agentleFS
mediajunkiemediajunkie/piper-morgan-product/.claude/skills/audit-cascade/SKILL.md

audit-cascade

Skill8 starsChanged 11 days agoApache-2.0View on GitHub

Perform systematic audit-and-correct between phases of multi-step work. Use when PM says "audit cascade", before transitioning from issue to gameplan, gameplan to agent prompts, or prompts to execution. Catches drift before it accumulates.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 7 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill when:

    line 25
  2. Distinctive:

    Create a markdown table comparing the template requirements against the document:

    line 100
  3. Distinctive:

    Before proceeding to the next phase:

    line 119
  4. Distinctive:

    Do NOT proceed with unfixed items

    line 122
  5. Distinctive:

    If a requirement seems inapplicable: STOP and ask PM. You cannot decide this yourself.

    line 124
  6. Distinctive:

    Only after all items are ✅, proceed to:

    line 135
  7. Distinctive:

    After completing an audit:

    line 181

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

audit-cascade

Perform systematic audit-and-correct steps between phases of multi-step work, using templates as checklists.

Reference: Pattern-049: Audit Cascade

When to Use

Use this skill when:

  • Writing or reviewing a GitHub issue before creating a gameplan
  • Writing or reviewing a gameplan before creating agent prompts
  • Writing or reviewing agent prompts before execution
  • Any multi-step work where drift could accumulate

Critical rule: You have ZERO AUTHORIZATION to mark any template requirement as "optional", "N/A", or "not applicable" without explicit PM approval. If a requirement seems inapplicable, STOP and ask.


The Cascade (3 Phases, 3 Audit Gates)

Write Issue → AUDIT → Write Gameplan → AUDIT → Write Prompts → AUDIT → Execute

3 audit gates catch drift before it compounds into the next phase. Each phase has a write step followed by an audit step.


Procedure

Step 0 (preamble): Confirm your worktree

Audit-cascade work is substantive by definition (Phase 0 audit + gameplan + audit + execute + audit), so the standing worktree-default applies. On Amber (Model A): confirm you're in your stable per-agent worktree (~/Development/piper-morgan-worktrees/{role}), not the shared checkout. On Desktop (Model B): confirm the session's ephemeral auto-worktree is active. Skip only if the audit work is short enough to qualify as an exception (single-file Phase 0 audit + immediate disposition, no implementation).

(Added 2026-08-23, closing a ~3-month-old disposition — mailboxes/cio/sent/memo-cio-to-host-cc-ppm-docs-arch-ceo-audit-cascade-worktree-preamble-disposition-2026-05-15.md. Original wording assumed the pre-Amber ephemeral-worktree-per-session model; reworded here for Model A's stable-worktree-reuse reality rather than landing three-month-stale text unchecked against current architecture.)

Step 1: Identify the Phase

What are you auditing?

PhaseTemplate Location
Issue.github/issue_template/ (feature.md, bug_report_alpha.md, e2e-bug.md)
Gameplanknowledge/gameplan-template.md
Agent Promptsknowledge/agent-prompt-template.md

Step 1b: Agent Prompts phase only — audit the dispatch tier

Added 2026-09-14, PM's ruling (via Janus, closing the week's Fable-ceiling investigation): when the phase being audited is Agent Prompts and the prompt dispatches a subagent, the audit matrix gets one more row. The principle, in PM's own words: "A subagent absolutely can run Opus but I think should rarely need Fable if the agent dispatching it is already using Fable to run the show, plan the work, write the gameplan, draft the prompt, audit everything, and hold the subagent accountable." The dispatcher holds the judgment; the subagent executes against it. Tier follows from that — a subagent doesn't need to re-derive judgment the dispatcher already applied and put into the prompt.

The check, worded exactly as PM specified: does this prompt carry enough specification that a cheaper tier could execute it? If not, is that deliberate, or is the prompt underspecified? That question improves the prompt whether or not it changes the tier, which is what makes it worth asking every time rather than only when a Fable dispatch looks suspicious.

Rough tier guide (not a rule to apply mechanically — judgment on the actual unit of work): Haiku for mechanical, fully-specified work (sweeps, inventories, format conversions — success is checkable without judgment); Sonnet for bounded implementation against clear criteria (most issue-level work); Opus for genuinely hard reasoning inside the unit (tricky debugging, a design call deliberately delegated, adversarial review — explicitly fine, not a ceiling to apologize for); the dispatcher's own top tier (Fable when PM ruled) rarely, and if you reach for it, treat that as a signal to look at the prompt first, not the tier.

