Manual
Repository manualdocs/USER_MANUAL.md

User Manual

Canonical documentation rendered from the repository docs/ folder.

manualSource file

User Manual

Purpose

This manual explains how to use the suite without reading all 257 prompts. It focuses on selecting the right outcome, setting honest authority, choosing an evidence lane, understanding generated artifacts, and knowing when work is not complete.

Mental model

The suite is not a bag of prompts. It is a routed system:

Reference
request
→ context and authority
→ prompt or scenario selection
→ evidence and brief
→ production or execution
→ verification
→ residuals and closure

Use one atomic prompt when one capability owns the result. Use a composite scenario when the result spans multiple distinct capabilities.

Starting a run

A useful directive contains:

Reference
RUN {PROMPT_ID|SCENARIO_ID|DEPARTMENT_PACK}
MODE {AUDIT_ONLY|PLAN_ONLY|DRAFT_ONLY|APPLY_SAFE|APPLY_APPROVED|VERIFY_ONLY}
ASSURANCE {FAST|STANDARD|HIGH_ASSURANCE}
ROOT {PROJECT_ROOT}
OUTCOME {OBSERVABLE_RESULT}
AUDIENCE {USERS_REVIEWERS_DECISION_MAKERS}
SCOPE {IN_SCOPE}
EXCLUDE {OUT_OF_SCOPE_AND_PROTECTED_SURFACES}
AUTHORITY {ALLOWED_WRITES_AND_EXTERNAL_ACTIONS}
EVIDENCE_LANE {FACTUAL|HYBRID|IMAGINATIVE}
EVIDENCE {AUTHORITATIVE_INPUTS}
MEDIA {SOURCE_AND_EXPORT_FORMATS}
SKILLS {AUTO|EXACT_IDS|NONE}
BUDGET {PROMPTS_CALLS_COST_TIME_PARALLELISM}
CONSTRAINTS {LEGAL_PRIVACY_BRAND_ACCESSIBILITY_TECHNICAL}
QUALITY {ACCEPTANCE_CRITERIA}

The most important fields are outcome, authority, evidence, and quality. Missing cosmetic preferences can usually be resolved later; missing authority or evidence may change the entire route.

Using the MD keyword

When root agent guidance is installed, begin a request with MD followed by an exact ID or ordinary task words:

Reference
MD C-95
MD research report
MD visual assets for a slide deck
MD audit fix verify

The agent should run deterministic keyword lookup first and inspect the selected target before execution. The keyword is a convenience layer; it does not weaken authorization, evidence, skill, loop, or verification rules.

See Root Agent Guidance and Keyword Routing Guide.

Choosing by outcome

Research and analysis

Use deep research, protocol, analytics, fact verification, or a department-specific investigative capability. Public or decision-critical claims normally require MD-82.

Writing and communication

Choose the artifact owner: article, script, academic paper, report, policy, pitch, communication, or social package. Add editorial review only when it materially improves the result.

Engineering

Use discovery or requirements before implementation when the target is unclear. Use true pairs for debugging, cleanup, refactoring, security remediation, migration, performance, reliability, delivery, and similar change surfaces.

Visual and interactive work

Choose presentation, dashboard, infographic, diagram, workshop board, web artifact, or frontend refinement. Verify source and exact export separately.

Organizational work

Use department packs to discover capabilities, then compile the smallest graph. Do not execute an entire department pack.

Modes

AUDIT_ONLY

Use when you want findings, evidence, or review without changes.

PLAN_ONLY

Use for plans, specifications, acceptance criteria, or approval preparation.

DRAFT_ONLY

Use for local, unapproved artifacts. A draft is not accepted, submitted, published, sent, or deployed.

APPLY_SAFE

Use for reversible local changes within explicit authority.

APPLY_APPROVED

Use only for the exact approved consequential action. A valid receipt, recovery path, and verification plan are required.

VERIFY_ONLY

Use for independent checks. The verifier should not alter the reviewed subject.

Assurance profiles

Choose the minimum assurance that matches the risk, not the shortest route you prefer.

  • FAST: low-risk local work.
  • STANDARD: current evidence, typed artifacts, review, verification, residuals.
  • HIGH_ASSURANCE: formal evidence, counterevidence, approvals, dry run, recovery, independent verification, lineage.

Reading marker-rich output

A good run exposes its reasoning products without revealing private chain-of-thought:

Reference
@EVIDENCE:market-01 authoritative market report, collected 2026-07-15
#FINDING:market-01 demand is concentrated in two segments
?UNKNOWN:market-02 regional price sensitivity is not measured
+ACTION:market-01 run a limited pricing experiment
=VERIFY:market-01 finding supported by two independent sources

Markers let you trace decisions and identify where uncertainty remains.

Understanding completion

Do not judge completion by length or confidence. Check the prompt's task-specific <completion_criteria>.

A run is not complete when:

  • a required artifact is missing;
  • a verification record is absent;
  • a failed action is hidden;
  • an unknown is treated as fact;
  • a residual has no owner;
  • an executive discovered new work and executed it without approval;
  • the source is correct but the export is broken;
  • publication or deployment was assumed rather than authorized.

Skills

SKILLS AUTO permits the router to consider registered adapters. It does not authorize installation or external effects. Unresolved locks and unreviewed permissions block automatic high-assurance use.

Use SKILLS NONE when native prompt execution is sufficient or supply-chain risk is unacceptable.

Common examples

Research-backed report

Reference
RUN C-95
MODE DRAFT_ONLY
ASSURANCE STANDARD
OUTCOME Produce a decision-ready report with reconciled exhibits and exact PDF QA
AUTHORITY local artifacts only
EVIDENCE_LANE FACTUAL

Safe cleanup

Reference
RUN C-04
MODE PLAN_ONLY
ASSURANCE HIGH_ASSURANCE
OUTCOME Identify junk, dead code, generated artifacts, and duplicate sources of truth without deleting anything
AUTHORITY read-only

The approved executive run can follow only after evidence proves non-use.

Brand and deck

Reference
RUN C-37 + C-41
MODE DRAFT_ONLY
ASSURANCE STANDARD
OUTCOME Create a coherent brand system and browser-native presentation
EVIDENCE_LANE HYBRID

Imaginative work

Reference
RUN C-33
MODE DRAFT_ONLY
ASSURANCE FAST
OUTCOME Draft a short story with a specified viewpoint and scene turn
EVIDENCE_LANE IMAGINATIVE

Do not force factual research into fiction unless the work intentionally makes factual claims.

Failure and residuals

A good run may end in:

  • failed because verification did not pass;
  • rolled_back because execution caused damage or uncertainty;
  • residual_open because approved work remains;
  • !STOP because evidence, authority, or recovery is insufficient.

These are honest outcomes, not system failures to conceal.

Where to learn more

Reviewing a paired plan before execution

When a planning prompt has an execution twin, the system will show the completed plan before doing execution work. Review the plan and provide changes, additions, removals, or refinements. The planner applies accepted feedback and presents the revised plan again. After you approve that exact version, the planner asks whether to invoke its named execution twin. Approval of the plan and permission to execute are separate choices.