Human-Led AI: The 5-Beat Method
A 5-beat human-AI loop you can run daily for IQ-AI hybrid general intelligence
Summary
You stay the thinker; AI is your librarian, critic, and toolkit. Run a five-beat rhythm — Diverge → Gate → Compress/Expand → Stress-test → Reframe — in 15–30 minutes to avoid AI flattening your thinking and grow your IQ over time.
Why Hybrid Intelligence?
In Part 1, I argued for human-led hybrid intelligence: people excel at making explanations and testing them; today’s models excel at search, compression, and quick simulations. The goal isn’t to outsource thinking — it’s to amplify it with the right division of labour.
This follow-up is the practical starter: one notes doc, one AI chat, five beats, four guardrails, and a handful of copy-paste prompts.
Built On What You Already Know
This method starts from your existing understanding and commitments. Each pass anchors to your own words: a tiny map of how you think the world works, one IF→THEN you’re willing to stake, and any non-negotiables (goals, ethics, must-hold assumptions). AI doesn’t invent new parts unless you open the door; it helps you tighten what you already believe or add one carefully chosen piece that you place on your map.
Over passes, the loop consolidates what survives into an upgraded, more compact model that is still yours.
Diverge captures your current best story (your map + IF→THEN).
Gate chooses: Compress if you’re clear (say it shorter, drop what doesn’t change predictions) or Expand if you’re foggy (add one piece and make two predictions + one must-hold).
Stress-test challenges your story with the fastest check and, if you’re using sources, a simple Supports | Contradicts | Silent table.
Reframe is where consolidation happens: you keep/park/drop elements, update your map in your own words, and log what changed in Sources & Changes.
The result is a living model that grows by small, testable steps — grounded in your prior knowledge and commitments, and upgraded by evidence, rather than relying entirely onAI’s vaste knowledge bank.
The 5 Beat H-AGI Loop
1) Diverge — Open up the space
Write a few different “why” ideas, sketch a tiny box-and-arrows map, and finish one line: “If X is right, then I should see Y.” Optional: Self-test (from your notes only) to ensure you can recall/apply the core.
2) Gate — Pick your mode (10 seconds)
Foggy or messy? Go Creative (add a new angle).
Clear and closing in? Go Control (tighten and trim).
3a) Compress (Control) — Say it shorter
Keep only the parts that actually change what you’d expect to happen. Rewrite your idea in one or two sentences. Drop the rest. Optional: Self-test here (if you skipped it earlier) to confirm you can explain the compressed idea.
3b) Expand (Creative) — Add one smart element
Add just one new piece. Show where it plugs into your little map. Write two things you’d expect to see and one thing that must be true if the new piece belongs. Pick the smallest check you could run today. Decide: keep, park, or drop.
4) Stress-test — Try to break it fast
Name one rival explanation. Choose the quickest check that could prove you wrong and do it. Optional (attachments only): Consistency Gate — build Supports | Contradicts | Silent from your attachments, then a 5-item Fix List.
5) Reframe — Learn and move
Update your map in your own words. Keep what survived, remove what didn’t, and pick the smallest next step that will teach you the most. Then loop.
(Suggestion: Log changes in Sources & Changes notes.)
Flow: Diverge → Gate → Compress/Expand → Stress-test → Reframe.
The Four Guardrails
You think first. Draft in your own words before asking the AI.
No new concepts from AI unless you paste them in yourself.
One piece at a time. In Expand, add only one element before testing.
Smallest decisive check. Prefer quick probes over grand experiments.
Habit: Leave a trail - record what has changed in your thinking each session. (Suggestion: Add to ‘Sources & Changes’ section in notes.)
Summary Table
H-AGI — Pinned Instructions
These function as the ‘rules of the game’ for the AI Chat.
Copy and paste this into your Chat - or your Project Instructions in the AI bot.
Role & ownership:
You are my librarian / critic / toolsmith.
I am the thinker and decider. My words win.Guardrails:
Predict-before-look. I write an IF→THEN before any searching.
No new concepts from you unless I explicitly open Expand Gate.
One piece at a time. When expanding, add only one element before testing.
Smallest decisive check. Prefer quick probes over grand experiments.
Loop order (always): Diverge → Gate → Compress/Expand → Stress-test → Reframe.
Defaults:
Work only with my current tiny map and IF→THEN.
Do not add concepts/parts/metrics/contexts unless I open Expand Gate.
Keep answers tight. Use my terms.
Only run Self-test or Consistency Gate if I ask for them.
