Assay Report · 2026-04-26 · sample

We found one thing
worth looking at first.

The report opens with the strongest concrete finding, not a dashboard. Read top to bottom. Each section ends with a recommended action.

assay_id
2026-04-26-local
scope
project · ~/.claude
assets scanned
46 agents · 22 skills · 17 commands · 9 hooks · 4 MCPs
findings
3 high · 4 medium · 7 low
model claims
5 proposed · 0 confirmed
harness
draft · not yet saved
Contents · 9 sections
01Inventorywhat was found—
02Influence mapwhat shapes Claude's behavior6 conflicts
03Hygieneduplicates, stale, malformed, orphaned4 items
04Trust & safetyrisky hooks · broad permissions · MCPs2 items
05Context qualityconflicts · contamination · vague rules5 items
06Operator modelinferred heuristics, workflows, boundaries5 claims
07Navigation harnesscontext Claude Code can use, with approvaldraft
08Workflow modesreplacement · augmentation · reimagination9 mapped
09Progression planone primitive · one experiment · one rerun signal1 active
02 · Influence map

What is shaping Claude's behavior, in order.

left · global · earliest  →  right · local · latest
user · ~/.claude
project · ./.claude
runtime · plugins / hooks
CLAUDE.md
autonomy: encouraged
./CLAUDE.md
autonomy: review-first  ⚠ conflict
PreToolUse · git-guard.sh
agents/ · 46 files
agents/ · 6 files (3 shadow user)
PostToolUse · format.sh
skills/ · 22
skills/ · 4
MCP · ra-collector (local)
hooks/ · 9
.claude/settings.local.json  ⚠ shadows user permissions
MCP · web-search (network)

Two surfaces present opposing rules to the same tool call. Claude resolves the tie by recency, not by intent. See finding 01.

03–05 · Top findings · 3 high severity
finding · 01  ·  context
Global autonomy rule conflicts with this project's review behavior.
high
source
~/.claude/CLAUDE.md · L42
./CLAUDE.md · L18
evidence
Global: "Default to autonomous edits unless asked to review."
Project: "Review and preserve before editing strategy docs."
Both surfaces reach Claude on every session in this project.

Why it matters. Effective behavior depends on conversational drift rather than a clear rule. Review behavior is non-deterministic.

Recommended action. Move the autonomy rule into the projects where you want it. Keep review conservatism local here.

confidence · high reversibility · safe detector source · EP 003
finding · 02  ·  trust
A pre-tool hook executes a shell pattern that touches .git internals.
high
source
~/.claude/hooks/pre-commit.sh · L12
evidence
Hook runs: "git reflog --all | grep -v origin"
Triggered on PreToolUse for any Bash invocation.
No confirmation gate before execution.

Why it matters. Hook is broadly scoped and reads .git history that may contain reverted secrets.

Recommended action. Narrow trigger to commit-related tool calls. Add a confirmation step.

confidence · medium reversibility · review-first

12 more findings in the full report. Hygiene: 4 items · Trust: 2 · Context quality: 5 · Low severity: 7.

06 · Operator model · proposed claims

Five claims drafted from your working environment.

The product does not assert who you are. You confirm, edit, or reject. Only confirmed claims are eligible for the navigation harness.

preference · inferred
You prefer strategy changes to preserve prior versions rather than overwrite them.
./CLAUDE.md asks for side-by-side narratives. Recent planning docs create *-v2 companions.
confidence · medium
trust boundary · inferred
Do not run destructive git operations without explicit approval.
AGENTS.md L8. Pre-tool hook git-guard.sh blocks force-push on every session.
confidence · high

3 more proposed claims. None confirmed yet. Confirmation is required before harness generation.

07 · Navigation harness · draft

Generated from confirmed claims only. None confirmed yet.

This is what the harness will look like once you confirm claims. It does not impersonate you. It tells Claude Code where to ask before acting, what to preserve, and what to stop short of.

# Operator Navigation Context — draft # Generated: 2026-04-26 · not yet saved · no claims confirmed ## Confirmed boundaries # (will appear here after confirmation) ## Working preferences # (will appear here after confirmation) ## Pending confirmation preference_draft: preserve prior strategy versions before replacement boundary_draft: do not run destructive git operations without approval status: requires confirmation · not active

Harness is saved only when you save it. Claude Code does not receive this context until installation is confirmed.

08 · Workflow modes

Per-workflow. No aggregate score.

9 workflows · first assay · no delta yet

Each workflow is classified as replacement, augmentation, or reimagination on its own terms. Deltas appear on the second assay. Regression is reported, not hidden.

codebase_navigation
augmentation
Claude assists with search but operator still navigates files.
—
strategy_editing
augmentation
Claude drafts faster but judgment preservation is manual.
—
pr_review
replacement
Same job, faster. Operator still reads all output.
—
incident_response
replacement
Drafts faster. Operator still owns triage and decision.
—

5 more workflows in the full table. Trajectory section appears on the second assay.

09 · Progression plan

One primitive. One week. One rerun signal.

rerun signal: 2026-05-03

The report does not end with advice. It ends with a contract: which workflow, which model primitive, which harness change, which experiment, and what the next assay should show if the change took.

Progression plan · 2026-04-26 strategy_document_editing
current mode Augmentation — Claude drafts and edits faster, but the operator still has to preserve strategic judgment manually.
primary blocker No preservation boundary. Claude can rewrite live strategy without making comparison possible.
next primitive Judgment-preservation boundary.
harness change nav.preserve_strategy_versions — preserve prior versions or create companion comparison documents before replacement.
experiment Run three strategy edits with this rule active · 2026-04-26 → 2026-05-03.
rerun signal Claude proposes comparisons before overwrites and produces cleaner decision trails.
Request your own assay → Join to get monthly reports
This is a sample

Your report will look like this.

This sample uses Deprecated's working environment as its dataset. Your assay will run against your own ~/.claude, your own agents, your own hooks. The findings, claims, and harness will be yours.

nota screenshot of a dashboard
notgeneric advice you could find anywhere
isevidence-backed findings from your actual working system
Request your assay →