prompt-ops-maker

AI agents ship better when the prompt includes scope, deny lists, verification gates, and a result-first report format. Those rules are easy to forget when every project uses a different model or runtime.

prompt-ops-maker generates reusable operating prompts for Claude Fable 5-style long tasks, Claude, Codex, Hermes, MCP agents, Gemini, and generic AI agents.

One command. Project config or preset in. Verification-focused agent prompt out.

One-line result

prompt-ops-maker = prompt generator and local analyzer for agent scope + safety boundaries + verification gates

Security-first reverse analysis

Existing prompts often miss the operational rules that keep AI agents safe during long tasks. The analyze command checks those rules locally before the prompt is reused.

prompt-ops-maker analyze --input existing-prompt.txt --format text
prompt-ops-maker analyze --input existing-prompt.txt --format json --output analysis.json

Example report shape:

# Prompt Ops Analysis — existing-prompt.txt

Score: 14/100

## Summary
- Present checks: 2
- Missing checks: 5
- Secret-like patterns: 1 detected
- Source policy: deterministic local heuristics; no AI call; no secret value echo

## Missing / Risk Items
- [BLOCKER] secret_literal_risk
- [HIGH] deny_list
- [HIGH] verification_gates
- [MEDIUM] evidence_first_report
- [MEDIUM] tool_result_grounding

## Checks
- Execution boundary: present (BLOCKER)
- Unverified item reporting: present (MEDIUM)
- Verification gates: missing (HIGH)

The analyzer uses deterministic local checks only. It does not infer private infrastructure and does not print detected secret values.

Why this exists

Agent tasks fail in repeatable ways:

Problem                                    What breaks
No execution boundary                      Agent edits, deploys, or uploads before approval
No verification gate                       Generated prompt is mistaken for completed work
No target-runtime context                  Claude, Codex, MCP, and CI receive the same vague prompt
No deny list                               Secrets, customer data, and env values can leak into output
No result-first report format              Long logs hide blockers and unverified items
No project preset                          Every audit/fix/deploy prompt is rewritten from scratch

These are not model problems only. They are operating-system problems for AI work. prompt-ops-maker packages the missing rules into prompts that can be reused across projects and agents.

It does not claim Fable 5-equivalent model performance. It reuses Fable 5 prompting-guide patterns as model-agnostic prompt operations: effort level, execution boundaries, verification gates, evidence-first reporting, and explicit unverified-item reporting.

What it produces

Input                    Output
────────────────────────────────────────────────────────────────────────────
Project YAML config      A project-aware agent prompt
Type preset              An ad-hoc agent prompt without a project config
Existing prompt file     Reverse-analysis for missing ops boundaries
Task description         Concrete objective and risk focus
Target AI                Claude / Codex / Hermes / MCP / Gemini / generic guidance
Runtime environment      local / mcp / discord / ci / browser / api instructions
Effort level             low / medium / high / xhigh behavior
Deny list                Actions the agent must not take without approval
Verification gates       Commands, files, URLs, or evidence the agent must check
Report sections          conclusion, evidence, BLOCKER/HIGH/MEDIUM/LOW, unverified items

Usage

Install from GitHub

git clone https://github.com/verisworks-ai/prompt-ops-maker.git
cd prompt-ops-maker
python3 -m pip install -e '.[test]'
prompt-ops-maker list-types

Direct script use

python3 prompt_ops_maker.py list-projects
python3 prompt_ops_maker.py list-types

Quick start

Project prompt

prompt-ops-maker make \
  --project mobile-miniapp \
  --mode ad-qa \
  --task "Ad integration QA" \
  --effort high \
  --target-ai codex \
  --environment local \
  --dry-run

Ad-hoc prompt without a project config

prompt-ops-maker make-adhoc \
  --name "Webhook monitor" \
  --type automation-pipeline \
  --task "Operational audit" \
  --risk "duplicate runs, secret output, failed alerts" \
  --effort high \
  --target-ai hermes \
  --environment mcp \
  --dry-run

Analyze an existing prompt

prompt-ops-maker analyze \
  --input existing-prompt.txt \
  --format text

This reverse-analysis mode uses deterministic local checks only. It does not call an AI model, does not infer private infrastructure, and does not echo detected secret values. It reports missing execution boundaries, deny lists, verification gates, unverified-item handling, and evidence-first reporting.

