- Home
- Skills
- Testing & Debugging
- 07 — Catastrophe Shepherd · Complete Edition
Works with the AI Agents you already use
07 — Catastrophe Shepherd · Complete Edition
Preserves controlled fault exploration, resource-pressure methods and failure reduction, then adds practical diagnostic planning, bug reports and automation recovery.
$25
07 — Catastrophe Shepherd · Complete Edition
Example session with this skill installed
Everyday request
The upload works sometimes and fails sometimes. Use the observations I have to narrow the possibilities and tell me the safest useful test to try next.
For a reproducible local example, use the included synthetic structured fixture:
{
"hypotheses": [
{
"id": "network",
"predicts": {
"other-site": false,
"fresh-session": false
}
},
{
"id": "session",
"predicts": {
"other-site": true,
"fresh-session": true
}
},
{
"id": "service",
"predicts": {
"other-site": true,
"fresh-session": false
}
}
],
"tests": [
{
"id": "other-site",
"instruction": "Check a permitted unrelated site to distinguish a general connection problem.",
"minutes": 1
},
{
"id": "fresh-session",
"instruction": "Try the already authorized site in a fresh signed-in session after preserving the draft.",
"minutes": 2
}
],
"observed": {
"other-site": true
}
}
Run the class's packaged helper on that fixture. Show the readable result, full structured output, executed Thirsty calculation and the limits of what was demonstrated.
- Read your context and instructions
- Compiled the output
Catastrophe Shepherd — real local worked example
Example type: Actual execution of the packaged structured helper on synthetic input. This is not a claimed live customer session or a full specialist-model benchmark.
Readable work product
Hypotheses still compatible with observations: session, service
Next test candidate: Try the already authorized site in a fresh signed-in session after preserving the draft. (estimated 2 minutes)
More evidence required; no cause confirmed
Complete structured result
{
"next_tests": [
{
"cost_minutes": 2,
"discrimination": 1,
"id": "fresh-session",
"instruction": "Try the already authorized site in a fresh signed-in session after preserving the draft."
}
],
"remaining_hypotheses": [
"session",
"service"
],
"scope": "Recommendations only; no device or service modified",
"status": "More evidence required; no cause confirmed"
}
Executed Thirsty-lang calculation
module conclave: core
drink result = 3 - 2
pour result
The helper checked the language result independently, encoded the complete structured result using the published TSCG-B codec, decoded it and compared the text. Running the standalone helper generates the result files and execution receipt. The example shown here was produced on the supplied synthetic input.
The deeper specialist use
Request: Investigate a job that might run twice after a lost acknowledgement. Design a bounded disposable test, a clean control and a reduction strategy.
Illustrative specialist response: State the invariant: one effect per job. Limit the fixture to declared fault points and retry attempts. Compare the clean path with acknowledgement loss, then remove unnecessary timing faults while preserving an observed duplicate-effect predicate. Record effects independently of worker acknowledgements and reconcile uncertain commitments before another retry. The deliverable is a replayable failure schedule plus evidence and recovery findings when actually executed—not a declaration that a production defect has already been found.
The preceding specialist response is an authored illustration, not an executed external action. The original class procedures and supporting skills remain included for an actual authorized assignment.
Next step
Replace the fixture with your own task data after checking that it fits the helper's input model, or use the conversational workflow directly with your AI host. Use a fresh output directory; the helper refuses to overwrite an existing result. Review factual classifications and missing inputs before relying on the result.
Connects securely to your tools. The creator never sees your data.
About this skill
Catastrophe Shepherd — Complete Edition
Hunt the rare failure. Stop the recurring headache.
Preserves controlled fault exploration, resource-pressure methods and failure reduction, then adds practical diagnostic planning, bug reports and automation recovery.
One class. 47 equipped skills. The original specialist method plus six practical everyday workflows. $25 as a one-time purchase.
For recurring software glitches, unreliable automations, and troubleshooting that goes in circles. For deeper work, the class retains its original specialist equipment and source-defined method: Which bounded fault schedule exposes a reproducible failure?
What you actually receive
- All 40 original skills, including the specialist entrypoint, original supporting procedures, language workbench and evaluation lab.
- Six added everyday skills, listed in full below, and the new structured helper and worked example.
- 1 restored named dependency/dependencies: epistemic-status-enforcer.
