P
PromptOSEnterprise AI Prompting
Interactive enterprise demo

Turn prompting into a delivery capability.

Build, score, test and govern enterprise-grade prompts in one dynamic workspace-designed for AI Project Managers, PMO leaders and delivery teams.

Prompt readiness
86
Status
Production candidate
8/8Anatomy
LowRisk
4.6×ROI
92%Coverage

Guided learning tutorial

A 50-slide, step-by-step training experience prepared for ElevenLabs cloned-voice narration, practice prompts, cost controls and role-specific examples.

01 / 50
Voice-guided learning mode Play your ElevenLabs cloned voice for the current slide, or use Auto-run to continue slide by slide.
Narration will appear here.
01 Welcome to PromptOS

You are in the right place. Today you build confidence.

This tutorial will give you the tools, approaches and practical judgment to create and deliver accurate, high quality tasks and solutions with AI.

Clarity

Turn vague requests into precise, reusable instructions.

Accuracy

Ground outputs in evidence, verify claims and reduce hallucination risk.

Speed

Reduce retries, latency and wasted effort through better prompt design.

Control

Manage token spend, governance, privacy and delivery quality with confidence.

By the end, users should be able to prompt with purpose, review with discipline and deliver outputs they can trust.
02 The mindset shift

Do not “ask AI a question.” Design a capable collaborator.

Passive user

Asks, hopes, accepts the first draft and regenerates when disappointed.

AI Project Manager

Defines the decision, supplies context, sets boundaries, tests quality and owns the outcome.

The model produces language. The professional remains accountable for accuracy, business alignment, risk and delivery quality.
03 Anatomy of a strong prompt

Eight components create repeatable quality.

1ObjectiveThe exact deliverable.
2RoleThe expert perspective.
3ContextBackground and assumptions.
4ConstraintsBoundaries and exclusions.
5SourceEvidence and grounding.
6FormatStructure and length.
7EvaluationAcceptance criteria.
8ToneAudience and voice.
04 Why weak prompts fail

Most AI quality failures are prompt-design failures.

Vague task

“Summarize this” leaves purpose, audience and decision use undefined.

No grounding

The model fills evidence gaps with plausible but unsupported language.

No review loop

The first draft is treated as final instead of being critiqued and validated.

No format

The output may be accurate but unusable for the business workflow.

No constraints

Scope expands, sensitive data leaks and critical boundaries disappear.

No audience

Language, depth and emphasis miss the stakeholder need.

05 Technique selection

Match the technique to the complexity and risk.

Zero-shot

Simple, clear and low-ambiguity tasks.

Few-shot

Outputs that must match approved patterns.

Decomposition

Complex analysis with separable steps.

Prompt chaining

Draft, critique, revise and validate workflows.

Use structured prompting for repeatability. Add critique-and-refine when the output is high stakes.
06 World-class example

From “update the risk register” to a governance-ready instruction.

Weak

Update the risk register based on the project update.

Strong

Act as a senior programme risk manager. Using only the approved status report and RAID log, identify new or changed risks, score probability and impact on a 1-5 scale, state evidence, assign an owner role, propose a dated mitigation and flag any risk requiring steering committee action.

07 Quality control

A prompt is not finished until the output is evaluated.

Factual accuracy

Are claims supportable from the approved source?

Scope coverage

Did every requested element appear?

Consistency

Do numbers, dates and conclusions agree?

Evidence

Can the reader distinguish facts, assumptions and inference?

Relevance

Is the output immediately useful to the target audience?

Decision readiness

Are owners, actions, dates and trade-offs explicit?

08 Data and privacy

Never paste data simply because the tool accepts it.

Personal data

Anonymize or aggregate identifiers.

Client-confidential data

Use approved enterprise environments only.

Credentials

Never enter passwords, keys or access tokens.

Privileged material

Use summaries cleared by the relevant control function.

09 Governance operating model

Governance turns an individual habit into an organizational advantage.

Standards

Shared templates, owners and approved patterns.

Validation

Mandatory checks for facts, scope, risk and audience.

Audit trail

Prompt, model, source, version, date and decision record.

