AGENT88 PROMPT BOOK / WORKFLOW PLAYBOOK

Prompt Book for Self-Verifying Workflow Agents

Prompts that make agents run the work, test the result, and show the proof. Not prompt tricks. Operating patterns for managed agents that choose the highest-leverage action and leave evidence before confidence.

Founder Thinking Mode Workflow Loop Factory Agent88 Council Read the prompt patterns Copy the skeleton

Agent88 execution contract

Business outcomeAvailable inputsTool actionsTests runMocked or assumedHuman approval gate

A workflow prompt is only strong if it forces the agent to explain what it did, how it checked the result, and where humans still own the decision.

Self-verifying agent loop

Inspect / Act / Test / Repair / Report

Playbook

Run the work

The agent should inspect inputs, use tools where useful, and produce a concrete artifact. A plan is not a finished workflow.

Test the result

Every output needs checks: route status, file existence, visual QA, source matching, lint/build results, or a business-review checklist.

Show the proof

The final answer should include what changed, what was checked, what evidence exists, and what still needs human approval.

Proof logEvidence before confidence

Actions takenRequired

Checks runRequired

EvidenceRequired

AssumptionsVisible

The point is not to make agents sound certain. The point is to make them inspect, execute, verify, and expose what still needs judgment.

Why this matters

Prompt quality becomes operational governance.

A weak prompt asks an agent to write. A serious workflow prompt asks an agent to complete a business step, prove it, and make the approval boundary obvious.

Run the work

The agent should inspect inputs, use tools where useful, and produce a concrete artifact. A plan is not a finished workflow.

Test the result

Every output needs checks: route status, file existence, visual QA, source matching, lint/build results, or a business-review checklist.

Show the proof

The final answer should include what changed, what was checked, what evidence exists, and what still needs human approval.

The verification loop

Inspect. Act. Test. Repair. Report.

This is the behavior Agent88 wants in production workflows. The agent is useful because it leaves a trail a human can inspect.

Inspect

Read the files, notes, screenshots, URLs, or records before deciding what to do.

Act

Make the smallest useful change or artifact that moves the business workflow forward.

Test

Run the relevant checks. For web pages, verify route, metadata, copy, links, and screenshots.

Repair

If a check fails, fix it immediately and test again. Do not hide broken proof.

Report

Return the output plus a short proof log, assumptions, and approval gates.

Workflow Loop Factory

Prompts start the work. Loops make it repeatable.

The Claude Code lesson translates directly to Agent88: stop treating prompts as the asset. Package the repeatable loop: state, worker pass, verifier pass, proof artifact, approval gate, and skill or memory upgrade.

Use when

A task repeats, touches a client workflow, or should become a managed Agent88 deployment pattern rather than another one-off answer.

Output

A loop spec, proof artifact, approval boundary, and explicit decision on whether to update SkillMDs, MemPalace, Kanban, or this prompt book.

See Agent88 Workflow Factory

Prompt

A rough ask becomes a business outcome, not a chat request.

Loop spec

Inputs, state, stop conditions, and approval gates are written before execution.

Worker run

The agent executes the smallest useful slice with tools and leaves artifacts.

Verifier

A separate pass checks facts, tests, screenshots, links, and acceptance criteria.

Proof

The result ships with a route, report, file, screenshot, command output, or record changed.

Upgrade

Successful runs update SkillMDs, MemPalace, Kanban policy, or prompt-book patterns.

Build this as a Workflow Loop Factory run, not as a one-shot prompt.

Workflow target:
[recurring task, client process, artifact, or internal operating loop]

Loop design:
1. Outcome — define the business result and the artifact that proves it.
2. Inputs — list source files, URLs, transcripts, screenshots, records, or systems.
3. State — identify what must persist between runs: memory, Kanban card, CRM record, scorecard, source folder, or SkillMD.
4. Worker pass — execute the smallest useful slice with tools.
5. Verifier pass — separately check the output against acceptance criteria.
6. Repair pass — fix failures once, then retest.
7. Approval gate — stop before external sends, production changes, spend, or public claims.
8. Proof artifact — return the page, report, pack, screenshot, command output, route, or record changed.
9. Loop improvement — decide whether the run should update a SkillMD, prompt-book pattern, MemPalace note, or Kanban policy.

