Prompting and Task Execution

Knowledge of

• The core prompt elements -- task/goal, context, audience & tone, format, constraints, and examples -- and what each contributes to a usable response
• Positive instructions ('write in plain language') as more reliable than long lists of prohibitions
• Placement of the most important instruction close to the actual task rather than buried in a preamble
• Reusing a liked past output as an example to steer tone and format
• The need to separate instructions from source data (e.g., a 'Source document:' heading) so the model doesn't confuse data with commands
• That Claude only knows what's in the prompt, Project knowledge, and its training data -- not an organization's unstated context

Skills in

• Assembling a prompt that states the task, context, audience/tone, format, and constraints explicitly
• Writing positively framed instructions instead of long negative lists
• Placing the key instruction near the actual task rather than in a buried preamble
• Supplying an example of desired style/format to steer output
• Clearly delimiting source data from instructions with a heading or label

Knowledge of

• Decomposition as breaking a goal into ordered, checkable steps rather than one large vague ask
• Sequential decomposition (prompt chaining): completing and reviewing one step before feeding it into the next
• Structured single-prompt decomposition: spelling out ordered sub-steps within one prompt for moderately complex tasks
• The connection between decomposition and the Delegation/Description ideas in Anthropic's AI Fluency framework
• Decomposition as the mechanism that makes intermediate verification possible

Skills in

• Choosing between sequential prompt chaining and a structured single prompt based on task complexity
• Sequencing a multi-part request into ordered sub-steps
• Reviewing an intermediate result before feeding it into the next step
• Avoiding solving a multi-part problem in one vague prompt when it should be sequenced
• Distinguishing decomposition (structure and order) from simply lengthening a prompt

Knowledge of

• Iteration as diagnosing a specific gap (format, detail, tone, factual error, length) and making a targeted change
• The one-change-at-a-time discipline that lets you attribute an improvement (or regression) to a specific edit
• The read-diagnose-change-recheck loop as the mechanism for improving output over successive turns
• The connection between iteration and the Diligence competency in AI Fluency -- staying responsible for judging and refining the result
• Why repeated regeneration without changing the prompt is not iteration

Skills in

• Reading an output against original intent to identify exactly what's wrong
• Making one targeted change per iteration (adding a constraint, an example, tightening audience, or splitting the task)
• Re-running and comparing against the prior version before keeping or discarding a change
• Avoiding reflexive regeneration in place of a diagnosed edit
• Avoiding changing multiple variables in a single iteration

Knowledge of

• The distinct prompt emphasis each task type rewards: analysis (source data plus a reasoning lens), research (structured findings, verifiable sourcing, explicit scope/recency), drafting (audience, tone, length, format, plus a style example), brainstorming (quantity and range, with filtering deferred)
• Asking Claude to reason before concluding as a quality lever for analysis and multi-step reasoning tasks
• Why that reasoning-first overhead is unnecessary for simple lookups or short drafts
• The need for verifiable sourcing on research-style asks, since confident-sounding output can still be unverified
• Why a single rigid prompt format does not fit every task type

Skills in

• Emphasizing source data and an explicit analytical lens for analysis tasks
• Requesting structured findings and verifiable sources with explicit scope/recency for research tasks
• Specifying audience, tone, length, format, and a style example for drafting tasks
• Inviting quantity and range and deferring filtering for brainstorming tasks
• Asking for reasoning before a final answer on analysis/multi-step tasks, while skipping that overhead for simple asks