Capability

Peer review, calibration and prompt champions.

10 Maturity journey

Scale from ad hoc use to measured capability.

1 - Ad hoc

Individual experimentation.

2 - Repeatable

Personal templates.

3 - Defined

Shared standards.

4-5 - Managed & optimized

Governed, measured and continuously improved.

11 Business value

Measure prompting as a delivery investment.

Time efficiency

Minutes saved per recurring task.

Output quality

Acceptance without material revision.

Task reach

Work previously too costly or slow to perform.

Value created

Hours saved × loaded labor cost × adoption.

12 Business-user playbook

Business users win with context, decision clarity and restraint.

Name the decision

Ask for the business outcome-not “ideas” or “a summary.”

Ground the answer

Identify the approved source and prohibit invented facts.

Control the output

Set audience, structure, length and materiality thresholds.

Budget the response

Request only the sections and detail the stakeholder needs.

13 Developer playbook

Technical prompts need precision at every interface.

Minimal reproducible context

Include interfaces, errors and relevant code-not the entire repository.

Demand structured output

Request a patch, schema, test matrix or exact file-level change.

Define acceptance tests

State security, performance, compatibility and test requirements.

Constrain invention

Do not permit undocumented dependencies, APIs or configuration.

14 Token economics

Prompt cost is driven by input, output, retries and scale.

Input tokens

Instructions, examples, source text, history, schemas and tool results.

Output tokens

Often priced higher-cap length and request concise structures.

Failed iterations

Ambiguous prompts create repeated calls and hidden rework cost.

Model routing

Use the least expensive model that reliably meets the quality threshold.

Cost per call = input tokens × input rate + cached input × cache rate + output tokens × output rate. Human rework usually matters more than raw token cost.
15 Prompt library economics

A governed library converts good prompts into reusable intellectual property.

Fewer retries

Approved patterns reduce failed generations and repeated clarification.

Lean context

Reusable modules remove duplicated instructions and irrelevant background.

Cache-ready blocks

Stable system instructions and schemas can use provider caching where supported.

Measured value

Track quality, tokens, cost, acceptance rate and time saved by prompt version.

16 Call to action

Move from experimentation to enterprise discipline.

This week

Adopt the eight-part master template and review one live prompt.

Within 30 days

Build a library for the five highest-frequency use cases.

Within 60 days

Launch peer reviews and quality calibration.

Within 90 days

Measure rework, time saved, quality and adoption.

At scale

Govern prompting as a reusable organizational asset.

Leadership outcome

Faster decisions, stronger controls and more reliable AI value.

17 Learning contract

Every slide teaches one usable workplace behavior.

Watch

Understand the concept and why it matters.

Copy

Lift the structure directly into a real enterprise prompt.

Practice

Run the prompt against a realistic task.

Govern

Save approved versions in the prompt library.

The training goal is not to make users clever at prompting. It is to make outputs predictable, reviewable, auditable and cost efficient.
18 The universal workflow

World-class prompting follows a repeatable six-step loop.

1. Define the job

What decision or deliverable will this support?

2. Supply only needed context

Relevant facts, approved source material and boundaries.

3. Specify output

Format, length, audience, tone and acceptance criteria.

4. Generate

Ask once with enough precision to avoid blind retries.

5. Critique

Check evidence, scope, consistency, risk and usefulness.

6. Save what works

Add reusable patterns to the governed prompt library.

19 Business example: status report

Convert status data into an executive decision brief.

Bad request

“Summarize the project.” This invites narrative, not governance.

Best-in-class request

“Create a 220-word steering committee brief using only the status report and RAID log. Include overall status, material progress, decisions required, top three risks, owners and next 30-day actions. Flag missing evidence.”

Teaching point: business users should prompt for the decision the stakeholder must make, not the document they happen to have.
20 Business example: decision paper

Force trade-offs into the open before leaders approve.

Role

Strategy advisor preparing a CIO decision brief.

Method

Compare options by cost, timeline, risk, reversibility and control impact.

Output

Recommendation, rationale, sensitivity analysis and approval conditions.

