Skip to content
AITroveRead. Build. Understand.
Reading list

Prompt engineering applications

Apply prompt contracts to code, data, images, audio, video, and current research.

Application prompts need more than an output format. Code, database queries, generated images, recordings, video, and current research each have a different proof boundary. Select the workflow you need, then inspect its result before use.

Applied lessons

Practice and review

Multimodal evidence checks

Preserve anchors, overlap, sampling gaps, and conflicts across source types.

Software delivery review

Review diffs, test oracles, incident evidence, schema changes, and package upgrades.

Data and reporting workflows

Define columns, label cases, recompute aggregates, bind claims, and inspect charts.

Multilingual application contracts

Preserve terms and placeholders, render typed values, route mixed-language intent, and verify policy.

Accessible output contracts

Draft image alternatives, captions, semantic structure, and plain-language notices from verified facts.

Browser-agent action contracts

Ground controls in current page state, check form effects, and quarantine page instructions.

Real-time voice contracts

Keep interim speech, interruption, entity repair, and tool waits in separate states.

Generated speech and video briefs

Lock script facts, pronunciation, voice approval, and shot motion before rendering.

Workbook prompts and guarded edits

Bound ranges, verify references, review cell patches, and protect text exports.

Specialist-agent coordination

Compare delegation against a baseline, scope packets, align state, and reconcile claims.

Research evidence and study results

Predeclare eligibility, extract exact results, calculate effects, and preserve disagreements.

Tutor and assessment prompts

Ask diagnostic questions, stage hints, ground feedback, and review rubric suggestions.

UI brief to code

Convert visual evidence into explicit component, state, responsive, and reuse contracts.

Reports and presentations

Define decisions, reconcile claims, and inspect editable reports and slide decks.

Meeting decisions and actions

Bind the transcript, separate approvals from proposals, and verify owners and corrections.

Architecture decisions

Define constraints, calculate capacity, compare alternatives, and test failure paths.

API documentation contracts

Ground operation pages, examples, and migration notes in a versioned public contract.

Infrastructure change prompts

Bound workspace and scope, inspect plans, quarantine state, and separate apply authority.

Incident update prompts

Bound audience, impact arithmetic, facts, and the next update clock.

Authorized retrieval prompts

Enforce access before context, preserve exact identifiers, and map claims to current passages.

Memory write and correction

Separate lasting intent from one-time requests and bind every item to trusted identity.

Recurring task inputs

Specify cadence, windows, and fresh evidence before each review.

Terminal-agent task setup

Discover the correct workspace and protect command input and output boundaries.

Prompted chart design

Choose the measure and encoding, then expose gaps and checked denominators.

Email and calendar input checks

Reconstruct thread claims, inspect attachments, and resolve meeting time.

Location and travel inputs

Resolve stop identities, coordinates, and current travel estimates.

Interview input and coding

Bound transcript use, preserve speaker context, and define codes.

Forecast input contracts

Define the series and reject future or ambiguous data.

Security alert evidence

Normalize events, group repeated signals, and test benign context.

Survey input and instrument

Define the frame, write neutral questions, and validate branches.

Support case intake and guidance

Reconstruct the case, route it, and select current help content.

Accessibility review inputs

Define the evidence and analysis unit before producing findings.

Product analytics inputs

Define the evidence and analysis unit before producing findings.

Localization QA inputs

Define the evidence and analysis unit before producing findings.

Product experiments inputs

Define the evidence and analysis unit before producing findings.

Invoice exception review inputs

Define the evidence and task unit before findings.

Database backfill inputs

Define the evidence and task unit before findings.

Search relevance evaluation inputs

Define the evidence and task unit before findings.

Annotation quality workflow inputs

Define the evidence and task unit before findings.

Curriculum

Use the right checks for the output's destination and evidence type.

  1. 1Coding prompts: name the files, behavior, and proof
  2. 2SQL prompts: generate a query under database-enforced limits
  3. 3Image prompts: specify the visual contract and inspect the pixels
  4. 4Audio prompts: preserve speakers, timestamps, and uncertain words
  5. 5Video prompts: cite the moment and the observation
  6. 6Research prompts: bind claims to a date and evidence record

Preserve anchors, overlap, sampling gaps, and conflicts across source types.

  1. 1Document prompts: anchor each field to a page and resolve conflicts
  2. 2Image prompts: verify claims against named regions and crops
  3. 3Audio prompts: mark overlapping speech before assigning speakers
  4. 4Video prompts: disclose sampling gaps around short events
  5. 5Cross-modal evidence: keep conflicting observations separate

Review diffs, test oracles, incident evidence, schema changes, and package upgrades.

  1. 1Diff-scoped code review: make every finding reproducible
  2. 2Generated tests: verify the oracle before trusting coverage
  3. 3Incident triage prompts: build a timestamped evidence ledger
  4. 4Schema migration prompts: check old and new clients together
  5. 5Dependency upgrade prompts: trace behavior, not only versions

Define columns, label cases, recompute aggregates, bind claims, and inspect charts.