- Class methods, contracts, source references, standalone task-packet tooling, executable helpers and worked examples, delivered through your connected AI.
- Standalone Thirsty-lang and TSCG-B tooling that uses the real
thirsty-lang==0.9.0runtime. Code execution requires Python 3.10+, a host execution tool and that separately installed package in the same environment. No runtime wheel is delivered through Agensi.
The specialist class: the original depth stays equipped
- Bind the allowed test environment and record a restorable baseline.
- Generate finite fault schedules over explicit operation checkpoints.
- Confirm each injection is permitted and has a stop condition.
- Run the real test path with controlled interruptions, timing or resource failure.
- Observe effects and durable state independently of the injected worker.
- Reconcile uncertain commitments before repeating or restoring dependent work.
- Return a minimized reproducible schedule, observed failure and verified recovery evidence.
Specialist deliverable: fault_schedule_evidence. Its source-defined result fields are baseline_binding, fault_schedule, injection_receipts, failure_trace, settlement_records, recovery_evidence. These guide the work product; an actual outcome requires the task inputs, authorized tools and evidence.
Six everyday workflows — the documented practical use cases
1. Narrow a problem that only happens sometimes
Record exact symptoms and context, propose competing hypotheses, and change one controllable condition at a time. Preserve successful controls.
2. Investigate a failing personal or work automation
Map triggers, input, intermediate state and final effect. Check whether a retry could duplicate the effect before suggesting one.
3. Choose a useful diagnostic step
Compare hypothesis predictions and supplied observations, then prefer a discriminating low-cost test. Recommend only steps within the user’s capability and actual authorization.
4. Write a bug report someone can reproduce
Provide observed and expected behavior, exact steps, environment and sanitized evidence. Keep suspected causes labeled as hypotheses.
5. Reproduce a failure in a disposable environment
Define an invariant and finite fault schedule, run a no-fault control, and minimize the actual failing sequence. Never inject faults into live systems by default.
6. Check whether a fix restored the important behavior
Repeat the failing case and the clean control, inspect persistent state, and report unresolved side effects. A single passing run is not a universal reliability claim.
Example 1 — what a specialist assignment can look like
You ask: Investigate a job that might run twice after a lost acknowledgement. Design a bounded disposable test, a clean control and a reduction strategy.
Illustrative response: State the invariant: one effect per job. Limit the fixture to declared fault points and retry attempts. Compare the clean path with acknowledgement loss, then remove unnecessary timing faults while preserving an observed duplicate-effect predicate. Record effects independently of worker acknowledgements and reconcile uncertain commitments before another retry. The deliverable is a replayable failure schedule plus evidence and recovery findings when actually executed—not a declaration that a production defect has already been found.
This example illustrates the class method. It is not an assertion that the proposed external investigation, model trial or system change has been executed.
Example 2 — an everyday request
The upload works sometimes and fails sometimes. Use the observations I have to narrow the possibilities and tell me the safest useful test to try next.
The work product: A ranked investigation plan, eliminated hypotheses, reproducible conditions where observed, and a recovery-aware next step.
The included local helper demonstrates a bounded part of this workflow. Its actual run on synthetic structured input produced:
Hypotheses still compatible with observations: session, service
Next test candidate: Try the already authorized site in a fresh signed-in session after preserving the draft. (estimated 2 minutes)
More evidence required; no cause confirmed
See the demo section below for its complete input, structured result and executed Thirsty calculation.
Example 3 — using the class template on your own material
Use Catastrophe Shepherd — Complete Edition.
Goal: [the concrete result I need]
Material: [paste or attach the actual records, notes, code or facts]
Scope: [what may be inspected or changed]
Constraints: [time window, dependencies, formats and important conditions]
Deliver: A ranked investigation plan, eliminated hypotheses, reproducible conditions where observed, and a recovery-aware next step.
Use the original specialist method when the problem requires it.
Ask only for missing facts that materially change the result.
Show what you actually did and what still needs a decision.
For a simple request, use only the relevant everyday mode. For advanced work, select the original specialist procedures. Keeping the equipment available does not require loading every skill into every interaction.
Every skill available in this pack (47 total)
-
algorithm-design-forge — Invent and implement problem-specific algorithms with verified invariants and adversarial tests.
-
architecture-decision-recorder — Edition XXII source-attributed supplemental agent procedure: Architecture choice and tradeoff recording.