Constraint

Do not quantify benefits unless the source provides figures.

Quality check

State which assumption would change the recommendation.

Cost discipline

Use a compact template rather than pasting the entire project archive.

21 Business example: stakeholder email

Use AI to improve clarity, tone and accountability.

State the issue

One clear sentence. No jargon shield.

Explain impact

What changes for the stakeholder and why it matters.

Own next steps

Actions taken, owner role and target date.

Ask for a decision

Make the sponsor’s requested action explicit.

Prompt pattern: “Draft, then critique for defensiveness, vagueness, blame language and missing decision asks.”
22 Developer example: secure code review

Technical prompting rewards minimal but complete evidence.

Include

Changed files, relevant function, error trace, threat model and constraints.

Exclude

Unrelated repository files, stale logs and speculative requirements.

Ask for

Findings ranked by severity with exploit scenario and exact remediation.

Prevent

No undocumented dependencies, no imagined APIs, no broad rewrites.

Validate

Request test cases and residual risk after the proposed fix.

Budget

Cap output to the top issues that materially affect security.

23 Developer example: API documentation

Documentation prompts must protect against invented interfaces.

Source

OpenAPI spec, auth notes and operational limits only.

Structure

Purpose, auth, parameters, examples, errors and troubleshooting.

Validation

Field names and types must match the source exactly.

Gap handling

Conflicts and missing limits go into a discrepancy list.

The highest-value developer prompts often ask the model to find contradictions before writing the final answer.
24 Developer example: test generation

Test prompts should produce coverage, not volume.

Traceability

Every test maps to a requirement or acceptance criterion.

Coverage classes

Positive, negative, boundary, role-permission and audit trail.

Determinism

Expected result must be observable and testable.

Ambiguity handling

Untestable rules become open clarifications.

Cost control

Generate by feature or requirement group, not the entire backlog at once.

Quality check

Ask for duplicate detection and missing coverage after draft generation.

25 Token waste patterns

Most token waste is caused by undisciplined context.

Pasting everything

Large source dumps with no relevance filter.

Long chat history

Old assumptions and stale instructions keep consuming context.

No output budget

The model writes more than the audience needs.

Retry loops

Cheap prompts become expensive after five failed attempts.

Train users to measure cost per accepted output, not cost per individual call.
26 Cost-efficient prompt design

Quality and efficiency improve together when the prompt is engineered.

Compress context

Replace raw history with a current-state summary and source links.

Set output limits

Use word, row, bullet or JSON field caps.

Use schemas

Structured output reduces wandering prose and revision cycles.

Route by risk

Use cheaper models for drafts and stronger models for high-stakes validation.

Chain selectively

Draft → critique → refine only when risk justifies extra calls.

Cache stable blocks

Keep reusable instructions, policies and schemas stable where provider caching is available.

27 Prompt library operating model

A prompt library is a cost, quality and control system.

Template owner

One accountable person approves changes.

Versioning

Track purpose, model, token budget and acceptance rate.

Usage metrics

Monitor runs, retries, cost, saved time and rejected outputs.

Retirement rules

Remove stale templates and prompts that drift from policy.

Library economics are simple: fewer failed attempts, less duplicated instruction text, safer model routing and reusable cache-ready context.
28 Certification practice

Before users graduate, make them prove the behavior.

Exercise 1

Rewrite a vague business prompt into the eight-part structure.

Exercise 2

Reduce a bloated prompt by 40% while preserving quality criteria.

Exercise 3

Choose the right technique and model route for a high-risk task.

Exercise 4

Score the output against accuracy, scope, evidence and relevance.

Exercise 5

Save the final approved template to the prompt library.

Graduation standard

The prompt is reusable, traceable, cost-aware and accepted with minimal rework.

29 Reliability doctrine

World-class AI delivery balances truth, speed and control.

Truth

Every material claim must trace to an approved source, a calculation or an explicit assumption.

Speed

Latency is designed, not wished away: route, cache, stream, parallelize and cap the work.

Control

Use policies, validation gates, ownership and escalation paths for high-risk outputs.

