Content based on claude.ai features as of June 2026 · Anthropic Claude Associate Blueprint v1.3
D1 Description
D2 Discernment
D3 Diligence
D4 Delegation
1
Module 1
Product & Model Selection
The Four Decisions — set before the first prompt
Entry Points
Chat
One-off interactions. No persistent context. Best for exploratory tasks.
ProjectsRecurs + stable context + consistent format (2 of 3). Persistent workspace.
ArtifactsDeliverable as a separate editable block — documents, reports, code.
ResearchDeep multi-source synthesis. Paid plans. For current, multi-source investigations.
4-Layer Capability Model
Projects
Carry context — standing instructions + knowledge base + conversation history
SkillsDefine procedures — account level, reusable, invoked automatically
Code ExecVerify computations — sandboxed, returns a guaranteed correct result
MemoryPersist continuity — facts that carry forward, Project-scoped for isolation
Model Selection
| Tier | When to use |
|---|---|
| Haiku | High-volume, structured, clear rubric. Speed + cost efficiency compound. |
| Sonnet | Default for professional work: drafting, synthesis, analysis. |
| Opus | Complex judgment, ambiguous multi-layered inputs, high-stakes quality. |
Context Degradation — 3 Responses
Restart
·
Summarise
·
Persist to Memory
Project Test: Build a Project when 2 of 3 apply: task recurs · stable context · consistent format. Model tier and time spent are NOT criteria.
2
Module 2
Prompting & Task Execution
The 5-Component Stack — Description competency
The 5-Component Prompt Stack
Role
Vocabulary, depth, assumptions — e.g. "as a financial analyst"
ContextBackground Claude cannot infer: audience, situation, prior decisions, source material. Most often missing.
TaskThe specific action with one primary verb
ConstraintsBoundaries: length, tone, inclusions, exclusions
FormatShape of result: table, list, memo, email
Diagnostic Iteration
| Symptom | Root Cause | Fix |
|---|---|---|
| Generic / off-base | Thin Context | Add background |
| Wrong question answered | Ambiguous Task verb | Sharpen instruction |
| Wrong length / tone / shape | Missing Constraint or Format | Add it |
| Close but one section off | Partial gap | Iterate that section only |
Strategy by Task Type
| Type | Tighten | Loosen |
|---|---|---|
| Analysis | Criteria, standards, scope | Phrasing |
| Research | Question, sources, citations | Synthesis approach |
| Drafting | Audience, tone, format | Word choice |
| Brainstorming | Goal + guardrails only | Quantity and direction |
Change ONE component per iteration. Stop when further rounds produce marginal change — diminishing returns is the signal to stop.
3
Module 3
Evaluating & Validating Output
Discernment + Diligence — you own what you ship
5 Behavioral Properties of Claude
- Responses vary — probability-based, not deterministic
- Confident tone ≠ accuracy — fluency is uniform regardless of correctness
- Context is a budget — degrades as sessions grow long
- Knowledge has a training cutoff — post-cutoff needs a live source
- Configured procedures still vary — Skills reduce, not eliminate variance
7 Hallucination Types
High risk
Fabricated specifics — precise-looking stats or citations that don't exist. The precision is the tell.
High riskCapability hallucination — claims to have taken an external action it didn't.
Watch forCompleteness failure — confident summary masking silence on missing data.
Watch forConfirmation bias — output frames evidence to support the premise.
CommonPlausible-but-unsupported claims · Confident tone masking uncertainty · Internal contradictions (long outputs)
4 Risk Thresholds (Diligence)
Stakes
Cost if wrong?
Cost if wrong?
Reversibility
Can it be undone?
Can it be undone?
Audience
Who sees it?
Who sees it?
Regulatory
Does a rule govern this?
Does a rule govern this?
4
Module 4
Workflow Integration & Solution Design
Delegation mapping — Delegation competency
Delegation Mapping — 3 Criteria
Reversibility
Can this step be undone if the output is wrong?
StakesWhat is the cost of an error at this step?
AccountabilityWho is answerable? Accountability does not delegate.
3 Step Classifications
| Classification | Profile | Example |
|---|---|---|
| AI-appropriate | Reversible, low stakes, mechanical | Extract clauses, classify tickets |
| Collaborative | AI drafts, human judges. Gate must be real. | Draft redline → lawyer reviews |
| Human-retained | Irreversible, accountability-bearing | Approve/reject, sign and send |
3 Common Mapping Errors
- Halo delegation — prior step went well → next step auto-delegated
- Collapsing collaborative — 'AI drafts, human reviews' → review gate never staffed
- Mapping the tool — building around features you like, not actual step risk
Escalation signal: When people rely on a Claude-built artifact as infrastructure → escalate to Developer or Architect. Prompt-and-iterate has run its course.
