From Prompt to Research Partner
From Prompt to Research Partner
From Prompt to Research Partner
Multi-agent research skill
Multi-agent research skill
Multi-agent research skill
Project Overview
Project Overview
NTT DATA’s Launch design practice had an ongoing problem: every project started with one to two weeks of manual industry research, and none of that work carried forward to the next project or the next person. The practice had already started building an AI tool to fix this through an internal initiative, and I took over the next phase of it: rebuilding an early prompt into a structured multi-agent skill, testing it with the team, and iterating through four versions based on what that testing showed.
What started as a working but inconsistent prompt became a tool with defined modes, enforced checkpoints, and branded output the team could actually rely on.
NTT DATA’s Launch design practice had an ongoing problem: every project started with one to two weeks of manual industry research, and none of that work carried forward to the next project or the next person. The practice had already started building an AI tool to fix this through an internal initiative, and I took over the next phase of it: rebuilding an early prompt into a structured multi-agent skill, testing it with the team, and iterating through four versions based on what that testing showed.
What started as a working but inconsistent prompt became a tool with defined modes, enforced checkpoints, and branded output the team could actually rely on.
Where this started
Where this started
Context
Context
Launch runs design and strategy engagements across industries, healthcare, insurance, media, manufacturing, enterprise software. Every new engagement starts the same way: someone has to build enough industry context to be useful in the first client conversation. Reading reports, mapping competitors, identifying trends worth raising.
That work happens over and over, at uneven depth, and rarely gets reused. I had felt this firsthand on past projects and heard the same complaint across the practice, it was not tied to one bad client meeting, it was just how things worked.
The practice was already addressing it through an internal initiative: an AI Supported Design Research Accelerator, structured around 30-day sprints. The first sprint produced V6.2, a single long prompt that took a designer through the research process in one conversation.
Launch runs design and strategy engagements across industries, healthcare, insurance, media, manufacturing, enterprise software. Every new engagement starts the same way: someone has to build enough industry context to be useful in the first client conversation. Reading reports, mapping competitors, identifying trends worth raising.
That work happens over and over, at uneven depth, and rarely gets reused. I had felt this firsthand on past projects and heard the same complaint across the practice, it was not tied to one bad client meeting, it was just how things worked.
The practice was already addressing it through an internal initiative: an AI Supported Design Research Accelerator, structured around 30-day sprints. The first sprint produced V6.2, a single long prompt that took a designer through the research process in one conversation.
The Problem
The Problem
V6.2 worked, in the sense that it produced output, but a single long-form prompt has real limits once real people start using it.
V6.2 worked, in the sense that it produced output, but a single long-form prompt has real limits once real people start using it.
Tone drifted
Tone drifted
Without explicit constraints, the model’s writing style was inconsistent session to session.
Without explicit constraints, the model’s writing style was inconsistent session to session.
Uneven depth
Uneven depth
The same prompt produced thorough output for some inputs and shallow output for others, with no way to control for it.
The same prompt produced thorough output for some inputs and shallow output for others, with no way to control for it.
No defined output
No defined output
Getting from research to something shareable still meant manual formatting.
Getting from research to something shareable still meant manual formatting.
These were not bugs in a single prompt. They were the predictable result of asking one block of instructions to handle context-gathering, research, writing, and formatting all at once, with no structure forcing the model to slow down at the right moments.
What I actually owned
What I actually owned
My Role
My Role
I joined after V6.2 had already been tested with the practice. From V7 onward, the architecture and iteration were mine: I rebuilt the single prompt into a multi-agent skill, designed the second round of testing, sent the test survey to about ten people, synthesized the six responses I got back, and used those findings to drive the tool through to V10.
People moved in and out of the broader initiative throughout, since this is a team accelerator and not a solo project, but the day-to-day design and build work from V7 through V10 was mine.
I joined after V6.2 had already been tested with the practice. From V7 onward, the architecture and iteration were mine: I rebuilt the single prompt into a multi-agent skill, designed the second round of testing, sent the test survey to about ten people, synthesized the six responses I got back, and used those findings to drive the tool through to V10.
People moved in and out of the broader initiative throughout, since this is a team accelerator and not a solo project, but the day-to-day design and build work from V7 through V10 was mine.