Trade-off

Never optimize speed by removing the controls required for the business risk.

The leadership question is not “Which model is best?” It is “What is the right reliability pattern for this task?”
30 Hallucination root causes

Hallucinations are usually caused by missing evidence and weak instructions.

Evidence gap

The model is asked for facts that are not in the source packet.

Ambiguous task

The prompt does not define what counts as a valid answer.

Over-broad context

Important facts are buried inside irrelevant material.

No refusal rule

The model is not told to say “not enough evidence.”

No citations

Unsupported claims look as credible as sourced claims.

No validation loop

The first draft is accepted without factual challenge.

31 Hallucination control stack

Control hallucination with a layered operating model.

1. Source boundary

Use only the attached source packet or approved retrieval results.

2. Evidence rule

Every decision claim needs a source reference or confidence label.

3. Refusal rule

If evidence is missing, the model must ask for it or list the gap.

4. Structured output

Separate facts, assumptions, inferences, risks and recommendations.

5. Independent check

Run a second validation pass focused only on unsupported claims.

6. Human sign-off

Named owner accepts outputs used in regulated or high-impact decisions.

32 Source packet design

The best anti-hallucination technique is better evidence packaging.

Source inventory

Name the files, dates, owners and allowed use of each source.

Relevance filter

Include only passages that can change the answer.

Conflict handling

Tell the model how to handle contradictory dates, figures or owners.

Freshness check

Mark stale material and require an “unknown” answer when currency matters.

Prompt pattern: “Use only Source A-D. Cite source IDs. If the sources conflict, show the conflict instead of resolving it silently.”
33 Gold-standard hallucination prompt

Make unsupported claims impossible to hide.

Prompt clause

“For each claim, mark it as Source-supported, Calculation, Assumption or Inference. If a claim is not supported, place it in an Evidence Gap table instead of presenting it as fact.”

Output columns

Claim, Evidence ID, Confidence, Business impact, Required validation, Owner, Decision consequence.

This changes the model’s job from sounding convincing to maintaining an evidence ledger.
34 Verification workflow

High-risk AI outputs need a draft-check-approve chain.

Draft

Generate the output with strict source and format controls.

Extract claims

List every factual, numerical, legal, financial or technical claim.

Verify evidence

Check each claim against the approved source packet.

Red-team

Ask what could be wrong, missing, stale or misleading.

Revise

Remove unsupported claims and downgrade uncertain language.

Approve

Human owner signs off according to the risk tier.

35 Hallucination metrics

Manage trust with measurable factuality controls.

Unsupported claims

Count unsupported factual claims per accepted output.

Citation precision

Measure whether cited sources actually support the sentence.

Evidence-gap handling

Track whether the model says “unknown” when it should.

Correction rate

Measure human edits caused by factual errors.

Escalation accuracy

Track whether high-risk uncertainty reaches the right owner.

Golden tests

Run known-answer cases before changing model, prompt or retrieval.

36 Latency root causes

Latency is the sum of model, context, output and orchestration.

Model size

Larger reasoning models often improve quality but increase response time.

Input context

Long prompts and large retrieved passages increase processing time.

Output length

Every extra token must be generated; verbosity is latency.

Tool calls

Search, retrieval, database, API and agent calls add network and compute delay.

Serial chains

Draft → critique → revise improves quality but adds sequential waiting.

Retries

Ambiguous prompts create hidden latency through rework and reruns.

37 Latency design patterns

Make AI feel fast by designing the system path.

Route by task

Use small, fast models for classification, extraction and formatting.

Cap output

Set word, row, bullet or JSON-field budgets.

Retrieve less

Limit top-k, chunk size and irrelevant source expansion.

Parallelize tools

Run independent searches, checks and API calls at the same time.

Cache stable context

Reuse system prompts, policy blocks, schemas and source summaries.

Stream early

Show first useful content quickly while longer work continues.

38 Model routing ladder

Use the smallest model that can meet the risk and quality bar.

Tier 1: Fast path

Classify, extract, rewrite, format and summarize low-risk content.