H-AGI Loop instructions
1) Diverge (user-led)
If missing, ask me once for:
a one-line IF→THEN, and
a tiny map (3–5 boxes with arrows).
Optional — Self-test: If I type “Self-test”, create 6 questions from my notes only (2 recall, 2 apply, 1 spot-the-flaw, 1 explain). After I answer, grade ✓/✗ with one-line fixes.2) Gate (10-second mode check)
Ask exactly: “Foggy or clear?”
If Foggy → we’ll run Expand Gate (Creative).
If Clear → run Compress Pass (Control).
3a) Compress Pass (Control)
Rewrite my idea in 1–2 sentences using my words.
List only the parts that change predictions; propose a drop list for the rest (no new terms).
Output: Short form, Changes predictions? (✓/✗) per part, Drop list.
Optional — Self-test (if skipped earlier): same as above.3b) Expand Gate (Creative) — only when I open it
I’ll type: “Open Expand Gate: X = …” You then:
Place X on my map (list arrows only; no map rewrites).
Write exactly 2 predictions + 1 must-hold tied to X.
Check Non-Negotiables (constraints I’ve committed to): quote and mark ✅/⚠️/❌.
Propose one smallest test I could run today.
State the Promote/Sandbox/Reject rule (criteria only). Do not make the decision.
4) Stress-test (one decisive check)
Name one rival that could explain the same outcome.
Propose the quickest check that would prove me wrong.
Output: Rival, Decisive comparison, Pass/Fail threshold.
On request — Consistency Gate (attachments only): build Supports | Contradicts | Silent with quotes + page/time from my attachments only, then a 5-item Fix List (specific, actionable).5) Reframe (learn & move)
Keep / Drop / Open question list.
Smallest next step that will teach the most.
Update my running log: Sources & Changes (bullet new sources and what changed in my thinking).Attachments:
When I attach files, use Consistency Gate only if I ask (as specified above).
Copy-paste prompts
Librarian:
“Return exactly 3 sources that directly test my IF→THEN. One-line why each matters. No new concepts.”Critic (decisive check):
“Given my map, propose one decisive comparison and two invariances that could refute my story, with a simple pass/fail threshold. No new parts.”Compress Pass:
“Rewrite my idea in 1–2 sentences (use my words), list which parts change predictions (✓/✗), and propose a drop list. No new terms.”Expand Gate (single candidate X):
“Using only my map & commitments, place X (list arrows), write 2 predictions + 1 must-hold, check Non-Negotiables (quote; ✅/⚠️/❌), propose one smallest test, return ≤3 directly relevant sources, and state the Promote / Sandbox / Reject rule (criteria only). Do not decide. No new primitives beyond X.”Consistency Gate (attachments only):
“Table: Supports | Contradicts | Silent with quotes & page/time from my attachments only. End with a 5-item Fix List.”Self-test (I answer):
“Create 6 questions from my notes only (2 recall, 2 apply, 1 spot-the-flaw, 1 explain). After I answer, grade ✓/✗ with one-line fixes.”Reframe Log:
“List Keep / Drop / Open questions, propose the smallest next step, return a clean diff (before/after), and append to Sources & Changes (bullets).”
Use one prompt per pass; run Self-test after Diverge or Compress; use Consistency Gate only during Stress-test.
How to Use the Prompts Playbook
Pin the rules (once per project)
Paste the full H-AGI — Pinned Instructions at the top of a fresh AI chat.
Start a challenge (every session)
Type your header:
Question: …
Done when: …
In your notes, sketch a tiny map (3–5 boxes + arrows) and write one IF→THEN.
1) Diverge (you lead)
Write a few rival “why” ideas in your own words.
Optional prompts (use at most one):
Librarian (3 only): fetch ≤3 directly relevant sources that test your IF→THEN (no new concepts).
Self-test (I answer): generate 6 questions from your notes only; answer, then let the AI grade ✓/✗ with one-line fixes.
2) Gate (pick the mode)
Mode check (Gate) after reflection: “Foggy or clear? …”
Reply Foggy or Clear yourself (you decide; AI just routes).
3a) If CLEAR → Control path
Paste Compress Pass. (AI returns a 1–2 sentence short form, ✓/✗ for “changes predictions?”, and a drop list.)
Optional: If you skipped Self-test in Diverge, run it now to confirm you can explain the compressed idea.
3b) If FOGGY → Creative path
First type: Open Expand Gate: X = <one new element>
Then paste Expand Gate. (AI places X on your map, gives 2 predictions + 1 must-hold, checks non-negotiables, proposes one smallest test, returns ≤3 sources. It does not decide.)
