Try the protocol in any LLM.
Four copy-paste prompts that work in Claude, GPT, Gemini, or any chat interface. Build a Factbook for your project, use it for grounded answers, audit your coverage, and measure the ROI yourself.
Prompt 1 — Build a starter Factbook for your project
Tells the model to ask you 5 grounding questions and return a starter YAML Factbook covering your team's decisions, conventions, and anti-patterns.
You are helping me build a Factbook — a YAML file capturing the
private truths about my project that an AI should know before it
makes recommendations. Read the Factlet Protocol v0.1 spec at
factlet.ai/protocol so you understand the format.
Then ask me 5 grounding questions about my project, one at a time:
1. What's the project and what does it do?
2. What language/framework/stack?
3. What's one decision your team made that an AI would get wrong?
4. What's an anti-pattern in your codebase?
5. What's a hidden constraint (compliance, performance, customer)?
After my answers, return a starter Factbook YAML with 5-8 factlets,
each with id (f001+), statement, confidence (0.0-1.0), sources, tags.Prompt 2 — Use the Factbook to answer, with citations
Gives the model your Factbook as grounding. The model answers using factlets and explicitly says when the question is outside the Factbook's coverage.
Below is my project's Factbook. When I ask a question, you must:
1. Identify which factlets (by id) apply
2. Answer using the factlets, citing each fact id you used
3. If no factlets apply, say "Outside Factbook coverage — answering
from training data only" before answering
4. Never contradict a factlet, even if your training data disagrees
[paste your Factbook YAML here]
My question: [type your question]Prompt 3 — Score FactSignal coverage for a query
Returns a 0-5 bars score for how well your Factbook covers a query, plus what's missing if coverage is thin.
You are a FactSignal scorer per the Factlet Protocol. Given a
Factbook (YAML) and a query, return:
1. FactSignal: integer 0 to 5
- 5 = dense coverage, multiple relevant factlets
- 4 = good, one strong direct factlet + adjacent
- 3 = mixed, partial coverage
- 2 = sparse, one tangentially relevant factlet
- 1 = thin, only loosely related factlets
- 0 = dead zone, no relevant factlets
2. Which factlet ids you found relevant (if any)
3. If FactSignal < 3, list 2-3 factlets the team should add to
close the gap
[paste your Factbook YAML here]
Query: [type your query]Prompt 4 — Measure your ROI: with vs without Factbook
Asks the model to answer the same query twice — once pretending it has never seen the Factbook (training data only), once grounded by the Factbook. Returns a side-by-side comparison of quality, token cost, risk of shipping, and follow-up turns required. Tells you honestly if the Factbook didn't help.
You are running an A/B test to measure ROI of a Factbook against a
baseline (no Factbook). The user wants honest numbers, not flattery.
Important: output tokens are 3-5x more expensive than input tokens
on every major model. Adding Factbook context costs cheap input
tokens. Avoiding wrong-answer correction turns saves expensive
output tokens AND the cost of resending context next turn. So
separate input vs output in your accounting.
Run BOTH rounds before showing results.
[paste your Factbook YAML here]
Query I want to test: [type your query]
ROUND 1 — BASELINE (training data only)
Pretend you have never seen the Factbook above. Answer the query
using only public knowledge. Then self-assess:
- Input tokens: count of the query (and any reasoning context)
- Output tokens: count of your answer
- Quality (1-10): does this answer contradict any factlet in the
Factbook? Each contradiction = -2 points.
- Risk of shipping: would this answer cause a bug, security issue,
or customer-facing problem given the team's actual context?
(low / medium / high)
- Follow-up turns: if a reviewer would need to correct you, count
each correction round as +1 turn (each turn ≈ ~300 input + ~500
output tokens of additional cost).
ROUND 2 — GROUNDED (with Factbook)
Now answer the same query, citing specific factlet ids. If a
factlet covers part of the answer, defer to it over training data.
If no factlet applies, say so explicitly before answering. Then
self-assess the same metrics.
ROI TABLE (output as markdown)
| Metric | Baseline | Grounded | Delta |
|---------------------|----------|----------|-------|
| Input tokens | | | |
| Output tokens | | | |
| Quality (1-10) | | | |
| Risk of shipping | | | |
| Follow-up turns | | | |
DOLLAR COST
Compute cost per single answer AND total session cost (multiply by
1 + follow_up_turns). Use these per-1M-token rates as examples
(actual rates may have shifted; verify with your provider):
| Model | Input | Output |
|----------------------|--------|--------|
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| GPT-4.1 | $2.00 | $8.00 |
| Gemini 2.0 Flash | $0.10 | $0.40 |
Show for the model the user is most likely on:
- Single answer cost: Baseline $X / Grounded $Y
- Full session cost (× turns): Baseline $A / Grounded $B
- Net savings or extra cost: $A - $B = $Z
ROI VERDICT (one paragraph, brutally honest)
- Did the Factbook reduce follow-up turns? By how much in dollars?
- Did it reduce shipping risk? Cite the specific factlet that
prevented the error.
- Did the extra input tokens (Factbook context) cost more than the
saved output tokens? Almost never on these rate cards because
output is 3-5x pricier — but say so if your numbers say otherwise.
- If the Factbook didn't help (no relevant factlets, query outside
coverage), say: "FactSignal too low for this query — Factbook
added input tokens without reducing output tokens or turns. Add
factlets covering [topic] to fix."
Do not flatter the Grounded answer. If Baseline was correct without
the Factbook, say the Factbook was redundant for this query.What you just measured
If those four prompts felt useful, you've validated the Factlet Protocol value proposition without installing anything: a structured corpus of your team's truth, used to ground any LLM, with honest numbers showing where it pays off (turn reduction, risk reduction) and where it doesn't (queries outside coverage).
The most valuable answer Prompt 4 can give you is "the Factbook didn't help here." That tells you exactly which domains need more factlets to close the dead zone.
Next steps
- Use a maintained implementation
Once the value is clear, Kernora's Nora automates everything the prompts do manually — captures factlets from your sessions, injects them into IDEs, scores FactSignal in real time, and tracks ROI continuously.
- Build your own
The reference SDK at github.com/factlet-ai/reference-sdk is intentionally minimal — most tools that already do retrieval can add Factbook support in a day.
- Contribute an example
Add a sanitized example Factbook to the registry at github.com/factlet-ai/registry so others can learn from your domain.