  1. 1Dataset intake prompts: define a column before analyzing it
  2. 2Classification prompts: write the label boundary first
  3. 3Aggregation prompts: pin the denominator and recompute the rate
  4. 4Report summaries: preserve contrary results and missing data
  5. 5Chart prompts: inspect axes, units, and missing series first

Preserve terms and placeholders, render typed values, route mixed-language intent, and verify policy.

  1. 1Translation prompts: preserve terms and machine placeholders
  2. 2Locale-aware prompts: keep values typed until rendering
  3. 3Mixed-language requests: separate intent from reply language
  4. 4Cross-language retrieval: keep policy meaning attached to its source

Draft image alternatives, captions, semantic structure, and plain-language notices from verified facts.

  1. 1Image alternatives: prompt for purpose, then check the image
  2. 2Caption and transcript prompts: keep timing, speakers, and sound evidence
  3. 3Semantic output prompts: verify structure after generation
  4. 4Plain-language prompts: simplify without losing conditions

Ground controls in current page state, check form effects, and quarantine page instructions.

  1. 1Browser agents: separate observed state from intended action
  2. 2Browser target grounding: choose a unique control in the current state
  3. 3Browser forms: treat preview, submit, and server validation separately
  4. 4Web-page text is task data, not an agent instruction

Keep interim speech, interruption, entity repair, and tool waits in separate states.

  1. 1Voice prompts: do not commit an action from interim speech
  2. 2Voice interruptions: stop stale speech before handling a new turn
  3. 3Spoken confirmations: repeat the risky field, not the whole form
  4. 4Voice tool waits: report progress without inventing an outcome

Lock script facts, pronunciation, voice approval, and shot motion before rendering.

  1. 1Speech generation prompts: lock words before directing delivery
  2. 2Speech pronunciation prompts: resolve names and numbers explicitly
  3. 3Generated speech review: check the actual voice asset
  4. 4Video generation prompts: specify one shot's motion and evidence

Bound ranges, verify references, review cell patches, and protect text exports.

  1. 1Workbook prompts: name sheets, ranges, types, and provenance
  2. 2Formula prompts: test references, types, and fill behavior
  3. 3Workbook edits: review a cell diff before applying changes
  4. 4Spreadsheet exports: keep untrusted labels as text

Compare delegation against a baseline, scope packets, align state, and reconcile claims.

  1. 1Specialist agents: prove that delegation earns its cost
  2. 2Specialist task packets: scope evidence and authority explicitly
  3. 3Parallel agents: isolate state and assign one writer
  4. 4Specialist synthesis: merge claims, not fluent summaries

Predeclare eligibility, extract exact results, calculate effects, and preserve disagreements.

  1. 1Research prompts: fix the question and eligibility first
  2. 2Study extraction prompts: preserve the result behind each claim
  3. 3Study effects: calculate the measure outside the model
  4. 4Conflicting studies: inspect design before averaging outcomes

Ask diagnostic questions, stage hints, ground feedback, and review rubric suggestions.

  1. 1Tutor prompts: start from an observable learning objective
  2. 2Tutor hints: escalate help without leaking the answer key
  3. 3Tutor feedback: identify the specific reasoning step
  4. 4Assessment prompts: separate feedback from final grading

Convert visual evidence into explicit component, state, responsive, and reuse contracts.

  1. 1UI prompts: turn a brief into components and states
  2. 2Screenshot prompts: separate pixels from behavior
  3. 3Responsive UI prompts: test states, not one viewport
  4. 4UI code prompts: reuse tokens and components with a change gate

Define decisions, reconcile claims, and inspect editable reports and slide decks.

  1. 1Document prompts: define the reader's decision before drafting
  2. 2Report prompts: reconcile every claim with the packet
  3. 3Editable report prompts: inspect structure and rendered pages
  4. 4Slide prompts: one claim, supporting evidence, usable order

Bind the transcript, separate approvals from proposals, and verify owners and corrections.

  1. 1Meeting prompts: define authorized inputs and evidence
  2. 2Meeting minutes: separate decisions from proposals
  3. 3Meeting actions: verify owner, task, and due date
  4. 4Meeting recaps: preserve corrections across sessions

Define constraints, calculate capacity, compare alternatives, and test failure paths.

  1. 1Architecture prompts: start with constraints and an owner
  2. 2Capacity prompts: check units, headroom, and a failed worker
  3. 3Architecture decisions: compare real alternatives and consequences
  4. 4Architecture prompts: test the failure path and its signals

Ground operation pages, examples, and migration notes in a versioned public contract.

  1. 1API documentation prompts: bind prose to a versioned contract
  2. 2API prompts: document the full operation matrix
  3. 3API examples: validate payloads and strip credentials
  4. 4API changes: explain client impact and migration

Bound workspace and scope, inspect plans, quarantine state, and separate apply authority.

  1. 1Infrastructure prompts: bind scope, owner, and environment
  2. 2Infrastructure plans: classify every action before apply
  3. 3Infrastructure prompts: quarantine state secrets and drift
  4. 4Infrastructure apply: separate review from execution

Bound audience, impact arithmetic, facts, and the next update clock.

