Revisit a previous deep dive, research brief, or Grain with fresh evidence and first-principles reasoning; identify what changed, what still holds, and what the earlier investigation missed, then update the existing artifact. Use for a fresh check, research refresh, or reassessment of prior findings.
Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
Are you the author of Deep Dive Refresh?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/greenfield-inc-deep-dive-refresh)
---
name: deep-dive-refresh
description: Revisit a previous deep dive, research brief, or Grain with fresh evidence and first-principles reasoning; identify what changed, what still holds, and what the earlier investigation missed, then update the existing artifact. Use for a fresh check, research refresh, or reassessment of prior findings.
---
# Deep Dive Refresh
Reassess a previous investigation as of now. Reconstruct the reasoning, test it against current evidence, and leave the reader with an updated answer and a clear account of material changes. This is an on-demand workflow; invoking it does not create a recurring schedule.
## Use the shared research foundation
Read and apply [Deep Dive](../deep-dive/SKILL.md) once. It supplies source discovery, Bright Data access and scraper selection, coverage tracking, evidence standards, artifact handling and fresh-reader review. Reuse credentials, cached sources and prior setup decisions; do not restart the access interview. Read [Tech Deep Dive](../tech-deep-dive/SKILL.md) as an additional lens when the prior investigation depends on implementation mechanics or technology performance, without creating a second research pass. The refresh question and budget govern scope: import the relevant technical analysis, not an automatic full-landscape survey. Specialist discovery requirements apply proportionally to the claims being refreshed; broaden only when the question or changed evidence warrants it.
These skills ship together; resolve references relative to their bundled directories. If a dependency is missing, locate it or report the incomplete installation. Do not pretend to have run the full workflow without its instructions.
## Recover the question and baseline
Locate the previous Grain/brief, source ledger, research date, revisions and any relevant user decisions. Reuse its workspace ID; do not create a replacement simply because the request is a refresh. If the prior artifact is ambiguous or unavailable, ask for its identity/content while gathering independent evidence that is useful without it. Do not fabricate the previous conclusions.
Identify the original question, audience, scope, geography/timeframe and intended decision. Preserve later user corrections. Note any real change in the user's needs separately from changes in external facts.
Extract the important claims, explanations, assumptions, numbers, recommendations and unresolved leads into a working inventory. Include the original source/version/date and why each claim mattered. Establish this baseline before editing, using a saved revision or local copy so the prior reasoning remains recoverable.
## Rebuild the reasoning from first principles
Before searching merely for updates to named sources, ask:
- What problem, question or decision were we actually trying to resolve?
- What facts, constraints, mechanisms or incentives must hold for the answer to be true?
- Which conclusions were supported by observations, which were inferred, and which relied on popularity, analogy or convenient assumptions?
- What counterexample or changed condition would invalidate each consequential conclusion?
- If encountering the subject today, what options, explanations or sources would we consider without being anchored to the old shortlist?
Use these questions to direct the fresh pass. For a technical claim, retrace the input → decision → action → outcome and dependencies. For other subjects, examine the causal explanation, population, context and alternative explanations appropriate to the question. First-principles reasoning is a way to expose assumptions; it does not replace empirical evidence.
## Recheck and expand
Prioritize claims by consequence, uncertainty and likelihood of change. Verify volatile facts and the assumptions carrying recommendations against current primary sources. Re-open the relevant original evidence, inspect new versions or implementation changes when material, and distinguish publication dates from the dates events or measurements occurred.
Use caches for provenance and stable historical evidence; do not treat an old scrape as a current verification. A new scrape timestamp alone does not prove the underlying content changed. Record the old and new observation dates.
Search beyond the old source list. Check new terminology, credible counterexamples, relevant directories/collections, practitioner discussions, competing explanations and previously inaccessible or uninvestigated leads. Use the shared scraper catalog for the sources that can answer the question. Do not limit the refresh to recent social posts or the same familiar authors.
Choose depth proportionally: a narrow fresh check can focus on consequential claims and their dependencies; a changed premise or major contradiction may require a broader reassessment. Track which claims were actually rechecked. Stop within the user's limits and identify unchecked areas, rather than relabeling the entire old report current.
## Record what the evidence changes
Classify important findings with supporting sources and dates:
| Status | Meaning |
| --- | --- |
| Reconfirmed | Current evidence still supports the earlier claim in its stated scope |
| Changed | An input, mechanism, result or relevant condition has materially changed |
| Contradicted / corrected | Evidence undermines the claim, or the earlier analysis was mistaken |
| Newly surfaced | Relevant evidence or an option was previously absent; say whether it is new in the world or merely newly found |
| Unresolved | Evidence is missing, conflicting or unavailable; absence is not disproof |
| Not rechecked | Retained historical material that was outside this refresh's coverage |
Do not infer that a claim was false simply because a link broke or a project disappeared. Preserve the earlier observation as dated history where appropriate. Distinguish an original analytical error from a genuine external change.
Recompute affected comparisons, cost models and priorities with stated baselines. Trace how a changed assumption affects the final conclusion; do not update a statistic while leaving an incompatible recommendation untouched. Retain a conclusion only if its reasoning still holds. It is valid to find no material change after an actual check; explain the scope checked without inventing novelty.
## Update the reader's answer
Lead with the current answer, the most important changes, and what those changes mean for the user's question. Include a concise dated “What changed” section with before → now → why it matters and sources. Explain any corrected first-principles model with a concrete example or diagram if helpful. Keep the primary reading path understandable without requiring the old report.
Update the existing artifact coherently: summary, definitions, examples, diagrams, calculations, recommendations and source receipts must agree. Preserve useful prior evidence and source attribution; move superseded conclusions into dated history or an expandable change record. Mark partial refresh coverage and unresolved questions explicitly. Do not silently overwrite uncertainty with a new confident narrative.
Apply the foundation's mandatory fresh-reader review to the revised artifact. Add teach-back questions: What changed and why? Which changes are new facts versus corrections to earlier reasoning? What still holds? What was not rechecked? Does the recommended next step follow from the current evidence? The reviewer receives only the artifact and audience, not this conversation or expected answers; the coordinator owns delegation.
Verify the saved revision and meaningful interactions/citations. Return the existing Grain link/path, the material changes or supported no-change result, and remaining coverage limits. Keep baseline, current source ledger and unresolved leads available for the next revisit. No application implementation, purchasing, outreach or scheduling is implied by a research refresh.