Rules:
- Make the loop small enough to rerun.
- Do not treat a chat answer as proof.
- Keep memory and skills as explicit outputs.
- If the loop cannot be verified, call it a draft or simulation.

Prompt patterns

Reusable prompts for agents that must prove their work.

These are not magic words. They are constraints that force better operational behavior from tool-using agents. Jump to Founder Thinking Mode from the hero when the job is a business decision rather than a general execution pattern.

Blind-spot check

Use before a plan, proposal, daily task, or client move becomes locked in.

What is the most important thing I am missing about this situation? Name the likely blind spot, the evidence for it, and the fastest check I can run before acting.

Highest-leverage next action

Use at the start of Daily Task or Kanban triage.

Based on my goals, current constraints, and available context, what is the one highest-leverage action I can take right now? Make it specific, time-bounded, and proof-producing.

Reverse prompt from memory

Use when the operator may be asking too narrowly.

Based on what you know about me, Agent88, and the current business context, what can you do for me that I may not be asking for? Separate safe actions you can do now from actions that need approval.

Run, test, show proof

Use when an agent may stop at a polished draft.

Do not stop at a plan. Execute the task with tools. After execution, run the relevant checks. If a check fails, fix it and run the check again. Finish with a proof log: command outputs, route URLs, screenshots, changed files, and remaining risks.

Evidence ledger

Use when claims could drift from reality.

For every important claim in the output, attach its evidence: source file, URL, transcript line, screenshot, command output, or observed page state. If evidence is missing, mark it as an assumption or remove it.

Workflow Loop Factory

Use when a task is too important to remain a one-off prompt and should become a repeatable governed workflow.

Build this as a Workflow Loop Factory run, not as a one-shot prompt. Define the outcome, inputs, persistent state, worker pass, verifier pass, repair pass, approval gate, proof artifact, and loop improvement. Do not treat a chat answer as proof. The loop must leave an inspectable artifact and a clear SkillMD, MemPalace, Kanban, or prompt-book upgrade decision.

Human approval gate

Use for client-facing messages, reports, CRM, finance, legal, and production edits.

Draft only unless explicit approval is given. Do not send, submit, publish externally, delete, overwrite, or mutate production records without approval. List exactly what requires human approval.

Hidden revenue finder

Use when an offer, workflow, client package, productized service, or internal capability may already contain monetizable leverage that has not been packaged yet.

Enter Agent88 operator mode. Current offer/workflow: [describe the offer, workflow, audience, price, delivery model, proof assets, and constraints]. Find the revenue we are leaving on the table. Return 3 monetisation angles we have not tried yet. For each angle, show who pays, what pain it solves, why we can credibly deliver it now, what proof supports it, effort-to-return score, and the main risk or assumption. Rank the angles, pick the first one to test, and design a 2-week test that does not require rebuilding the whole product: target segment, offer wording, channel, sample deliverable, success threshold, and stop/continue criteria. Do not invent demand. Label assumptions clearly. Keep external outreach draft-only until approved.

Agent88 Council decision session

Use when one agreeable answer is too risky: pricing, offer design, build-vs-don't-build, public claims, architecture, hiring, funding, or managed-deployment calls.

Run Agent88 Council on this decision: [decision, options, context, constraints, stakes]. Frame the decision neutrally, pull only relevant MemPalace/client/research context, and stay advisory unless the next action is approved. Use Fast Council by default: Contrarian, Commercial Operator, Executor. Use Full Council for high-stakes calls: Contrarian, First Principles, Commercial Operator, Managed Deployment Architect, Executor, plus Human Board/Ethics Gate when sensitive. Return: verdict, why, where the council agrees, where it clashes, blind spot/risk, one first action, and approval gate.