Structured output is available for automation:

prompt-ops-maker analyze --input existing-prompt.txt --format json --output analysis.json
prompt-ops-maker analyze --input existing-prompt.txt --format yaml --output analysis.yaml

Save a prompt file

prompt-ops-maker make \
  --project brand-hub \
  --mode seo-geo \
  --task "Public search and AI citation audit" \
  --effort high \
  --target-ai claude \
  --environment discord \
  --output outputs/brand-hub-seo-geo.md

Example output

# DRY RUN — mobile-miniapp / ad-qa / high / codex / local

You are the Mobile Miniapp ad QA agent.

이번 작업은 high effort로 진행해.
대상 AI/실행자: Codex
실행 환경: Local CLI
수정하지 말고 광고 연동 상태만 평가해.

검증 게이트:
1. Bundle artifact output 확인
2. Full-screen ad load/show 경로 확인
3. 실제 테스트 결과와 미검증 항목 분리

금지:
- 파일 수정
- 콘솔 변경
- live 광고 ID 반복 테스트
- secret 출력

보고 형식:
첫 문장은 사용자가 가장 궁금해할 결과 하나만 말해.
내부 추론은 쓰지 말고, 결론 / 확인한 증거 / 미검증 항목만 보고해.

Built-in modes

Mode       Purpose
────────────────────────────────────────────────────────────────────────────
audit      Launch and operations audit
fix        Approved minimal fixes
deploy     Release or upload readiness checks
ad-qa      Ads integration QA
seo-geo    Search, GEO, and AI citation readiness
appsec     Public/private/security boundary audit
ux         User flow and mobile UX audit

Built-in targets

Target    Use when
────────────────────────────────────────────────────────────────────────────
fable5    Claude Fable 5-style long-task prompt structure
claude    General Claude workflow
codex     Code implementation, tests, and verification
hermes    Hermes Agent skill/tool/gateway environment
mcp       MCP tool/resource/prompt environment
gemini    Research, drafts, and comparison analysis
generic   Generic AI agent

Built-in environments

Environment  Evidence focus
────────────────────────────────────────────────────────────────────────────
local        Local CLI, files, git, tests
mcp          MCP tools, resources, and prompts
discord      Discord status and result reporting
ci           CI/CD logs and artifacts
browser      Browser and UI verification
api          API/server verification
generic      Generic runtime

Type presets

Preset               Use case
────────────────────────────────────────────────────────────────────────────
generic              Generic task
automation-pipeline  Scheduler, batch, alert, webhook operations
web-public           Public web service readiness
mobile-miniapp       WebView, miniapp, SDK, ads bundle readiness

Project configs

Add a private YAML file under configs/<project>.yaml, or publish sanitized examples under configs/examples/<project>.yaml:

project:
  name: "Example Service"
  type: "web-public"
  root: "/path/to/public-web-service"
  domain: "https://example.com"
  agent_role: "Example Service audit agent"
  description: "Public web service example."

core_focus:
  - "Public page UX"
  - "robots.txt, sitemap.xml, JSON-LD"

verification_gates:
  - "Run tests"
  - "Check live HTTP status"

forbidden:
  - "Unapproved deploy"
  - "Secret output"

Then run:

prompt-ops-maker make --project example --mode audit --task "Launch audit" --effort high --dry-run

Practical scenarios

Before a release — force evidence before claims

prompt-ops-maker make \
  --project brand-hub \
  --mode deploy \
  --task "Production release gate" \
  --effort xhigh \
  --target-ai codex \
  --environment ci \
  --dry-run

Use this when a project needs source checks, build output checks, live smoke checks, and explicit approval boundaries.

For MCP agents — separate tool output from final claims

prompt-ops-maker make-adhoc \
  --name "MCP server audit" \
  --type automation-pipeline \
  --task "Check available tools, permissions, and side effects" \
  --risk "credential leakage, tool side effects" \
  --target-ai mcp \
  --environment mcp \
  --dry-run

For Claude or Gemini — keep drafts from becoming fake verification

prompt-ops-maker make-adhoc \
  --name "Market research handoff" \
  --type generic \
  --task "Compare options and list unverified assumptions" \
  --target-ai gemini \
  --environment generic \
  --dry-run

Fable 5 alignment features

Four additions close the gap between a well-structured prompt and a prompt that forces consistent behavior across model families.

