Skip to content

9. Prompting & Customizing the Agents

← FAQ & Glossary · Next: Your Master Career Dossier →

Every agent in Maestro is driven by a system prompt — a structured instruction set that tells the AI exactly what to do. These prompts live in the prompts tab of your database, and you can edit them without touching any code. This page explains the framework they follow, why it works, and how to customize them safely.

💡 You don't have to touch prompts to use Maestro. The defaults are tuned and ship ready to run. Read this only when you want to change an agent's behavior — e.g. a different resume style, stricter fact-checking, or a new tone.

Where prompts live

Each of the ten agents has one row in the prompts tab:

Column Meaning
agent_id Which agent (agent_1 … agent_10)
default_markdown The shipped prompt — leave this alone
override_markdown Your custom prompt. If set, it replaces the default for that agent.
updated_at, description Metadata 
prompts tab
┌──────────┬──────────────────────┬──────────────────────┬─────────────────────┐
│ agent_id │ default_markdown      │ override_markdown     │ description          │
├──────────┼──────────────────────┼──────────────────────┼─────────────────────┤
│ agent_1  │ ## ROLE\nYou are a…   │  ← (empty: uses       │ Resume Builder       │
│          │ [the shipped prompt]  │     default)          │                      │
│ agent_5  │ ## ROLE\nYou are a…   │ ## ROLE\nYou are a…   │ Critic               │
│          │ [the shipped prompt]  │ [YOUR version wins]   │                      │
└──────────┴──────────────────────┴──────────────────────┴─────────────────────┘
                                          ▲
                          Put your edits HERE, not in default_markdown.
                          Blank override = the default is used.

The override pattern is your safety net: the original prompt always stays intact in default_markdown. To revert a customization, just clear your override_markdown cell.

The agent roster at a glance

agent_id Agent What its prompt controls
agent_1 Resume Builder How resumes are drafted and structured
agent_2 Cover Letter Builder Cover letter style and content
agent_3 Resume Verifier How strictly claims are fact-checked
agent_4 Cover Letter Verifier Fact-checking for cover letters
agent_5 Critic What counts as a weakness worth fixing
agent_6 Resume Polisher How the first refinement applies critique
agent_7 Resume Scorer How fit scores are assigned
agent_8 Discovery Scorer How discovered jobs are scored (two axes)
agent_9 Resume Fine-Refiner How later refinements apply your instructions
agent_10 Job Ranker How the fast first-pass ranking works

The prompt framework

Every Maestro prompt follows the same skeleton. This consistency is deliberate — it's what makes the system predictable, debuggable, and safe to edit. Learn the skeleton once and you can read or modify any agent.

## ROLE              ← who the agent is, in one or two sentences
## CONTEXT           ← what inputs it receives, and which is the source of truth
## INSTRUCTIONS      ← the concrete output structure (builders)
   or
## WHAT TO EVALUATE  ← what to inspect (verifiers / critics / scorers)
## METHOD            ← the step-by-step procedure to follow
## TONE & VOICE      ← style rules (builders only)
## SEVERITY GUIDANCE ← how to rank issues (verifiers / critics)
   or
## SCORING GUIDANCE  ← how to assign numbers (scorers)
## CONSTRAINTS       ← hard "do NOT" rules — the guardrails
## EDGE CASES        ← how to handle tricky inputs
## SUCCESS CRITERIA  ← what a good output looks like
## EXAMPLES          ← worked good/bad examples

Not every agent uses every section — a builder has INSTRUCTIONS and TONE & VOICE; a verifier has WHAT TO EVALUATE and SEVERITY GUIDANCE instead. But the spine is always the same, in the same order.

Why each section earns its place

Section The problem it solves
ROLE Sets the persona so the model adopts the right expertise and standards from the first token.
CONTEXT Names the inputs and — critically — declares the single source of truth (the master document). This is what stops fabrication.
INSTRUCTIONS / WHAT TO EVALUATE Pins the output to an exact structure, so results are consistent and the pipeline can parse them.
METHOD A numbered procedure turns a vague task into repeatable steps, which dramatically improves reliability.
CONSTRAINTS Explicit "do NOT" rules are the guardrails. "Do NOT invent facts" is the single most important line in the whole system.
EDGE CASES Pre-answers the ambiguous situations (sparse JD, employment gaps, thin experience) so the model doesn't improvise badly.
SEVERITY / SCORING GUIDANCE Calibrates judgment to a shared scale, so a "critical" issue means the same thing every run.
SUCCESS CRITERIA Gives the model a target to self-check against before it answers.
EXAMPLES One concrete good-vs-bad pair teaches more than a paragraph of description.