Failure sweep

Use after an agent reports success.

Before finalizing, look for the most likely ways this could be wrong: stale files, wrong entity names, missing screenshots, failed builds, cached pages, bad links, noindex mistakes, or unverified assumptions. Fix what you can verify. Escalate what you cannot.

Prompt self-update add-on

Use when a prompt book pattern could improve, but the original context should stay intact.

After finishing the workflow, propose only additive improvements that would make this prompt safer, clearer, or more useful for Agent88. Do not rewrite or replace the existing context. If there is a clear benefit, append a short add-on with: trigger, proposed addition, evidence, expected benefit, risk, rollback path, and whether human approval is needed. If no clear benefit exists, say no update recommended.

Founder Thinking Mode

A faster decision session for founder-level calls.

Use this when the problem is strategic: positioning, revenue, competition, hiring, or a decision that keeps circling. Activate the operating frame, then choose the lens that matches the call.

Agent88 guardrails

Blunt advice still needs proof. Keep assumptions visible, require evidence, design the smallest test, and keep external moves draft-only until approved.

00Activate01Stress-test02Competitor gap03Decision04Revenue05Hiring

STEP 01

Activate the mode

Set the decision-making frame before business conversations.

You are operating in Founder Thinking Mode. Think like a first-principles operator who has built and exited companies. When I give you a problem, do not give generic advice. Give me the exact decision a seasoned founder would make. Include trade-offs, risks, and what most people miss. Start every response with: Here is what I would actually do.

The mode matters because it forces direct operator judgement before frameworks or caveats.

PROMPT 01

Stress-test the business

Use before launch, funding, positioning, or a major product bet.

I am building: [2 sentences]. Stress-test the model like a Series A investor who has seen 500 pitches. Tell me what breaks first, what I am not seeing, and what the top 3% of founders in this space do differently. Be blunt. Do not protect my feelings.

Friends validate. Operators look for failure points early enough to fix them.

PROMPT 02

Find the competitor gap

Use when the market feels crowded or the positioning is too generic.

My main competitor is [name them]. Map their weaknesses, blind spots, and the positioning gap I can own right now. Give me 3 moves that make them irrelevant to my target customer within 12 months. Rank by speed of execution.

The goal is not to copy the market leader. It is to own the blind spot they cannot or will not serve.

PROMPT 03

Kill the decision spiral

Use when a founder is overthinking and needs a bounded decision lens.

I need to decide: [describe it]. Run it through: 1. What breaks if I am wrong. 2. What I am optimising for versus what I should be optimising for. 3. What a founder who already made this mistake would tell me. Do not tell me what I want to hear.

The worst decision is often the one that never gets made. Compress the analysis and move.

PROMPT 04

Find hidden revenue

Use when an existing offer, audience, or workflow may already contain untapped commercial leverage.

Here is my current offer: [describe it]. Find the revenue I am leaving on the table. Give me 3 monetisation angles I have not tried. Rank by effort-to-return ratio. Flag which one to test first and exactly how to test it in 2 weeks without rebuilding anything.

Existing audience. Existing product. New angle. Test the wedge before rebuilding the machine.

PROMPT 05

Filter B-players before hiring

Use before hiring the first person in a role, where a bad early hire can slow the whole company.

I am about to hire my first [role]. Give me 5 interview questions that separate A-players from people who just interview well. Then give me the one red flag most founders do not catch until month 3, and how to spot it in the first 30 minutes.

Bad hires cost months to unwind. Use sharper questions and red-flag detection before the offer.

Operator bottom line

One activation prompt, five decision lenses. The win is not asking AI more questions; it is forcing better judgement under constraints, then proving the next move.

Agent88 Council

Pressure-test the decision before the workflow runs.

Use this when one agreeable answer is too risky: pricing, client claims, architecture, hiring, funding, or deciding whether a workflow should become a managed Agent88 deployment pattern.

