-
-
Manage Client Discovery
Inspiration
Client discovery rarely starts from a clean brief. Important context is distributed across request documents, issue trackers, meeting notes, and prior decisions. Teams either repeat questions that are already answered or move into requirements before they understand the most consequential unknowns.
Manage Client Discovery was inspired by the need for a safer, more useful way to turn that scattered context into focused client conversations without allowing AI-generated assumptions to become project truth.
What it does
Manage Client Discovery is a Codex plugin that turns accessible project evidence into practical client-interview materials. The user chooses where the project information lives. The plugin reads only authorized sources, preserves evidence references, and uses GPT-5.6 to distinguish facts, stakeholder requests, hypotheses, decisions, and unknowns.
Before a meeting, it produces two separate deliverables: a lightweight request for existing artifacts that can be reviewed asynchronously, and a timed live-interview plan focused on experience, judgment, tensions, exceptions, and tradeoffs. This keeps the meeting from becoming a document-completion exercise.
After the interview, the plugin extracts proposed facts, decisions, unresolved questions, and tasks from the notes. It does not silently rewrite project knowledge. The user reviews the proposals first, and only confirmed learning becomes part of the saved project state. The next interview plan then builds on that confirmed learning instead of asking the same questions again.
The same Skill can be started without arguments at any stage. It inspects the saved workflow state, explains the detected next step, and asks for the next user decision before continuing.
How we built it
The project is an installable Codex plugin containing one state-driven Skill, focused workflow references, and Python standard-library scripts for deterministic state changes. GPT-5.6 performs evidence interpretation, uncertainty classification, gap discovery, and question prioritization. The Codex plugin orchestrates source access, artifact generation, reusable state, validation, and confirmation gates.
Codex also accelerated the design and implementation process. It helped convert the workflow into an explicit state machine, scaffold the plugin, implement regression and end-to-end tests, and run fresh-session demos. The finished project passes 11 automated tests, plugin validation, Skill validation, and a fresh-session end-to-end workflow.
Challenges we ran into
The main challenge was balancing useful autonomy with client-data safety. The plugin must resume intelligently, but it must not select sources, change authentication, publish externally, or treat interview notes as confirmed facts without user intent. Separating generated proposals from confirmed project state made that boundary explicit and testable.
Another challenge was improving interview quality rather than producing a longer checklist. The final design separates facts that belong in pre-meeting document review from questions that benefit from hearing a person's lived experience and judgment.
A fresh-session demo also revealed that an unselected visible file could be treated as a source. We corrected the workflow contract, added a regression test, and validated the fixed behavior in a new session.
Accomplishments that we're proud of
- A resumable, no-argument workflow that adapts to saved project state
- Explicit user selection of project sources
- Evidence records that preserve source locators and provenance
- Separate pre-interview and live-interview deliverables
- Confirmation-gated reflection and follow-up planning
- A complete synthetic demo loop and deterministic automated tests
- Fresh-session validation using a local brief and public GitHub Issues
What we learned
Good discovery support is not mainly about generating more questions. It is about deciding what should be learned asynchronously, what deserves live conversation, what evidence supports each assumption, and what the team is allowed to treat as confirmed knowledge.
We also learned that a conversational workflow needs deterministic boundaries. GPT-5.6 is valuable for interpreting evidence and prioritizing questions, while explicit state transitions and human confirmation make the resulting workflow understandable, repeatable, and safe.
What's next for Manage Client Discovery
Post-submission work will refine conversational boundaries, including clearer separation between source registration and source verification. Future integrations may package high-value source connectors when real usage demonstrates that a specific connection should be standardized.
Links
- Code repository: https://github.com/aym-sekiguchi/manage-client-discovery
- Public demo video: https://youtu.be/Hz6y_OIOuxE
Log in or sign up for Devpost to join the conversation.