Tier 2: Standard path

Business analysis with moderate ambiguity and approved context.

Tier 3: Expert path

High-stakes reasoning, trade-offs, architecture and regulatory interpretation.

Tier 4: Human path

Legal, client-impacting, safety-critical or unsupported factual decisions.

Model routing is both a cost strategy and a latency strategy. Escalate only when the work truly requires it.
39 Operational latency SLOs

Production AI needs service-level thinking.

Time to first token

How quickly the user sees useful response progress.

p50 / p95 latency

Median and worst-normal response times by task and model route.

Tokens per second

Generation speed after the model starts responding.

Retry rate

How often users rerun because the output failed first time.

Fallback rate

How often the system downgrades or escalates to another route.

User-perceived wait

Whether progressive disclosure, streaming and status updates reduce frustration.

40 Reliability graduation standard

The user is trained when they can reduce hallucination and latency together.

Trust

Uses source boundaries, citations, confidence labels and evidence-gap tables.

Speed

Chooses model routes, token budgets, caching and parallel execution deliberately.

Judgment

Knows when to slow down for validation and when to use a fast path.

Operations

Tracks factual errors, latency, retries, cost and acceptance rate by use case.

Library

Saves proven reliability clauses as reusable template blocks.

Governance

Escalates uncertainty before AI output becomes business action.

41 Token spend doctrine

Token spend is managed before build, not after the invoice.

Planning artifact

Create a Token Strategy Charter for every material AI workload.

Named owner

Assign business, platform, finance and risk accountability before launch.

Three scenarios

Project expected, stretch and runaway consumption before engineering begins.

Kill criteria

Define automatic pause thresholds for runaway spend, latency and quality failure.

Best-practice frame: manage tokens as a controllable operating expense attached to business value, not as an unavoidable technical bill.
42 Context balance sheet

Every token in the context window must earn its place.

Asset tokens

Instructions, evidence, examples or schemas that directly improve the answer.

Liability tokens

Resident content that might help but is not clearly load-bearing.

Expense tokens

One-time content that should be consumed, summarized and removed.

Passenger tokens

Irrelevant history, stale sources and broad dumps that increase cost and distract the model.

Eviction rule

Specify when old context, tool results and chat history leave the working set.

Source filter

Filter once upstream rather than asking the model to filter every call.

43 Five spend levers

World-class teams manage spend through five controllable levers.

1. Model selection

Use the lowest-cost model that reliably meets the quality and risk bar.

2. Workflow architecture

Reduce unnecessary agents, serial chains, tool loops and orchestration overhead.

3. Data pre-work

Clean, rank, chunk and filter before retrieval reaches the model.

4. Prompt literacy

Train users to specify context, format, evidence and output budget.

5. Human judgment

Spend review effort where it prevents expensive downstream errors.

Portfolio governance

Review these levers quarterly for every production workload.

44 Orchestration multiplier

The largest hidden cost is usually calls per user action.

Definition

Orchestration Multiplier = total tokens consumed for one user-visible action divided by the tokens required by a single well-prompted call.

Target behavior

Prefer a single tool-using call when possible. Authorize multi-agent decomposition only when the sub-problems are parallel, substantial and require synthesis.

1.0×

Single disciplined call.

2-3×

Supervisor-worker flow.

6-10×

Multi-agent decomposition.

Runaway

Unbounded recursion, fan-out and retries.

45 Cache and library discipline

Caching only works when reusable content is stable.

Cache candidates

System prompts, policy blocks, schemas, style guides and large stable instructions.

Cache killers

Timestamps, user IDs, request IDs and variable text inside the cached prefix.

Break-even logic

Cache only when repeated use exceeds the write premium and time-to-live limits.

Library benefit

Approved templates create stable prefixes, fewer retries and easier model routing.

Monitoring

Track cache reads, writes, hit rate, cost avoided and invalidation causes.

Rule

Do not claim caching savings unless the hit rate proves it.

46 Model routing economics

A cascade saves money only when the math proves it.

Fast tier

Classification, extraction, reformatting, short summaries and routing.

Workhorse tier

Moderate ambiguity, business analysis, synthesis and reusable enterprise workflows.