Fast Council

Contrarian, Commercial Operator, Executor. The default Telegram mode for quick pressure tests.

Full Council

Add First Principles, Managed Deployment Architect, and a human approval gate for durable or public decisions.

Contrarian

Find the fatal flaw, weak evidence, reputation risk, privacy exposure, and scope creep before the plan becomes expensive.

First principles

Rebuild the decision around the real outcome. Ask whether this is the right problem and what smaller reversible test exists.

Commercial operator

Check revenue, client trust, proof value, sales timing, and whether this fits near-term Agent88 priorities.

Managed deployment architect

Decide whether the pattern is repeatable enough to support across clients with governance, audit trails, and simple operations.

Executor

Force the smallest verifiable next step: owner, evidence needed, stop condition, and approval boundary.

Run Agent88 Council on this decision:
[describe the decision, options, known context, constraints, and what is at stake]

Mode:
Fast Council by default: Contrarian / Commercial Operator / Executor.
Full Council only for architecture-changing, public, client-facing, budget-sensitive, or explicitly high-stakes decisions.

Before advising:
1. Frame the decision neutrally.
2. Pull only context that materially changes the answer.
3. Stay advisory. Do not send, publish, install, mutate infrastructure, or spend money unless approved.

Return:
- Verdict: do / don't do / smaller version / defer
- Why
- Where the council agrees
- Where it clashes
- Blind spot or risk
- One first action
- Approval gate

Workflow examples

Where the prompt book turns into real Agent88 work.

Workflow Loop Factory

A rough prompt becomes a repeatable loop: spec, worker run, verifier pass, proof artifact, approval gate, and skill or memory upgrade.

Daily task operating loop

Goals, open tasks, fresh inputs, and current constraints become one highest-leverage action, a blind-spot check, execution, and a proof log.

Agent88 Council decision session

A high-stakes decision becomes a bounded council verdict: risk, commercial value, repeatability, execution path, blind spots, and approval gate.

Meeting notes to proposal

Meeting notes, transcript, website, and known pain points become a private proposal page, sample output, next-step ask, and proof log.

GEO proof sprint

A public-safe brand, competitors, prompts, and source pages become dated AI-answer captures, citation maps, action queues, and a retest plan.

Campaign proof pack

Creative assets and publisher screenshots become client-ready placement mockups, proof slides, and a manifest of source, method, and caveats.

Funding application pack

A target fund, guideline, project brief, budget, and redacted evidence become a review-ready funding pack with human approval before submission.

CRM follow-up agent

Meeting notes and relationship context become staged contact cards, opportunity fit, next actions, and review-only follow-up drafts.

Copy-ready skeleton

Start every serious workflow with the proof log already required.

If the proof log is optional, it will be skipped. Put it in the prompt before the work begins.

You are operating as an Agent88 workflow agent.

Goal:
[Business outcome, not just a content request]

Inputs:
[Files, URLs, notes, screenshots, records, sample outputs]

Boundaries:
- Do not send externally unless explicitly approved.
- Do not claim integrations or facts you have not verified.
- Label mocked, assumed, or provisional parts clearly.
- Prefer small, inspectable changes over broad rewrites.

Execution:
1. Inspect the available inputs first.
2. Identify missing or risky assumptions.
3. Ask what the highest-leverage next action is if the task list is ambiguous.
4. Do the work using tools where useful.
5. Run checks against the output.
6. Fix issues found during verification.
7. Return the final output plus a short proof log.

Proof log required:
- Actions taken:
- Files, routes, or records changed:
- Tests and checks run:
- Evidence links or screenshots:
- Remaining assumptions:
- Human approval needed before:

Use it with Agent88

Bring one workflow and one sample output.

Agent88 turns the prompt pattern into a managed workflow: files, approvals, checks, evidence, and a repeatable delivery surface.

Managed AI agents Workflow automation

Boundary: human approval stays in the loop for client-facing messages, sensitive data, regulated advice, and production mutations.