Why this framework works

Three principles make it effective:

  1. Truth is structural, not hoped-for. The system doesn't ask the model to be honest — it architects honesty. The master document is declared the sole source of truth in CONTEXT, fabrication is forbidden in CONSTRAINTS, and a separate Verifier agent (Agent 3) independently fact-checks the result. Honesty is enforced at three layers, not one.

  2. Specific beats vague, everywhere. Vague prompts produce vague output. The framework forces specificity at every turn: a numbered METHOD, an explicit output structure, calibrated severity scales, and worked EXAMPLES. The Critic's own prompt says it best: "vague feedback produces vague fixes."

  3. Separation of powers. No single agent both writes and judges its own work. The Builder drafts, the Critic critiques, the Refiner fixes, the Verifier fact-checks, the Scorer rates. Each prompt has one job and does it well — which is easier to get right, easier to debug, and harder to fool.

A real example: the Resume Builder (Agent 1)

Here's how the skeleton looks filled in (excerpted from the shipped default):

## ROLE
You are a professional resume writer. You produce a tailored resume that
matches a specific job description while staying faithful to the candidate's
actual career history…

## CONTEXT
You will receive:
1. A job description (JD) for a specific role.
2. The candidate's master career document — the authoritative source for
   every role, achievement, metric, and credential the candidate actually has.
Your output is a tailored resume in markdown, drawn entirely from facts in
the master document. You select and emphasize relevant content; you never
invent it.

## INSTRUCTIONS
Produce the resume using this standard structure, in order:
1. Headline …
2. Professional Summary …
3. Professional Experience …

## CONSTRAINTS
- Do NOT invent facts; every claim must trace to the master document.
- Do NOT fabricate or alter metrics, titles, employers, or dates.
- Do NOT inflate the candidate's seniority …

## EXAMPLES
Example A — bullet point (weak to strong):
WEAK:   "Responsible for managing the team and improving processes."
STRONG: "Managed a 6-person team and streamlined the intake process,
         reducing turnaround time by 30%."
WHY: the strong version uses an action verb, specifies scope, and
     quantifies the result.

Notice how CONTEXT names the master document as authoritative, CONSTRAINTS forbids invention three different ways, and EXAMPLES shows the exact quality bar. That's the framework doing its job.

How to customize a prompt

Step 1 — Read the default first

Open the prompts tab, find your agent's row, and read its default_markdown end to end. You can't safely edit what you don't understand, and the defaults already handle most cases.

Step 2 — Copy it into override_markdown

Copy the entire default_markdown into the same row's override_markdown cell. Always start from the full default — never write an override from scratch, or you'll lose the guardrails that keep the system honest.

Step 3 — Make your change inside the framework

Edit the section that owns your change. Keep the skeleton intact.

You want to… Edit this section
Change resume tone or length TONE & VOICE and/or INSTRUCTIONS
Add/remove a resume section INSTRUCTIONS
Make fact-checking stricter or looser SEVERITY GUIDANCE and CONSTRAINTS in Agent 3
Change what the critic flags WHAT TO EVALUATE and SEVERITY GUIDANCE in Agent 5
Adjust how fit scores are assigned SCORING GUIDANCE in Agent 7 / Agent 8
Add a rule the agent keeps breaking Add a line to CONSTRAINTS
Teach the agent a tricky case Add a bullet to EDGE CASES

Step 4 — Test on one job

Run a single build (or discovery) and inspect the result. If it's worse, clear the override_markdown cell to instantly revert to the default.

Golden rules for editing

  • ✅ Always keep the source-of-truth and no-fabrication rules. If you remove "do NOT invent facts" from a builder, you break the whole honesty guarantee.
  • ✅ Add constraints as explicit "do NOT" lines. Models follow concrete prohibitions far better than vague preferences.
  • ✅ Update the EXAMPLES if you change the output. A stale example confuses the model more than no example.
  • ✅ Change one thing at a time, then test. Multiple simultaneous edits make failures impossible to diagnose.
  • ❌ Don't delete whole sections unless you truly understand the consequence — each one is load-bearing.
  • ❌ Don't make builder prompts judge themselves — that's the verifier's and critic's job, by design.

Prompting tips that transfer

These principles apply whether you're editing Maestro's prompts or writing your own anywhere:

  • Be specific and detailed. "Quantify results where the master document provides numbers" beats "make it impressive."
  • Use positive and negative examples. Show the good version and the bad version, and say why.
  • Encourage step-by-step reasoning. A numbered METHOD produces more reliable output than a single instruction.
  • State the output format explicitly. If you need structured output, describe its shape exactly.
  • Name the constraints. The things you don't want are as important as the things you do.

For a deeper reference on prompting, see Anthropic's guide: https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview.

← FAQ & Glossary · Next: Your Master Career Dossier →