Do not skip this for the model your OWN dispatching session runs. A fan-out silently inherits the dispatcher's model unless the model parameter is passed explicitly — this was the actual mechanism behind the incident that produced this ruling (one seat's 30 dispatches exhausted a shared tier ceiling that two unrelated, non-dispatching seats then paid for). State the tier explicitly in the audit matrix row rather than leaving it implicit.

Step 2: Create Audit Matrix

Create a markdown table comparing the template requirements against the document:

## Audit: [Document Name] against [Template Name]

| Template Requirement | Status | Notes |
|---------------------|--------|-------|
| [Requirement 1] | ✅ / ⚠️ / ❌ | [Details] |
| [Requirement 2] | ✅ / ⚠️ / ❌ | [Details] |
...

Status legend:

  • ✅ Present: Requirement fully satisfied
  • ⚠️ Partial: Requirement addressed but incomplete or unclear
  • ❌ Missing: Requirement not addressed

Step 3: Fix ALL Issues

Before proceeding to the next phase:

  • Fix every ⚠️ (partial) item
  • Fix every ❌ (missing) item
  • Do NOT proceed with unfixed items

If a requirement seems inapplicable: STOP and ask PM. You cannot decide this yourself.

Step 4: Document the Audit

Save the audit matrix as a working document:

  • Location: dev/YYYY/MM/DD/
  • Naming: {issue-number}-{phase}-audit.md
  • Example: 583-gameplan-audit.md

Step 5: Proceed to Next Phase

Only after all items are ✅, proceed to:

  • Issue audit complete → Write gameplan
  • Gameplan audit complete → Write prompts
  • Prompts audit complete → Execute

Example Audit Matrix

## Audit: #583 Gameplan against gameplan-template.md

| Template Requirement | Status | Notes |
|---------------------|--------|-------|
| Issue number referenced | ✅ | #583 in header |
| Problem statement | ✅ | "Chat not persisting on refresh" |
| Five-whys analysis | ⚠️ | Only 3 whys - need 2 more |
| Success criteria | ✅ | 3 measurable criteria |
| Test strategy | ❌ | Missing entirely |
| Phases with estimates | ✅ | 4 phases, 2-3 hours |
| Rollback plan | ❌ | Missing |
| Dependencies listed | ✅ | None identified |

### Action Required
Before proceeding to prompts:
1. Complete five-whys to root cause
2. Add test strategy section
3. Add rollback plan

Anti-Patterns

Don't Do ThisWhyDo This Instead
Skip audit under time pressureThis is exactly when drift occursAudit takes 5-10 min; rework takes hours
Mark requirements as "N/A" yourselfYou may be wrong about applicabilityAsk PM if requirement seems inapplicable
Audit without the template openYou'll miss requirementsAlways have template visible
Audit multiple timesOne thorough audit is sufficientDo it once, do it right
Proceed with ⚠️ or ❌ itemsDrift compounds at each phaseFix everything before proceeding

Quality Checklist

After completing an audit:

  • Template was open during entire audit
  • Every template requirement has a row in the matrix
  • No ⚠️ or ❌ items remain unfixed
  • No requirements marked "N/A" without PM approval
  • Audit matrix saved to dev/YYYY/MM/DD/
  • Ready to proceed to next phase

Key Insight

LLMs struggle to follow templates while creating but excel at auditing against templates afterward.

This asymmetry is why audit cascade works: instead of trying harder to follow templates during creation (which doesn't work), we accept imperfect creation and systematically audit afterward (which does work).

The word "audit" triggers systematic verification behavior. Use it.


Pattern Families

This skill is part of the Completion Theater Family (045/046/047/049):

  • Pattern-045: Diagnoses the gap (tests pass, users fail)
  • Pattern-046: Enforces 100% completion (no expedience rationalization)
  • Pattern-047: Enables pause when uncertain (Time Lord Alert)
  • Pattern-049: Audits at every phase boundary — this skill operationalizes 049

These patterns form a reinforcing system. Audit cascade (049) is the operational mechanism that catches drift before it compounds. Apply the full family for multi-phase work.

See PATTERN-FAMILIES.md for full family index.


Related

  • Pattern-049: Audit Cascade - Full pattern documentation
  • Pattern-046: Beads Completion Discipline - No expedience rationalization
  • create-session-log - Session documentation
  • close-issue-properly - Issue closure with evidence

Skill version: 1.0 Created: 2026-01-23 Scope: Cross-role

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.