Expert tier

Complex reasoning, regulated analysis, architecture trade-offs and high-impact work.

Escalation rate

If too many requests escalate, the cascade may cost as much as flat routing.

Misroute cost

Cheap first passes are expensive if they create rework or missed risk.

Quarterly test

Compare cost, acceptance, latency and risk against a simpler routing strategy.

47 Spend telemetry

Manage what matters: cost per accepted output.

Token ledger

Input, output, reasoning, cache read/write and tool-result tokens by workload.

Usage ledger

Calls, users, retry rate, failed attempts and peak concurrency.

Quality ledger

Acceptance rate, correction time, unsupported claims and escalation events.

Architecture ledger

Orchestration multiplier, agent depth, fan-out and tool calls per action.

Financial ledger

Expected, stretch, runaway and actual spend against budget.

Decision ledger

Ratify, revise, retire or redesign every material workload.

48 Anti-patterns that burn budget

Most token overruns have a recognizable fingerprint.

Recursive sub-agent

Token spike without traffic growth; fix depth, fan-out and circuit breakers.

Cached timestamp

Cache writes equal call count; move variable fields out of the cached prefix.

Phantom RAG

Retrieval adds tokens without improving accuracy; A/B test without retrieval.

Eternal conversation

Input tokens rise with every turn; summarize and evict stale history.

Opus habit

Premium model used without eval evidence; downgrade where quality is equivalent.

Unbounded system prompt

Prompt grows without review; convert stable blocks into governed reusable assets.

49 Token spend operating model

Token control needs a role, a ledger and a cadence.

Token Architect

Owns the economics discipline, exception review and portfolio dashboard.

MBOM

Model Bill of Materials records model, prompt, sources, workflow, owner and review topology.

Daily / weekly

Monitor thresholds, exceptions, cache hit rates, routing and latency.

Monthly / quarterly

Review budget variance, realized value, model tier and retire-or-revise decisions.

Executive KPI: alpha-net per dollar of inference spend-not blind cost reduction.
50 Monday diagnostic

The practical test: can you explain where every dollar goes?

1. Top workloads

List the top 10 AI workloads by monthly token spend.

2. Owners

Name the business owner, platform owner, finance partner and risk sign-off.

3. Unit economics

Calculate cost per accepted output and human rework avoided.

4. Waste fingerprint

Check OM, cache hits, retries, stale context, RAG quality and model tier.

5. Kill switch

Confirm runaway thresholds and automatic pause rules exist.

6. Action

Ratify, revise, retire or redesign each workload within the quarter.

Enterprise value cockpit

A live view of the adoption-to-value gap, prompting quality and delivery impact.

AI adoption
76%
Near universal
Bring-your-own-AI
78%
Governance exposure
Fully scaled
<30%
Value-capture gap
Prompt-led savings
2.0h
Per task assumption

Prompt Lab

Compose a structured prompt, watch quality update in real time and generate a reusable enterprise-ready output.

0
Start building your prompt Complete the eight anatomy fields to improve quality, auditability and reuse.
Prompt tokens0
Estimated call cost$0.0000
Your generated prompt will appear here.

Weak prompt

“Summarize the project status.”

No roleNo audienceNo constraintsNo format

Likely outcome

Generic output, variable depth, high rework and weak governance traceability.

Strong prompt

You are a senior project manager. Produce a three-paragraph executive summary for the steering committee covering progress, decisions, top risks and next actions. Use only the supplied status data, stay under 200 words and finish with named owners and dates.

RoleObjectiveContextConstraintsFormat

Likely outcome

Focused, decision-ready output with predictable quality and lower revision effort.

Select a technique

Choose a card to see where it works best and what to watch for.

Governance readiness
0 of 4 controls enabled

Business and developer playbooks

Two roles, two risk profiles and two different ways to produce quality without wasting tokens.

Input tokens0
Output budget0
Cost per run$0.0000
Monthly model cost$0.00

Token and cost command center

Estimate prompt tokens, output budget, per-call cost and scaled spend. Then optimize for quality per dollar-not merely the shortest prompt.