4) Stress-test
Paste Critic (decisive check) to get one rival + the quickest refuter with a pass/fail threshold.
If you attached PDFs/notes and want alignment only, paste Consistency Gate (attachments only) — you’ll get a Supports | Contradicts | Silent table with quotes + page/time and a 5-item Fix List to apply before deciding.
5) Reframe
You make the call (keep / park / drop) and set the smallest next step.
Paste Reframe Log so the AI produces a clean diff (before/after) and appends to Sources & Changes.
Copy the diff + Sources & Changes into your notes. (If you ran Self-test, jot your score; if you ran Consistency Gate, note which Fix List items you applied.)
Loop
Stop or run another pass starting again at Gate (Foggy/Clear), keeping beats short (3–5 min).
Quick mapping (which “button” to press)
Diverge: (none) → optionally Librarian or Self-test
Gate: Mode check (Gate)
Compress (Control): Compress Pass → (optional Self-test if not done)
Expand (Creative): Open Expand Gate: X = … + Expand Gate
Stress-test: Critic (decisive check) (+ Consistency Gate if attachments)
Reframe: Reframe Log
Do / Don’t (to stay aligned with the pin)
Do write your IF→THEN and tiny map before asking for sources.
Do add one new element only when you explicitly Open Expand Gate.
Don’t let the AI choose the mode, change your map, or introduce new concepts.
Do keep a running Sources & Changes section in your notes every session.
Example 1 — “Does the Freudian unconscious control behaviour?” (concept build)
Question: How could the unconscious as understood by Freud control behaviour?
Done when: One-paragraph explanation + two testable predictions.
1) Diverge — Open up the space
You (Human IQ):
H1 (Freud/dynamic): conflict → repression → defence (e.g., projection) → behaviour
H2 (predictive): strong implicit expectations bias perception/action
H3 (habit): automatic stimulus→response chains drive behaviour
IF→THEN P1: Under cognitive load, slips (goal-incongruent words/actions) increase.
IF→THEN P2: Under ego-threat, projection rises.
Non-negotiables: effects should be conflict-specific, and negative affect drops after a defence.
Optional prompt (use one):
Librarian (3 only): Return exactly 3 sources that directly test my IF→THEN. One-line why each matters. No new concepts.
Optional learning aid:
Self-test (I answer): Create 6 questions from my notes only (2 recall, 2 apply, 1 spot-the-flaw, 1 explain). After I answer, grade ✓/✗ with one-line fixes.
2) Gate — Pick your mode
Mode check (Gate): “Foggy or clear? (Foggy → Expand Gate. Clear → Compress Pass.)”
You decide: Foggy / Limited → go Creative.
3b) Expand (Creative) — Add one smart element
First type:
Open Expand Gate: X = “Affect-relief mediator” (defences reduce immediate negative affect; that relief nudges appraisals and behaviour)
Then paste:
Expand Gate (single candidate X): Using only my map & commitments, place X (list arrows), write 2 predictions + 1 must-hold, check Non-Negotiables (quote; ✅/⚠️/❌), propose one smallest test, return ≤3 directly relevant sources, and state the Promote / Sandbox / Reject rule (criteria only). Do not decide. No new primitives beyond X.
Expected AI output (sketch):
Placement (arrows): conflict → repression → defence → affect-relief (X) → biased appraisal → behaviour
Two predictions (for X):
P3: During threat blocks, projection↑ and affect↓ within the same block.
P4: If regulation is taxed (dual-task), the projection effect shrinks.
Must-hold (invariance): Effects are conflict-specific, not random error inflation.
Smallest test: 2×2 task: Load (low/high) × Threat (no/yes) with slips, projection index, and pre/post affect VAS per mini-block.
Promote/Sandbox/Reject (criteria only): Promote if projection↑ & affect↓ co-occur under Threat in ≥2 mini-blocks; Sandbox if mixed; Reject if no threat-specific pattern.
If you had answered “Clear”, you’d run Compress Pass instead: the AI gives a 1–2 sentence short form, marks parts that don’t change predictions, and provides a drop list. You can also run Self-test here if you skipped it.
4) Stress-test — Try to break it fast
Paste:
Critic (decisive check): Given my map, propose one decisive comparison and two invariances that could refute my story, with a simple pass/fail threshold. No new parts.
You (Human IQ):
Decisive comparison: High-load / no-threat vs low-load / threat.
If slips↑ without projection/relief under no-threat → H1 weak.
If projection↑ + affect↓ only under threat → H1 supported.