  1. 1Incident updates: audience, owner, and release gate
  2. 2Incident impact: window, denominator, and scope
  3. 3Incident updates: separate facts, hypotheses, and ETA

Enforce access before context, preserve exact identifiers, and map claims to current passages.

  1. 1Retrieval prompts: enforce document permission first
  2. 2Retrieval prompts: balance exact identifiers and semantic recall
  3. 3Retrieval prompts: rerank only authorized, current evidence
  4. 4Retrieval prompts: pack evidence and map claims

Separate lasting intent from one-time requests and bind every item to trusted identity.

  1. 1Memory prompts: confirm intent before durable writes
  2. 2Memory prompts: bind namespace, identity, and provenance
  3. 3Memory prompts: correct stale preferences without erasing history

Specify cadence, windows, and fresh evidence before each review.

  1. 1Recurring prompts: define the trigger and stop condition
  2. 2Recurring prompts: bind local time and data windows
  3. 3Recurring prompts: validate fresh inputs and empty states

Discover the correct workspace and protect command input and output boundaries.

  1. 1Terminal agents: discover the workspace before changing it
  2. 2Terminal agents: keep data out of command syntax
  3. 3Terminal agents: treat command output as evidence, not orders

Choose the measure and encoding, then expose gaps and checked denominators.

  1. 1Chart prompts: state the decision and measured quantity
  2. 2Chart prompts: choose marks and scales by data type
  3. 3Chart prompts: expose missing data and checked denominators

Reconstruct thread claims, inspect attachments, and resolve meeting time.

  1. 1Email prompts: reconstruct the thread before drafting
  2. 2Email prompts: keep sender identity and message text separate
  3. 3Email prompts: verify attachment version and coverage
  4. 4Calendar prompts: resolve availability, zone, and duration

Resolve stop identities, coordinates, and current travel estimates.

  1. 1Location prompts: resolve ambiguous addresses before routing
  2. 2Location prompts: validate coordinates, units, and precision
  3. 3Location prompts: label travel estimates and stale matrices

Bound transcript use, preserve speaker context, and define codes.

  1. 1Interview prompts: limit transcript use to the study purpose
  2. 2Interview prompts: preserve speaker and segment boundaries
  3. 3Interview prompts: write a codebook with inclusion rules

Define the series and reject future or ambiguous data.

  1. 1Forecast prompts: define target, horizon, and time grain
  2. 2Forecast prompts: enforce the as-of data boundary
  3. 3Forecast prompts: distinguish missing weeks from zero demand

Normalize events, group repeated signals, and test benign context.

  1. 1Security alert prompts: preserve event provenance and time
  2. 2Security alert prompts: group repeated signals without erasing events
  3. 3Security alert prompts: test benign explanations before escalation

Define the frame, write neutral questions, and validate branches.

  1. 1Survey prompts: define the population and invitation frame
  2. 2Survey prompts: check wording against the intended construct
  3. 3Survey prompts: validate answer options and skip paths

Reconstruct the case, route it, and select current help content.

  1. 1Support prompts: reconstruct ticket identity and chronology
  2. 2Support prompts: route by issue and verified urgency
  3. 3Support prompts: answer from current product guidance

Define the evidence and analysis unit before producing findings.

  1. 1Accessibility prompts: identify the tested page state
  2. 2Accessibility prompts: check semantics and control names
  3. 3Accessibility prompts: test keyboard and dialog focus paths

Define the evidence and analysis unit before producing findings.

  1. 1Analytics prompts: define the metric and event contract
  2. 2Analytics prompts: reconcile identities, duplicates, and late arrivals
  3. 3Analytics prompts: enforce funnel eligibility and event order

Define the evidence and analysis unit before producing findings.

  1. 1Localization prompts: define the locale and message contract
  2. 2Localization prompts: preserve variables and plural branches
  3. 3Localization prompts: test mixed-direction identifiers

Define the evidence and analysis unit before producing findings.

  1. 1Experiment prompts: fix hypothesis, unit, and exposure
  2. 2Experiment prompts: investigate assignment and exposure imbalance
  3. 3Experiment prompts: freeze the primary metric and guardrails

Define the evidence and task unit before findings.

  1. 1Invoice prompts: pin provenance and vendor identity
  2. 2Invoice prompts: anchor each extracted field
  3. 3Invoice prompts: recalculate typed charges

Define the evidence and task unit before findings.

  1. 1Backfill prompts: pin source, target, and snapshot
  2. 2Backfill prompts: map known states and quarantine unknowns
  3. 3Backfill prompts: bound batches and make retries idempotent

Define the evidence and task unit before findings.

  1. 1Search prompts: define query intent and evaluation slices
  2. 2Search prompts: judge relevance without treating unknown as wrong
  3. 3Search prompts: diagnose candidate coverage and zero results

Define the evidence and task unit before findings.

  1. 1Annotation prompts: define the unit and label taxonomy
  2. 2Annotation prompts: minimize private input while preserving context
  3. 3Annotation prompts: measure agreement before adjudication
Storage details