0%

Token counts are estimates using approximately four characters per token for English prose; code, tables, non-English text, system messages, tools and provider tokenizers vary. Use the provider tokenizer or API usage record for billing-grade counts.

Estimated monthly model spend
$0.00
Claude Sonnet 5
Input tokens / call0
Output cap / call500
Cost / call$0.0000
Annual spend$0.00
Input share0%
Retry cost / month$0.00

Cost of representative prompts on the selected model

PromptAudienceInput tokensOutput budgetEstimated cost / run100 runs
1. Remove context that cannot change the answer

Do not paste entire decks, logs or threads when a focused extract, retrieval query or source reference will do.

2. Set an explicit output budget

Ask for the exact format, number of rows and maximum length. Output tokens are often the more expensive side of the call.

3. Reduce retries-not necessary precision

A 400-token prompt that works once is usually cheaper than a 100-token prompt repeated four times.

4. Route by task complexity

Use efficient models for classification, extraction and formatting; reserve premium reasoning for difficult or high-risk work.

5. Cache stable context where supported

System instructions, policies, schemas and large repeated reference blocks can be cache candidates. A library makes these blocks consistent and cache-ready.

6. Measure quality-adjusted cost

Track cost per accepted output, not cost per call. Include human correction time, failed attempts and downstream risk.

Pricing presets are provided for demonstration and are dated July 11, 2026. Provider pricing, discounts, regions, long-context tiers, reasoning tokens, tools and batch rates can change; the editable rate fields remain the source of truth for this calculator.

Token spend manager

A practical governance cockpit for estimating, reducing and controlling AI inference spend by workload.

Spend doctrine
Plan
Charter before build
Operating KPI
C/AO
Cost per accepted output
Waste signal
OM
Calls per user action
Control rhythm
QBR
Ratify / revise / retire

Workload spend simulation

Monthly users100
Actions / user / month20
Input tokens / call3,000
Output tokens / call700
Average attempts / accepted output1.4×
Orchestration multiplier2.5×
Cache hit share30%
Upstream context reduction20%
Projected monthly inference spend
$0

This simulation includes model route, context size, output budget, retries, orchestration multiplier, caching and upstream filtering.

Annual run rate$0
Cost / accepted output$0
Monthly accepted outputs0
Before controls$0
Controls avoided$0
Primary leverReduce context
1. Write a Token Strategy Charter

Define purpose, business owner, expected/stretch/runaway scenarios, review topology, kill criteria and reporting cadence before build.

2. Maintain a Model Bill of Materials

Record every workload’s model tier, prompts, source set, workflow pattern, review rules, owner and last review date.

3. Audit orchestration multiplier

Measure total tokens per user-visible action. Replace unnecessary agents and serial chains with a single tool-using call where possible.

4. Apply eviction discipline

Set turn caps, summarization triggers, tool-result truncation and relevance scoring so stale history stops consuming context.

5. Filter upstream

Remove stale, duplicate, low-authority and irrelevant material before indexing or retrieval. Do not pay the model to filter every time.

6. Track cost per accepted output

Include retries, human correction, hallucination remediation and latency impact-not just raw request cost.

Token spend control checklist

Governed prompt library

Capture approved prompts once, reuse them at scale and measure quality, tokens, cost and acceptance by version.

Library prompts0
Approved0%
Planned monthly runs0
Estimated monthly spend$0

Add or save a prompt

Custom entries are stored locally in this browser using localStorage. No content is transmitted by this standalone demo.

Library savings simulator

Cost avoided through reuse

Team prompts / month1,000
Ad-hoc attempts / task2.0×
Library attempts / task1.1×
Reusable-context cache hit60%
$0
Estimated monthly model cost avoided. Human rework savings are additional.
Ad hoc$0
Library$0
01Capture

Start with high-frequency, high-rework tasks and a named business owner.

02Review

Score structure, data controls, evidence, token budget and expected output.

03Approve

Record version, approved model, owner, sensitivity, expiry and test results.

04Measure

Track usage, tokens, model cost, retries, acceptance and time saved.

