Designing Program Creation to Make Search Work
Search was the visible request. The deeper problem was incomplete upstream data. I redesigned creation around flexible structure, conditional complexity and AI suggestions that never remove human control.
Search quality had to be designed upstream.
A six-field form created records, but not the consistent context that discovery, filtering and later operations needed.
Consistency for search should not force specialised programs into the wrong category.
- 01
Hybrid taxonomyControlled terms, searchable multi-select and a reviewed custom path.
- 02
Conditional complexityReveal operational detail only when it becomes relevant.
- 03
Human-controlled AISuggest, explain and recover—never commit automatically.
The obvious request was search. The consequential decision was to redesign creation.
The old flow was fast because it captured little: a name, banner, broad focus area, description, visibility and a certificate checkbox. Audience, geography, ownership and certificate rules were handled outside the product.
Three findings changed the product direction.
Broad categories could not describe specialised programs
Direction: keep controlled terms for search, but support multiple selections and a governed custom value.
Audience and location were not single-value questions
Direction: support mixed professions and local, multi-state, national, global or online reach.
AI was useful as a starting point, not an authority
Direction: make suggestions contextual, editable, replaceable, regenerable and optional.
Research credibility: 14 sessions · English + Hindi · ~2h audio · AI-assisted transcription manually checked.
Free text was too inconsistent. A rigid taxonomy was too narrow. The answer was a hybrid.
Select all relevant focus areas
Reconstructed interaction model based on the documented design; confidential production screens are not shown.
AI could reduce blank-page effort only if the user stayed in control.
Suggest structure from visible program context.
Accept or edit.
Ask for context instead of hiding an inference.
Clarify or continue manually.
Offer alternatives and preserve entered context.
Regenerate, replace or skip.
Keep output as an unapproved draft.
Review and confirm.
Save existing work and keep the workflow open.
Complete the task manually.
A guided setup flow makes the new structure usable.
Define
Sector, focus, audience, location, objective and visibility.
Create identity
AI-assisted or manual name and banner paths.
Set rules
Topics, contacts and conditional certificate detail.
Review ownership
Leads, view/edit permissions and final confirmation.
Create the program identity
Start with a suggestion or continue manually.
Certificate rules appear only when needed
Eligibility, issuer and accreditation do not burden every program.
Visibility and team access stay separate
Discovery, joining, viewing and editing are treated as different decisions.
Users confirm the complete program before publishing
Generated and manual choices meet in one accountable final review.
What changed in the design—and what remains unverified.
A stronger foundation for discovery
The workflow defines richer, more consistent inputs for search, filtering, registration and future automation.
The search request moved upstream
I reframed discovery as a creation and data-model problem, bringing product, engineering and program-team constraints into the same decision.
No invented performance claim
Public post-launch metrics are unavailable, so intended benefits are not presented as measured impact.
After launch I would measure: completion and drop-off, metadata completeness, AI accept/edit/skip behaviour, and search zero-result or reformulation rate.
Two lessons I would carry forward.
What I learned: enterprise simplification is not always fewer steps; it is putting consequential decisions in the right order.
What I would validate next: whether people understand the taxonomy and AI provenance quickly, and whether the manual fallback feels equally complete.
Depth for the interview conversation.
Research method and evidence integrity
I synthesised spoken questions, hesitation, self-correction and direct feedback from English and Hindi audio. AI-assisted transcription accelerated processing, but I manually checked findings, excluded unclear speech and did not infer visual behaviour from audio-only evidence.
Supporting synthesis themes
Evidence clustered around search taxonomy, audience and location, AI trust and control, and certification operations. Repeated patterns were kept separate from one-off preferences so frequency did not become a substitute for severity.
Other decisions influenced by research
Certification became a conditional workflow with eligibility, issuer and accreditation detail. Visibility, joining, viewing and editing became separate questions instead of one public/private setting.
How AI supported the design process
AI helped with transcription, early pattern surfacing and content exploration. The recordings remained the source of truth; evidence, interpretation and design response were kept distinct.