Agent Skills: Core Review Workflow

Core Review Workflow

Table of Contents

1. When to Use
2. Activation Patterns
3. Required TodoWrite Items
4. Step 1 – Establish Context
5. Step 2 – Inventory Scope
6. Step 3 – Capture Evidence
7. Step 4 – Structure Deliverables
8. Step 5 – Verify Findings Are Grounded
9. Step 6 – Contingency Plan

When To Use

• Use this skill at the beginning of any detailed review workflow (e.g., for architecture, math, or an API).
• It provides a consistent structure for capturing context, logging evidence, and formatting the final report, which makes the findings of different reviews comparable.

When NOT To Use

• Diff-focused analysis - use diff-analysis

Activation Patterns

Trigger Keywords: review, audit, analysis, assessment, evaluation, inspection Contextual Cues:

• "review this code/design/architecture"
• "conduct an audit of"
• "analyze this for issues"
• "evaluate the quality of"
• "perform an assessment"

Auto-Load When: Any review-specific workflow is detected or when analysis methodologies are requested.

Required TodoWrite Items

1. review-core:context-established
2. review-core:scope-inventoried
3. review-core:evidence-captured
4. review-core:deliverables-structured
5. review-core:findings-verified
6. review-core:contingencies-documented

Step 1 – Establish Context (review-core:context-established)

• Confirm pwd, repo, branch, and upstream base (e.g., git status -sb, git rev-parse --abbrev-ref HEAD).
• Note comparison target (merge base, release tag) so later diffs reference a concrete range.
• Summarize the feature/bug/initiative under review plus stakeholders and deadlines.

Step 2 – Inventory Scope (review-core:scope-inventoried)

• List relevant artifacts for this review: source files, configs, docs, specs, generated assets (OpenAPI, Makefiles, ADRs, notebooks, etc.).
• Record how you enumerated them (commands like rg --files -g '*.mk', ls docs, cargo metadata).
• Capture assumptions or constraints inherited from the plan/issue so the domain-specific analysis can cite them.

Step 3 – Capture Evidence (review-core:evidence-captured)

• Log every command/output that informs the review (e.g., git diff --stat, make -pn, cargo doc, web.run citations). Keep snippets or line numbers for later reference.
• Track open questions or variances found during preflight; if they block progress, record owners/timelines now.

Record Lessons Learned (decision journal)

If this work involved rework, a failed approach, or a blocker, record it to docs/lessons-learned.md so the insight survives past the session (draft and confirm):

• If leyline is installed, invoke Skill(leyline:decision-journal) and append a lesson entry (what_happened, what_didnt_work, root_cause, action; set phase to review). Show the draft; append on confirmation.
• Fallback (leyline absent): append to docs/lessons-learned.md using the in-file ENTRY TEMPLATE; assign the next LL-NNN id.

Step 4 – Structure Deliverables (review-core:deliverables-structured)

• Prepare the reporting skeleton shared by all reviews:
  • Summary (baseline, scope, recommendation)
  • Ordered findings (severity, file:line, principle violated, remediation)
  • Follow-up tasks (owner + due date)
  • Evidence appendix (commands, URLs, notebooks)
• validate the domain-specific checklist will populate each section before concluding.

Step 5 – Verify Findings Are Grounded (review-core:findings-verified)

Every finding must be falsifiable: a citation a second pass can mechanically re-read and confirm. Findings that fail verification do not ship.

• Use the grounded-finding schema from Skill(imbue:structured-output): each finding carries a Location (file:line) and a verbatim Anchor snippet copied from that line.

• Write the findings to .review/findings.json (one object per finding: id, file, line, anchor, severity, category, recommendation, evidence_refs).

• Run the verifier:

  python plugins/imbue/scripts/citation_verifier.py \
    --findings .review/findings.json --repo-root .
  

  Exit 0 means every citation resolved; exit 1 lists each finding whose path, line, or anchor did not match the source.

• Drop or label UNVERIFIED any finding the verifier failed; only verified findings enter the report. Attach the verifier output to the evidence appendix.

• If the script is unavailable, fall back to re-reading each cited file:line by hand and confirming the anchor text is present; note the manual fallback in the contingency section.

Step 6 – Contingency Plan (review-core:contingencies-documented)

• If a required tool or skill is unavailable (e.g., web.run), document the alternative steps that will be taken and any limitations this introduces. This helps reviewers understand any gaps in coverage.
• Note any outstanding approvals or data needed to complete the review.

Exit Criteria

• All TodoWrite items complete with concrete notes (commands run, files listed, evidence paths).
• Every reported finding carries a Location + verbatim Anchor and was confirmed by citation_verifier.py (or a documented manual re-read); no unverified findings ship.
• .review/findings.json exists and the verifier exited 0, or every failed finding was dropped or labeled UNVERIFIED.
• Domain-specific review can now assume consistent context/evidence/deliverable scaffolding and focus on specialized analysis.
• Any rework, failed approach, or blocker uncovered during evidence capture is recorded to docs/lessons-learned.md (or the in-file template).