-
build-reproducibility-receipts — Create verifiable local build and test receipts linking source snapshots, commands and exact artifact bytes.
-
catastrophe-shepherd-automation-rescue — Investigate a failing personal or work automation.
-
catastrophe-shepherd-bug-report-builder — Write a bug report someone can reproduce.
-
catastrophe-shepherd-fault-fixture — Reproduce a failure in a disposable environment.
-
catastrophe-shepherd-intermittent-problem — Narrow a problem that only happens sometimes.
-
catastrophe-shepherd-recovery-check — Check whether a fix restored the important behavior.
-
catastrophe-shepherd-safe-next-test — Choose a useful diagnostic step.
-
ceta-authority-timeline-autopsy — Reconstruct possible CETA event and authority histories across occurrence, recording, observation, and validity time, and identify distinguishing evidence.
-
ceta-class-router — Choose a Thirstforged Conclave sub-agent class for a CETA task, prepare its bounded assignment and exact skill inventory, and distinguish instruction packets from native profile activation.
-
ceta-evidence-cross-examination — Cross-examine a Thirstforged Conclave result against raw observations, exact source bindings, expected predicates, and actual reviewer identity before integration.
-
ceta-frame-authority-confusion — Build bounded TSCG/TSCG-B and policy specimens that separate valid framing and checksums from legitimate actor, scope, freshness, and execution authority.
-
ceta-incentive-failure-societies — Run bounded deterministic agent-society simulations that expose incentive conflicts, metric gaming, missing stakeholders, and collective failure despite local compliance.
-
ceta-interrupted-effect-reconstruction — Reconstruct interrupted or cancelled effects from receipts and effect-owner observations, preserving unknown outcomes and preventing unexamined retries.
-
ceta-rare-failure-cultivation — Search reproducible crash, cancellation, delay, ordering, and partial-progress schedules in disposable fixtures to expose rare CETA failures.
-
ceta-resource-pressure-reduction — Minimize a discovered fault or resource-pressure failure into a reproducible causal case while preserving measurement provenance and uncertainty.
-
ceta-utf-contract-map — Map a CETA task to the six Universal Thirsty Family tiers and exact isolated Thirsty-Lang 0.9.0 tools, preserving the difference between analysis, integrity, and authority.
-
codex-approval-trace — Use for trace decision, requester, scope, expiry, and evidence receipts.
-
codex-authority-resolver — Use for resolve governing clause identity and applicable authorization without assuming precedence.
-
codex-autonomy-review — Use for evaluate competence evidence and legitimate authority for autonomy changes.
-
codex-build-reproducer — Use for inventory local build ingredients and reproducibility conditions.
-
codex-claim-integrity — Use for prevent observed/inferred/proposed/implemented states from being conflated.
-
codex-disaster-recovery — Use for review backup evidence and recovery runbooks without performing recovery.
-
codex-embodied-safety — Use for review physical or sensor actions, limits, and independent stop pathways.
-
codex-evidence-ledger — Use for validate referenced evidence paths and source identity.
-
codex-failure-reproducer — Use for create reproducibility inputs from logs and source; do not invent successful runs.
-
codex-observability-auditor — Use for inspect recorded monitoring/log evidence without claiming telemetry deployment.
-
codex-offline-readiness — Use for audit offline-first configuration and absence of required live-saas assumptions.
-
codex-recovery-rollback — Use for assess rollback artifacts, durable continuity and recovery prerequisites.
-
codex-repo-preservation — Use for protect untracked work, check provenance and refuse destructive cleanup.
-
codex-root-cause-debugger — Use for inspect actual exceptions, syntax and git state to isolate observed failures.
-
codex-runtime-health — Use for inspect health-related logs and configs as static evidence, not live monitoring.
-
codex-scope-guardian — Use for prevent invented or silently widened objectives and repository scope.
-
codex-source-integrity — Use for check bundled source document checksums and protect observed provenance.
-
conclave-catastrophe-shepherd — Preserves controlled fault exploration, resource-pressure methods and failure reduction, then adds practical diagnostic planning, bug reports and automation recovery.
-
conclave-evaluation-lab — Design held-out comparisons of the equipped agent and baseline workflows with matched models, tools, budgets, and recorded evidence.