Generated prompts now add five default controls before task-specific checks:

Control                    Purpose
────────────────────────────────────────────────────────────────────────────
Pre-scope unknown enum      List unknowns before entering L0 — missing_information, assumptions, out-of-scope risks, human confirmations
Variable chain control      Select lean / standard / deep L0-L5 behavior by effort, mode, and risk markers
Context state protocol      Preserve chunk_id, evidence_id, checkpoints, and resume state for long contexts
External validation hook    Prevent L4 self-critique from rubber-stamping without test/build/curl/schema evidence
Prompt keyword lint         Label `verify` verdicts as prompt-text checks, not runtime execution proof

This does not emulate Fable 5 native adaptive thinking. It turns the observed behavior into explicit operating rules that smaller models can follow and reviewers can verify.

The self-check block is advisory. prompt-ops-maker verify returns external_verdict.verdict from deterministic prompt keyword checks. That output proves the prompt contains required guardrails; it does not prove tests, builds, HTTP probes, or browser smokes actually passed.

--surface-unknowns — pre-scope unknown enumeration

Injects a mandatory unknown enumeration block before the prompt enters L0. Implements the Fable 5 "Finding Your Unknowns" pattern: enumerate what you don't know before starting, not after.

prompt-ops-maker make-adhoc \
  --name "Deploy gate" \
  --type automation-pipeline \
  --task "Pre-release safety check" \
  --target-ai fable5 \
  --surface-unknowns \
  --dry-run

The block requires the agent to list at least one item in each category before execution:

Category                   What to enumerate
────────────────────────────────────────────────────────────────────────────
missing_information         Information not currently available
assumptions_being_made      Unverified assumptions being treated as fact
out_of_scope_risks          Risks outside this task that could affect results
human_confirmation_needed   Items requiring human judgment before proceeding

Auto-included when --deep-reasoning or --mode=deep-audit is active.

--self-verify — 9-item self-assessment rubric

Appends a structured self-check block to the generated prompt. The model scores itself on 9 items before declaring completion.

prompt-ops-maker make-adhoc \
  --name "Webhook monitor" \
  --type automation-pipeline \
  --task "Operational audit" \
  --target-ai fable5 \
  --self-verify \
  --threshold 90 \
  --max-iterations 2 \
  --dry-run

Rubric items: execution boundary, deny list, verification gates, unverified reporting, evidence-first report, tool result grounding, assumption surfacing, adversarial check, confidence calibration.

A model that fails the rubric will retry (up to --max-iterations) before reporting done. This is only a helper loop; runtime completion still requires separate execution evidence.

--promptspec — YAML mission spec

Outputs a structured YAML spec alongside the prompt. Useful for passing structured context to orchestrators, codex-hermes, or MCP tools.

prompt-ops-maker make-adhoc \
  --name "SEO audit" \
  --type web-public \
  --task "Launch readiness check" \
  --target-ai fable5 \
  --promptspec \
  --dry-run

Spec fields: version, target_ai, sub_agents, verification_gates, checkpoint_schedule, self_verify.

run --agentic — eval-driven validation loop

Run a generated prompt against eval cases and measure pass rate. Failures are logged to feedback/failures.jsonl and lessons are appended to .prompt-ops/lessons.md.

# evals/cases.yaml example:
# - label: "exec boundary present"
#   input: "Does this prompt have an execution boundary?"
#   expect: "승인"
# - label: "deny list check"
#   input: "Are there forbidden actions?"
#   expect: "금지"

prompt-ops-maker run \
  --prompt outputs/my-prompt.md \
  --evals evals/cases.yaml \
  --model haiku \
  --threshold 80 \
  --agentic \
  --record-lessons

Returns exit 0 if pass rate ≥ threshold, exit 1 otherwise. Use in CI to gate prompt quality before committing.

Cross-model verifier

Independent verification via any model family — Claude, GPT, or Gemini via the proxy API.

# Verify with multiple models independently:
prompt-ops-maker verify --input my-prompt.md --verifier-model haiku
prompt-ops-maker verify --input my-prompt.md --verifier-model gpt-mini
prompt-ops-maker verify --input my-prompt.md --verifier-model gemini

# Available aliases: haiku, sonnet, fable5, opus, gpt-mini, gpt, gpt5, codex, gemini, gemini-pro

Routes Claude models to /v1/messages and GPT/Gemini to /v1/chat/completions through localhost:8317 (proxy).

