GLM 5.3 Prompt Generator

The work GLM 5.3 was specifically post-trained for.

## Goal
[THE END STATE, e.g. all integration tests passing against the new schema]

## Environment
Working directory: [PATH]. Repo layout: [BRIEF LAYOUT]. Stack: [LANGUAGE, FRAMEWORK, VERSIONS].

## Available tools
- Shell: run any command in the working directory
- File system: read and write project files
- [OTHER TOOLS]

## Plan
Decompose roughly as: [STEP 1] → [STEP 2] → [STEP 3]. Adjust if the repo contradicts this.

## Constraints
- Do not modify [PROTECTED PATHS].
- Run the test suite after each change rather than at the end.
- Stop and report if [STOP CONDITION].

## Success Criteria
[e.g. `npm test` exits 0, no new lint errors, diff touches fewer than N files]

Full spec up front — GLM does badly on drip-fed requirements.

## Goal
Refactor [WHAT] across the repository so that [DESIRED END STATE].

## Context
[PASTE REPO TREE OR THE RELEVANT FILES]

## Constraints
- Language and version: [X]. Framework: [Y]. Formatter/linter: [Z].
- Preserve all public APIs unless listed below as intentional breaks: [BREAKS].
- No new dependencies.

## Output Format
1. Migration plan: files touched, in what order, and why.
2. Full revised contents for each changed file.
3. Anything deliberately left alone, with the reason.

## Success Criteria
Compiles clean, existing tests pass, no behaviour change outside the listed breaks.

GLM 5.3 tops CyberGym — give it a threat model and an evidence bar.

## Goal
Audit the code below for exploitable vulnerabilities.

## Threat model
Attacker profile: [e.g. unauthenticated internet user]. Assets at risk: [ASSETS]. Trust boundary: [BOUNDARY].

## Scope
In scope: [PATHS]. Out of scope: [PATHS]. Do not attempt to exploit anything outside the scope.

## What to hunt
Prioritise [CLASSES, e.g. injection, broken access control, deserialisation, secrets handling].

## Evidence standard
Every reported finding must include: the vulnerable line, a concrete exploitation path, the impact, and a fix. Mark anything speculative as "unconfirmed" rather than reporting it as a finding.

## Output Format
Findings ordered by severity, then a short summary of what you checked and found clean.

## Code
[PASTE CODE]

Paste the corpus first, instruction after — the 1M-token pattern.

[PASTE THE FULL CODEBASE, LOG SET OR DOCUMENT CORPUS]

---

## Goal
Based on everything above, [THE TASK, e.g. map every place user input reaches the database].

## Constraints
- Reference file paths and line numbers for every claim.
- Where the material is ambiguous, say so rather than guessing.

## Output Format
[e.g. a table of Source → Path → Sink, then a narrative summary]

## Success Criteria
Complete coverage — explicitly state if you believe any area of the input was not fully examined.

Verification criteria stated up front, which GLM follows reliably.

## Goal
Reproduce and fix this bug: [SYMPTOM].

## Context
Expected: [EXPECTED]. Actual: [ACTUAL]. Repro steps: [STEPS]. Environment: [ENV].
[PASTE RELEVANT CODE AND STACK TRACE]

## Plan
1. Write a failing test that captures the bug.
2. Find the root cause — not the nearest symptom.
3. Fix it.
4. Confirm the new test passes and no existing test breaks.

## Constraints
- Minimal diff. Do not refactor surrounding code.
- If the root cause is in a dependency, say so rather than patching around it.

## Success Criteria
New test fails before the fix and passes after; full suite green.

Ambiguity-killing structure, GLM's preferred input shape.

## Goal
Design [WHAT, e.g. a REST API for the billing service].

## Context
Consumers: [WHO]. Expected load: [LOAD]. Existing conventions: [CONVENTIONS]. Data model: [MODEL].

## Constraints
- Style: [REST/GraphQL/gRPC], [AUTH SCHEME], versioning via [SCHEME].
- Must support [HARD REQUIREMENTS].
- Must not [PROHIBITIONS].

## Output Format
1. Endpoint table: method, path, purpose.
2. Request/response schemas for each.
3. Error model.
4. Three design decisions you made and the alternative you rejected.

## Success Criteria
Every consumer use case listed above is satisfiable with the endpoints given.