To turn saved research into content ideas without hallucinating, ask AI to identify patterns and tensions inside a defined source collection, then require every proposed angle to name its supporting sources, counterevidence, and missing proof. Treat the output as an idea map, not a set of facts ready to publish.
The best ideas often come from relationships between sources: a repeated customer phrase, a contradiction between official advice and real behavior, or a mechanism that explains several isolated observations.
Begin with a research question
A large library produces generic ideas because the system has no reason to prefer one relationship over another. Choose a project and define the audience problem.
For example:
What prevents solo operators from reusing research across ChatGPT, Claude, and coding agents?
Add sources that can answer different parts of the question: user complaints, official product behavior, detailed workflows, counterexamples, and your own experience.
The project boundary prevents a content ideation prompt from pulling unrelated but semantically similar notes.
Ask for patterns with evidence
Use a structured ideation prompt:
Search the project and propose ten content angles. For each angle, state the core claim, target reader, supporting sources, strongest counterpoint, and the evidence still missing. Do not use a claim that cannot be connected to a saved source or clearly labeled as our interpretation.
This creates ideas that can be audited. A weak angle may sound exciting but reveal that it depends on one opinion. A quieter angle may have support across several independent sources.
Distinguish four kinds of statements
Label each proposed claim as one of these:
- Source fact: Directly supported by an identifiable source.
- Cross-source pattern: An observation created by comparing several sources.
- Creator interpretation: Your explanation or point of view.
- Open hypothesis: A claim that still needs evidence or testing.
Hallucination risk increases when these categories are blended. A pattern should not be presented as a measured fact, and an interpretation should not be attributed to a source that never made the claim.
Score ideas for usefulness, not novelty alone
Evaluate each angle across five questions:
- Is the audience problem real and specific?
- Is the claim meaningfully different from existing coverage?
- Can the project support the claim?
- Can you add firsthand experience or a demonstration?
- Does the idea lead naturally to a useful next step?
An original idea with no evidence may still become valuable after an experiment. Label it as a hypothesis and design the experiment instead of publishing certainty.
Build a claim-to-source outline
For the selected idea, create an outline where every major section includes the claim, evidence, interpretation, and desired reader takeaway.
| Section | Claim | Evidence | Interpretation | Gap |
|---|---|---|---|---|
| Problem | Research gets copied between AI chats | User examples and workflow observations | The project lacks a shared memory layer | Need a concrete before-and-after example |
| Mechanism | Retrieval reduces repeated setup | MCP architecture and project workflow | Context should be retrieved, not pasted | Need a measured time comparison |
| Action | Start with one bounded project | Firsthand product workflow | Smaller scope improves retrieval | Test with a real project |
This outline makes unsupported leaps visible before drafting.
Verify before publishing
Open every source used for a factual claim. Check the passage, date, authorship, and surrounding context. Trace repeated claims back to the original source when possible.
If the evidence changed, update the draft and the project. If the source cannot be verified, remove the claim or present it as an unconfirmed observation.
Save the result back to the project
After publishing, store the final piece, audience response, and any correction. Future ideation can distinguish what the research predicted from what actually resonated.
This creates a compounding loop: sources generate ideas, published work creates feedback, and feedback improves the next research question. AI accelerates the loop without becoming the authority behind every claim.