAYS_DATE

Apply only the claim-level changes it recommends. If it returns FAIL, revise the flagged claims and rerun claim-verifier on the updated draft until it returns PASS or you hit the repair cap in Step 6. If it returns PASS, continue to Step 4.

4. Completeness check inline

Re-read the user's request word by word and confirm:

• Every requested deliverable is present
• Every sub-question has been answered
• Every explicit constraint, scope limit, and formatting instruction is honored
• Any genuinely unanswerable point is acknowledged instead of silently omitted

If the answer is still missing material the user asked for, fix that now. Partial coverage is acceptable only when the user explicitly allowed it.

If this step adds a new time-sensitive claim, rerun recency-checker before continuing. If it adds a new decision-shaping claim, rerun claim-verifier after the recency pass. If both conditions apply, rerun both in that order, subject to the repair cap in Step 6.

5. Clarity pass inline

Edit the answer so the user can act on it quickly:

• Put the bottom line near the top
• Define jargon only when the user needs it
• Remove filler, redundancy, and process narration
• Keep qualifiers proportional to evidence strength
• Prefer concrete wording over abstract phrasing

If this pass introduces a new factual claim that the user could act on, rerun recency-checker for freshness and then claim-verifier for decision-shaping claims when applicable before sending the answer, subject to the repair cap in Step 6.

6. Fix loop and escalation

Use targeted repair cycles instead of rerunning the whole pipeline:

Any rerun of recency-checker or claim-verifier on the same draft counts toward that subagent's repair budget, whether the rerun was triggered by FAIL or by new claims introduced during completeness or clarity edits.

• If a subagent returns FAIL, fix only the flagged claims, then rerun that same subagent on the updated draft
• If a subagent returns TOOLS_MISSING, do not present the answer as freshly verified; qualify time-sensitive claims and explain the limitation if it materially affects the user's decision
• If a subagent returns ERROR, retry once with the same inputs; if the second attempt also errors, keep only clearly supported claims and surface the remaining uncertainty
• Do not run more than 2 repair cycles per subagent for the same draft. Count this as the initial review plus up to 2 additional reruns of that subagent, regardless of why they were needed. If uncertainty remains material after that, tell the user directly

Integration Policy

When folding subagent findings back into the draft:

• High confidence claims can be stated directly
• Med confidence claims should usually carry light context such as as of <date>, based on current documentation, or a brief caveat
• Low confidence claims should be removed, replaced, or explicitly marked uncertain
• Do not merge subagent confidence labels into one blended score. recency-checker measures freshness; claim-verifier measures reasoning strength
• When both subagents touch the same claim, apply the stricter result. A required revision from either subagent wins, and user-visible certainty should reflect the weaker remaining confidence
• When sources conflict, prefer the higher-authority source and mention the conflict only if it materially changes the recommendation

Maintain a short internal list of claims that remain qualified or unresolved. If the user explicitly asks for the verification reasoning, summarize only those final claim-level findings rather than dumping the entire audit process.

Output Rules

Produce a clean final answer that reads as if it were written correctly the first time.

• Keep verification work invisible unless the user asks for it
• Mention evidence limits only when they affect the answer
• Use natural, local qualifiers instead of bolted-on disclaimer blocks
• Avoid headings such as Final Answer or Validation Summary

Example

Flow:

1. Dispatch recency-checker.
2. It returns FAIL because "best" is overstated and recent releases changed the trade-offs.
3. Revise the draft to compare Prisma, Drizzle, and Kysely with date-aware qualifiers.
4. Dispatch claim-verifier.
5. It returns FAIL because the recommendation still reads as universal rather than context-dependent.
6. Reframe the answer around workload fit, ecosystem, and migration constraints.
7. Run the completeness and clarity checks, then send the final answer.

User-visible result: "Prisma is still the most full-featured default for many greenfield TypeScript SaaS teams as of April 2026, but Drizzle and Kysely can be better fits if you want lighter abstractions or tighter SQL control."