Constraints
Constraints
Context window gravity
Context window gravity
In a single-file prompt, every instruction shares one context window. As a session gets longer, early instructions lose weight, the model pays less attention to rules stated at the top by the time it reaches the bottom. This is the core reason a long prompt cannot reliably hold complex behavior over a full session.
In a single-file prompt, every instruction shares one context window. As a session gets longer, early instructions lose weight, the model pays less attention to rules stated at the top by the time it reaches the bottom. This is the core reason a long prompt cannot reliably hold complex behavior over a full session.
Checkpoint control failure
Checkpoint control failure
Telling Claude to pause at a checkpoint and wait for confirmation is not the same as it actually pausing. Stronger wording did not fix this. It needed architecture, not better phrasing.
Telling Claude to pause at a checkpoint and wait for confirmation is not the same as it actually pausing. Stronger wording did not fix this. It needed architecture, not better phrasing.
No JavaScript output
No JavaScript output
The practice shares reports through SharePoint, which does not support JS-based rendering. Any interactive-feeling output had to be built as a single static HTML file.
The practice shares reports through SharePoint, which does not support JS-based rendering. Any interactive-feeling output had to be built as a single static HTML file.
Mixed AI fluency
Mixed AI fluency
Some people on the team use Claude daily. Others rarely touch it. Anything that depended on the user already knowing how to prompt well was a dead end.
Some people on the team use Claude daily. Others rarely touch it. Anything that depended on the user already knowing how to prompt well was a dead end.
How it took shape
How it took shape
Process and Key Decisions
Process and Key Decisions
Started treating it as a product
Started treating it as a product
The single biggest shift from V6.2 to V7 was structural. Instead of one long instruction set, I split the work into an Orchestrator and four named agents, Context, Landscape, Deep Dive, and a formatting stage, each responsible for one part of the process. The Orchestrator asks four questions up front (what you’re building, who it’s for, how much depth, what format) before any research starts. That single decision removed most of the ambiguity that caused inconsistent depth in V6.2.
The single biggest shift from V6.2 to V7 was structural. Instead of one long instruction set, I split the work into an Orchestrator and four named agents, Context, Landscape, Deep Dive, and a formatting stage, each responsible for one part of the process. The Orchestrator asks four questions up front (what you’re building, who it’s for, how much depth, what format) before any research starts. That single decision removed most of the ambiguity that caused inconsistent depth in V6.2.
Replaced style guidance with hard constraints
Replaced style guidance with hard constraints
Early versions told the model to write “like the WSJ.” That worked until something else competed with it, and then it got ignored. In V7, I replaced aspirational direction with an explicit anti-patterns list: no em dashes, no “it’s not X, it’s Y” constructions, no AI filler phrases. Naming the specific behavior to avoid did more work than describing the tone to aim for.
Early versions told the model to write “like the WSJ.” That worked until something else competed with it, and then it got ignored. In V7, I replaced aspirational direction with an explicit anti-patterns list: no em dashes, no “it’s not X, it’s Y” constructions, no AI filler phrases. Naming the specific behavior to avoid did more work than describing the tone to aim for.
Integrated gate enforcement into the architecture
Integrated gate enforcement into the architecture
Telling the model to “hard stop” at a checkpoint did not reliably work, even when I made the language stronger. What helped was physically separating steps across agents, so the model could not move to the next stage without a real handoff. This was a structural fix, not a phrasing fix.
Telling the model to “hard stop” at a checkpoint did not reliably work, even when I made the language stronger. What helped was physically separating steps across agents, so the model could not move to the next stage without a real handoff. This was a structural fix, not a phrasing fix.
Implemented a verification step before output
Implemented a verification step before output
A clean, branded report looks finished, which makes it easy to trust before anyone checks the facts in it. I added a step where key statistics and sources surface for the user to review before the final HTML report is generated, so the polish does not outrun the accuracy.
A clean, branded report looks finished, which makes it easy to trust before anyone checks the facts in it. I added a step where key statistics and sources surface for the user to review before the final HTML report is generated, so the polish does not outrun the accuracy.