5
Module 5
Configuration & Knowledge Management
The right home for each type of content
4 Configuration Mechanisms
| Mechanism | What goes here | Where it lives |
|---|---|---|
| Standing Instructions | Behaviour rules: tone, format defaults, verification habits | Inside a Project |
| Knowledge Base | Facts, reference documents, policies | Inside a Project |
| Skills | Repeatable multi-step procedures | Account level — reusable across all Projects |
| Scoped Memory | Evolving project state: decisions, stakeholder names | Project-isolated |
Instruction Precision Test
Vague (fails silently)
"Make the reports good and accurate."
Two people read it differently.
"Make the reports good and accurate."
Two people read it differently.
Precise (changes output)
"For every figure, state its source. Mark unverified figures. Lead with a one-sentence headline."
"For every figure, state its source. Mark unverified figures. Lead with a one-sentence headline."
Routing Rule
Rule → Instruction
·
Reference → Knowledge Base
·
Procedure → Skill
Monthly maintenance cadence. Waiting until output quality drops means drift has already reached deliverables. 'It used to work' = stale config.
6
Module 6
Governance, Risk & Responsible Use
Daily practice — not just audit checkboxes
3 Use-Case Classifications
Fully appropriate
Internal FAQ from approved docs, public summaries, formatting tasks
With reviewCandidate screening, customer responses, analysis feeding decisions. Requires a defined gate.
InappropriateFinal medical/legal determinations, unreviewed eligibility decisions, autonomous external actions
Data Sensitivity — Green / Yellow / Red
Green
Published material, anonymised data, cleared internal docs. Safe to use.
YellowInternal docs not meant to leave the company, named individuals, unreleased material. Review first.
RedRegulated data (health/financial/gov), credentials, third-party confidentiality. Stop and escalate.
Skill Trust — 3 Checks
- Source: Who published it? Unknown = escalate to org review
- Reach: What does it inherit? Skills get full session access — not just the task scope
- Appropriateness: Right tool? Or more capability than the task needs?
Incognito ≠ compliance. Incognito controls persistence only. It does not override org data retention. It does not make red-tier data safe to upload.
7
Module 7
Troubleshooting & Optimization
Cheapest fix first — step 5 is last, never first
4 Root-Cause Failure Patterns
| Pattern | Signal | Fix |
|---|---|---|
| Under-specification | Wrong on first response | Add missing component(s) |
| Context overload | Degrades over time in session | Restart or summarise |
| Wrong feature/model | Specific, repeatable error type | Use the right tool |
| Stale configuration | "It used to work" | Run M5 maintenance checklist |
Diagnostic Sequence — Cheapest First
- Re-read prompt vs. 5 components: under-specified?
- Check conversation length: context overloaded?
- Check feature and model: needs Code Execution? Wrong tier?
- Check configuration: instructions, knowledge, Skills current?
- Question whether the task is a fit at all (last resort)
3 Friction Signals → Fixes
Repetition
Same paste every run → saved context or standing instruction
CorrectionSame flaw every output → configuration change
VarianceDifferent results, same task → shared Skill or knowledge base
Capture the fix: phrasing that works → standing instruction. Multi-step fix → Skill. Leaving it in a one-off chat means rediscovering it every cycle.
8
Module 8
Course Summary & Next Steps
The arc — and where Associate scope ends
AI Fluency Framework — 4 Competencies
D1
Description — Telling Claude precisely what you want. Prompting discipline.
D2
Discernment — Critically evaluating output against requirements, sources, standards.
D3
Diligence — Owning and verifying output. Accountability does not delegate.
D4
Delegation — Deciding which steps Claude does, which stay human, which are collaborative.
The 7-Module Arc
Each skill assumes the one before it.
M1 Selection
→
M2 Prompting
→
M3 Evaluation
→
M4 Integration
M5 Configuration
→
M6 Governance
→
M7 Troubleshooting
Certification Boundary
Associate
Individual practitioner. Prompt-and-iterate. This course.
DeveloperAPI, SDKs, agents, tool development. Escalate when prompt-and-iterate is insufficient.
ArchitectEnterprise architecture, integration design, governance at scale.
Recognising the Associate/Developer boundary and escalating across it is itself an Associate-level skill.