05Improve or retire

Promote winning versions and remove stale, duplicate or low-value prompts.

World-class prompt gallery

Detailed enterprise examples across programme delivery, governance, communications, technology and control functions.

Prompt review studio

Paste any prompt, diagnose the design gaps, receive a structured score and generate an enhanced enterprise version.

The enhanced prompt will appear here.
0
Not yet assessed

Eight design dimensions scored using practical enterprise heuristics.

Prompt risk heatmap

See how common design failures drive rework, hallucination, compliance exposure and stakeholder distrust.

Failure modeOutput qualityRework costHallucinationComplianceTrust
Vague objectiveHighHighMediumLowMedium
Missing context & roleHighMediumMediumLowMedium
No output formatMediumHighLowMediumLow
No constraintsMediumHighMediumMediumLow
No source groundingHighHighCriticalHighHigh
No review or validationHighMediumHighHighHigh

Output evaluation lab

Build a repeatable review habit before an AI output is shared, submitted or used in a decision.

Quality scorecard

Hallucination warning signs

Trust recommendation
0
Do not use yet

Complete the scorecard and flag any warning signs. High-risk content should be regenerated or independently verified.

Data, privacy and tool routing

Protect sensitive information and route each task to the right enterprise environment.

Personal identifiable data

Do not paste names, account details, addresses or protected identifiers into unapproved tools.

Client-confidential information

Use approved, contractually governed enterprise environments and minimum necessary data.

Credentials and secrets

Never provide passwords, API keys, access tokens or security answers.

Legally privileged material

Do not paste legal advice or privileged documents without explicit authorization.

Anonymize before prompting

Replace names and identifiers with roles, categories or synthetic values.

Use the approved environment

Match data classification and task sensitivity to the enterprise tool.

Split the workflow

Use AI for structure and analysis, then add sensitive details outside the model.

Preserve source lineage

Record the approved inputs, prompt version, model and reviewer.

Illustrative tool selector

This is a governance-demo routing model, not a live product entitlement check.

Hallucination and latency command center

Train users to choose the right reliability pattern: source-grounded outputs, explicit uncertainty, human validation, fast routing and measured service levels.

Trust target
0
Unsupported material claims
Latency target
p95
By use case and route
Primary lever
Tokens, tools and retries
Control model
4
Risk-tier routes

Hallucination risk trainer

Hallucination risk score
Calculating...

Latency design simulator

Input tokens6000
Output tokens800
Tool / retrieval calls3
Estimated response0s
p95 planning view0s
Recommended lever-

World-class reliability prompt block

RELIABILITY RULES
1. Use only the approved source packet and cite the evidence ID for every material claim.
2. Separate Fact, Calculation, Assumption, Inference, Recommendation and Evidence Gap.
3. If evidence is missing, stale or conflicting, do not resolve silently. Put it in the Evidence Gap table and ask for the missing source.
4. Before finalizing, run a verification pass: unsupported claims, stale dates, conflicting numbers, invented owners, invented APIs and overconfident wording.
5. Keep the response inside the output budget. If the answer cannot fit safely, return an executive summary plus a validation queue.

Prompting maturity model

Move from ad-hoc personal practice to an optimized organizational capability.

01

Ad Hoc

Inconsistent prompts, variable quality, little reuse.

02

Repeatable

Personal templates and some consistent patterns.

03

Defined

Shared standards, templates and review expectations.

04

Managed

Governance, calibration, metrics and quality controls.

05

Optimized

Prompting is strategic, measurable and continuously improved.

Select a maturity level to view the next recommended move.

Prompting ROI calculator

Model the annual value of better prompts using task volume, time saved, labor cost and adoption.

Users40
Tasks per user / month8
Minutes saved per task35
Loaded hourly cost$80
Adoption65%
Estimated annual value
$116,480

Value created through faster completion, reduced rework and improved output consistency.

Annual hours saved1,456
Monthly value$9,707
Adopted users26

30-60-90 day implementation roadmap

A pragmatic rollout plan for converting prompting from isolated skill into governed team capability.

Copied