Built guidance into the conversation
Built guidance into the conversation
Sprint 2 testing showed the same pattern every time: once someone opened Claude, they did not go back to a setup doc. So onboarding could not live outside the session. I moved everything a new user needed into the Orchestrator’s opening questions, where it could not be skipped.
Sprint 2 testing showed the same pattern every time: once someone opened Claude, they did not go back to a setup doc. So onboarding could not live outside the session. I moved everything a new user needed into the Orchestrator’s opening questions, where it could not be skipped.
Solution
Solution
By V10, the tool is a Claude skill with five working pieces:
By V10, the tool is a Claude skill with five working pieces:
• Orchestrator that routes the session and enforces format and checkpoint rules
• Orchestrator that routes the session and enforces format and checkpoint rules
• Three research modes, Quick Scan, Standard Brief, Deep Research, that adjust depth and citation detail to the occasion
• Three research modes, Quick Scan, Standard Brief, Deep Research, that adjust depth and citation detail to the occasion
• Two workflows, a Trend Report and a Client Primer, that share the same underlying agents
• Two workflows, a Trend Report and a Client Primer, that share the same underlying agents
• Branded, static HTML output that renders correctly in SharePoint with no JavaScript
• Branded, static HTML output that renders correctly in SharePoint with no JavaScript
• A context handoff block at the end of every report, summarizing client, industry, and trends covered, so the next time the user wants to start another session, they can pick up where the last one left off.
• A context handoff block at the end of every report, summarizing client, industry, and trends covered, so the next time the user wants to start another session, they can pick up where the last one left off.
None of this removes research judgment from the process. It gives every designer on the team, regardless of how much they already know about prompting, a structured way to get to a usable first draft of context in one session instead of a week.
None of this removes research judgment from the process. It gives every designer on the team, regardless of how much they already know about prompting, a structured way to get to a usable first draft of context in one session instead of a week.
What the testing showed
What the testing showed
Outcome
Outcome
I ran a second testing round with six designers across healthcare, insurance, media, manufacturing, and enterprise software. The pattern that came back was consistent: the output itself was strong, but the model’s process discipline was not yet where it needed to be.
I ran a second testing round with six designers across healthcare, insurance, media, manufacturing, and enterprise software. The pattern that came back was consistent: the output itself was strong, but the model’s process discipline was not yet where it needed to be.
4.5/5
4.5/5
4.5/5
Visual output
Highest-scoring dimension; testers called the branded HTML report polished and on-brand
4.0/5
4.0/5
4.0/5
Output quality
One tester said the insights matched conclusions their team reached after weeks of conversation
2.5/5
2.5/5
2.5/5
Gate compliance
Clearest weak spot; model still skipped required checkpoints in some sessions
4.5/5
4.5/5
4.5/5
Likelihood to reuse
Most testers said they would use it again on real work
I am treating gate compliance as unfinished, not fixed. The architecture changes in V9 and V10 improved it, but the testing data does not yet show it fully resolved, and I would say so plainly in any conversation about this project.
I am treating gate compliance as unfinished, not fixed. The architecture changes in V9 and V10 improved it, but the testing data does not yet show it fully resolved, and I would say so plainly in any conversation about this project.
Reflection
Reflection
Telling a model what to do is not the same as it doing it.
Telling a model what to do is not the same as it doing it.
I spent real time writing stronger instructions before I accepted that some behaviors need structural enforcement, not better phrasing. That is a different design skill than writing a good prompt, closer to designing a system than writing copy.
I spent real time writing stronger instructions before I accepted that some behaviors need structural enforcement, not better phrasing. That is a different design skill than writing a good prompt, closer to designing a system than writing copy.
A polished output format raises the bar for accuracy.
A polished output format raises the bar for accuracy.
Testers in industries they knew well caught inaccuracies. Testers in unfamiliar industries did not. A clean report builds trust fast, and that trust is only earned if the content underneath holds up.
Testers in industries they knew well caught inaccuracies. Testers in unfamiliar industries did not. A clean report builds trust fast, and that trust is only earned if the content underneath holds up.
This is not finished.
This is not finished.
Sprint 3 is the next step, and gate compliance is the open problem I would want to solve before calling this production-ready for the full fifty-person team.
Sprint 3 is the next step, and gate compliance is the open problem I would want to solve before calling this production-ready for the full fifty-person team.
Projects