Single-Project AI Assistant
Turn one project's complete evidence base into a persistent assistant
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 99%
Create a persistent AI assistant for one project rather than a general assistant spanning unrelated work. Name the project after a concrete goal, establish one folder, and require contributors to place every relevant update, experiment, transcript, plan, and decision there. Load that evidence into an approved project feature such as Claude Projects, a custom GPT, or an equivalent system. The assistant can then retrieve experiments from a given month, compare impact, explain failures, expose dependencies, generate follow-ups, and maintain working tables. Because the boundary is singular and the evidence is comprehensive, users spend less time searching Slack, email, and project-management tools. Human review remains necessary for consequential conclusions and source attribution.
Origin
Extracted from Marketing Against The Grain when Kieran Flanagan described building project assistants for growth and other initiatives.
Core principles
- 01Attach each assistant to one clearly bounded project
- 02Centralize every relevant update and artifact
- 03Optimize for completeness before elaborate folder structure
- 04Ask the assistant to compare outcomes and explain failures
- 05Maintain the repository as work changes
How to run it
- 1
Bound the project
Define one concrete goal, scope, time horizon, and responsible group for the assistant.
Pro tip Use the goal itself as the project name.
Watch out Do not mix unrelated initiatives merely because they share contributors.
- 2
Create the repository
Establish a single folder or approved source location for all project evidence.
Pro tip Prioritize a simple contribution rule over a complicated taxonomy.
Watch out Scattered source material will recreate the retrieval problem.
- 3
Ingest complete context
Add updates, experiments, transcripts, plans, decisions, dependencies, and outcomes to the persistent AI project.
Pro tip Include failed experiments and superseded decisions with clear dates.
Watch out Incomplete context can produce confident but historically wrong answers.
- 4
Query operational history
Ask the assistant to retrieve events, compare experiments, explain outcomes, identify dependencies, and prepare follow-ups.
Pro tip Request links or references to the underlying artifacts.
Watch out Verify consequential summaries before acting.
- 5
Maintain live interfaces
Use the assistant to update tables, follow-up lists, experiment summaries, and status views as work progresses.
Pro tip Explicitly mark completed or removed items to prevent stale actions.
Watch out A model-generated interface may not write changes back to the source system.
- 6
Close the learning loop
Record final outcomes and lessons so future team members can understand what happened and why.
Pro tip Ask what knowledge would be missing for someone who did not participate.
Watch out Do not archive the project without preserving decision rationale.
In the wild
A team creates a project called increase activation rate by 50% in 2025. Contributors place experiment plans, January results, meeting transcripts, updates, and customer-flow documents in one folder. The assistant identifies the three January experiments, compares their impact, explains why two failed, and generates an updated follow-up table for the project lead.
→ The manager retrieves project history and learning without manually searching multiple communication systems.
Common mistakes
Creating a universal assistant
Combining many unrelated projects weakens retrieval precision and makes project-specific instructions conflict.
Uploading only formal deliverables
Missing updates, failed experiments, and meeting decisions prevents the assistant from reconstructing why outcomes occurred.
Assuming write-back
An assistant may update a view inside its console without synchronizing the underlying project file or system.
Is it for you?
Best for
Managers overseeing complex projects with many updates, dependencies, experiments, and contributors.
Not ideal for
Projects containing data that cannot be placed in the selected AI environment or whose scope is too broad to remain coherent.
From the transcript
“The key is to attach them to a singular project.”
“Everything goes in a folder. Then everything that folder goes into project assistant, it is pretty incredible.”
“If you're saying, like, tell me about the three experiments in January that we did, the one that basically led to the biggest impact and…”
From the episode
How a $1B+ Crypto Company Really Uses AI in Marketing