-
conclave-language-workbench — Use Thirsty-lang as the primary local language and TSCG-B to encode/decode exact UTF-8 task and result text.
-
concurrency-performance-engineer — Implement and measure safe concurrency, latency and memory-budget improvements.
-
epistemic-status-enforcer — Apply explicit epistemic status to meaningful claims.
-
polyglot-process-bridge — Implement a robust polyglot subprocess protocol using stdin/stdout, JSON Lines, structured errors, timeouts and shutdown.
-
polyglot-stack-selector — Select practical combinations of programming languages and runtime boundaries for one application.
-
proof-obligation-generator — Edition XXII source-attributed supplemental agent procedure: Execution proof obligation construction.
-
property-fuzz-test-engineer — Implement executable property-based tests, deterministic fuzz harnesses and failure minimization for source code.
-
python-production-coder — Implement and repair working Python applications, libraries and asynchronous services.
-
thirsty-work-orchestrator — Edition XXII source-attributed supplemental agent procedure: Role graph scheduling and evidence handoff.
-
type-contract-architect — Create implementable domain types and API contracts before coding.
Thirsty-lang and TSCG-B in practice
Run the packaged helper with Python 3.10+ to generate a readable result, structured JSON, the executed .thirsty calculation, a binary .tscg-b result and an execution receipt. The helper executes supported arithmetic in Thirsty-lang and independently checks the result; Python handles structured input and the domain algorithm. TSCG-B preserves result text through binary encode/decode. A valid frame does not itself prove the text is true or authorize an action.
Preservation and validation
Every original skill identity and the specialist methods remain, together with the six everyday additions and repaired named dependencies. The Agensi upload layout adapts executable scripts and reference delivery; historical source files and preservation records remain in the creator's original archive.
The repaired collection passed 409 local delivery and execution checks, including all 19 standalone class helpers, real Thirsty-lang calculations, TSCG-B round trips, exact task packets, delivery limits and negative cases. Independent checks also passed all 52 distinct audit-tool bindings and their source-index lookups. Six compiler-specific examples were not executed because their compiler dependencies were absent. These results do not establish every AI host's behavior, marketplace approval, live buyer retrieval, frontier-model superiority or revenue.
Start using it
Connect your AI to Agensi and ask: “Use my Agensi skill 07 — Catastrophe Shepherd · Complete Edition. My goal is [result]. Here is my actual material: [notes, records or code].” The pack provides an index; your AI loads the class entrypoint skills/conclave-catastrophe-shepherd/SKILL.md, then loads the supporting skills relevant to your task. Purchasing the pack does not automatically run all its skills at once.
Text and reasoning workflows run through your existing AI. Standalone scripts additionally require Python 3.10+ and a code execution tool. Thirsty-lang and TSCG-B execution requires thirsty-lang==0.9.0 installed separately in that environment; your AI should explain missing requirements before starting. Scripts are fetched unchanged through Agensi, executed temporarily and removed. Task result files can be delivered to you; the purchased skill is not installed as a downloadable local folder.
Image, video, voice and data services from Agensi's optional tools catalog are separate paid services. This pack does not require adding those services merely to use its text workflows or Python helpers. External actions, compiler tasks and account access still require the actual relevant host tools and your permission.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Testing & Debugging.
This is a skill/class package for an existing compatible AI host. It does not supply model weights, hosted compute, API credits, automatic account connections or permissions. The included helpers accept bounded structured inputs; they are not general document parsers or substitutes for semantic review. Their actual synthetic-fixture outputs are distinguished from authored specialist illustrations. Local preservation/tool checks do not establish correctness on arbitrary tasks, frontier-model superiority, customer results or revenue. Optional UTF analysis features may require additional dependencies. Marketplace ingestion of the final multi-skill ZIP remains subject to the receiving platform’s actual review.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 5 days ago
- Software
Node.js
- Optional
Python 3.10+thirsty-lang==0.9.0
Text workflows use your connected AI. Code needs Python 3.10+ and an execution tool. Thirsty-lang and TSCG-B need thirsty-lang==0.9.0 installed in that environment. No hosted model or API credits are included.
Permissions
Read Files, Write Files and Terminal/Shell support the optional local examples and authorized project work. Work stays within the user-approved scope. Browser, network, services and credentials are task-dependent and are not granted by this package. A draft, recommendation or prepared payload is not evidence of an external action.