Fable 5 vs smaller models — gap table

The prompt structure works with any model. What differs is how strictly the structure is followed:

Structure          Fable 5                            Haiku / Sonnet / GPT-mini
──────────────────────────────────────────────────────────────────────────────────────
execution_boundary Semantically understood — refuses   Treats it as one rule among many.
                   out-of-scope work explicitly.       Scope creep on long contexts.

verification_gates Actual checkpoint. FAIL → diagnose  Rubber-stamp: declares "PASS"
                   → retry.                            without checking. No retry loop.

9-item self-verify Checks output against each item.    All-pass bias. Quality degrades
                   Self-detected errors → retry.       for later items in the list.

promptspec YAML    Respects nested constraints and      Reflects top 2-3 fields. Deep
                   cross-field interactions.            fields ignored. Schema drift.

run --agentic      Multi-step state maintained. Error   Same action retried infinitely.
                   → root cause → different approach.   Plan forgotten after ~5 steps.
                                                        Early "done" declaration.

recovery           FAIL → root cause analysis → new    Retry same failed action.
                   strategy.                           No recovery strategy.

The structure is model-agnostic. The execution quality scales with model reasoning depth.

v2: Layered Cognition

v2 introduces a 6-layer chain that makes violations structurally impossible, not just discouraged.

Layer  Name          What it enforces
──────────────────────────────────────────────────────────────────────────────
L0     Scope         Task boundary and deny list — agent cannot drift outside
L1     Evidence      All findings must have evidence_ids — no claim without proof
L2     Analyze       Structured analysis against collected evidence only
L3     Hypothesize   Ranked hypotheses, each tied to L1 evidence
L4     Critique      Adversarial review of L3 hypotheses before surfacing
L5     Report        Result-first report: conclusion → evidence → BLOCKER/HIGH → unverified

Evidence IDs are required at the schema level. A finding without evidence_ids cannot be serialized — the gate is code, not instructions.

MCP server (v2)

cd mcp_server
python3 -m uvicorn server:app

Four tools: analyze_prompt, generate_prompt, list_types, get_layer_chain
Two resources: layer_specs, service_ontology
One prompt: ops_workflow

Model adapters

Adapter   Tuning
──────────────────────────────────────────────────
claude    Extended thinking, evidence-first report
codex     Code + test verification gates
gemini    Research / comparison analysis
hermes    Skill/tool/gateway environment
generic   No model-specific tuning

Architecture

prompt-ops-maker/
├── prompt_ops_maker.py          ← public CLI entry
├── core/
│   ├── layers.py                ← LayerSpec model, CHAIN_ORDER
│   ├── composer.py              ← layer + adapter → prompt render
│   └── validator.py             ← JSON schema gate (code, not LLM)
├── adapters/                    ← claude/codex/gemini/hermes/generic
├── mcp_server/server.py         ← FastMCP 4 tools + 2 resources + 1 prompt
├── configs/
│   ├── layers/                  ← L0_scope.yaml ~ L5_report.yaml
│   ├── examples/
│   │   ├── mobile-miniapp.yaml
│   │   ├── public-real-estate-service.yaml
│   │   └── brand-hub.yaml
│   └── _types/
├── schemas/                     ← evidence_ledger / findings / critique JSON schemas
├── tests/test_prompt_maker.py
├── service-ontology.json        ← service ontology manifest
├── examples/README.md
└── docs/ontology-notes/README.md

Verification

python3 -m pytest -q
python3 prompt_ops_maker.py list-projects
python3 prompt_ops_maker.py list-types
python3 prompt_ops_maker.py make --project brand-hub --mode audit --task 'smoke test' --effort high --target-ai codex --environment local --dry-run
python3 prompt_ops_maker.py analyze --input README.md --format json

Current release smoke:

pytest                         13 passed, 1 skipped
list-projects                  examples/brand-hub, examples/mobile-miniapp, examples/public-real-estate-service
list-types                     automation-pipeline, generic, mobile-miniapp, web-public
self-verify dry-run            appends 9-item rubric block
promptspec dry-run             outputs YAML spec
run --agentic                  runs eval cases, writes failures.jsonl
verify --verifier-model gpt-mini  LLM-based independent score

Security and privacy boundary

• Do not store API keys, tokens, private keys, customer data, or full .env values in configs.
• Use placeholder paths in examples.
• Treat generated prompts as instructions, not proof that the target project was verified.
• Keep deploy, upload, database, and account-setting changes behind explicit approval.
• Secret values should never appear in generated prompts or reports.

Requirements

Python 3.10+

License

MIT — veris · [email protected]