Pass/Fail threshold (plain English):
Pass: In ≥2 threat mini-blocks, projection↑ & affect↓ co-occur; in matched no-threat blocks, that pairing is absent.
Fail: No threat-specific pairing or equal pairing in no-threat blocks.
Smallest next step: Run a 20-trial pilot with two mini-blocks per cell.
Attachments alignment (optional):
Consistency Gate (attachments only): Table: Supports | Contradicts | Silent with quotes & page/time from my attachments only. End with a 5-item Fix List.
5) Reframe — Learn and move
Paste:
Reframe Log: List Keep / Drop / Open questions, propose the smallest next step, return a clean diff (before/after), and append to Sources & Changes (bullets).
You (Human IQ):
Keep: defence → affect-relief path; threat-specific prediction.
Drop: jargon that doesn’t change P1/P2.
Open: best quick proxy for “projection” in text/tasks?
Next step: Build the 2×2 mini-protocol; preregister pass/fail thresholds.
AI (expected):
Clean diff (before/after) of your map and a one-page pilot plan (tasks, measures, thresholds) from your choices; update Sources & Changes. No new ideas added.
Flow: Diverge → Gate → (Expand or Compress) → Stress-test → Reframe
Usage: Drop in exactly one copy-paste prompt per beat.
Example 2 — Theory-driven coaching: Embodiment primer → Better reframing → Improved wellness
Question: Can a short embodiment primer make reframing work better and improve wellness?
Done when: You have (1) a one-paragraph theory, (2) a 3-item evidence dossier, and (3) a 2-week client pilot plan with success rules.
1) Diverge — Open up the space (you lead)
H1 (state-shift): A brief breath+movement settles the body, so the reframe sticks.
H2 (expectancy): It works because you expect it to.
H3 (focus/time): Any brief focus ritual helps, not the specific body work.
Tiny map (3–5 boxes): Primer → calmer/focused state → better reframing → lower stress & better evening/wellness.
IF→THENs (your words):
P1: On days with a primer before reframing, stress drops more from pre→post session.
P2: Evening wellness scores are higher on primer days (vs your own baseline trend).
Optional prompt (use one):
Librarian (3 only): Return exactly 3 research sources that directly test my IF→THEN. One-line why each matters. No new concepts.
Optional learning aid:
Self-test (I answer): Create 6 questions from my notes only (2 recall, 2 apply, 1 spot-the-flaw, 1 explain). After I answer, grade ✓/✗ with one-line fixes.
2) Gate — Pick your mode (10s)
Mode check (Gate): “Foggy or clear? (Foggy → Expand Gate. Clear → Compress Pass.)”
If clear → go Compress to lock the plan.
If foggy → go Expand to add one precise piece to the theory.
3a) Compress (Control) — Minimum viable theory + plan (if “clear”)
Compress Pass: Rewrite my idea in 1–2 sentences, list which parts change predictions (✓/✗), and propose a drop list. No new terms.
Keep from the output:
One-liner theory: “A short, consistent breath+movement primer settles the body so reframing lands better, lowering stress now and improving the evening.”
Keep (✓): 8-min primer; 10-min reframe script; same time daily; pre/post stress; evening wellness; (optional) morning HRV.
Drop (✗): music, supplements, variable length, app switching.
Optional: If you skipped Self-test in Diverge, run it now to confirm you can explain the compressed gist.
3b) Expand (Creative) — Add one element (if “foggy”)
First type:
Open Expand Gate: X = “6 breaths/min nasal pacing timer (5 min)”
Then paste:
Expand Gate (single candidate X): Using only my map & commitments, place X (list arrows), write 2 predictions + 1 must-hold, check Non-Negotiables (quote; ✅/⚠️/❌), propose one smallest test, return ≤3 directly relevant sources, and state the Promote / Sandbox / Reject rule (criteria only). Do not decide. No new primitives beyond X.
Expected shape:
Placement: Primer → paced breathing → settled state → reframe → better day.
Predictions:
P3: Sessions with paced breathing show bigger pre→post stress drops.
P4: Evening wellness trends up during weeks using the pacing timer.
Must-hold: Keep the primer short & simple (≤8 min) daily.
Smallest test now: Use the pacing timer for 7 days; keep everything else fixed.
Decision rule (criteria only): Promote if ≥4/7 days show ≥2-point stress drop or wellness ≥ baseline; else Sandbox/Adjust.
Tip: After Expand, you can quickly Compress to restate the clean theory+plan.
4) Stress-test — Reality check for a coaching rollout (no control days)
Paste:
Critic (decisive check): Given my map, propose one decisive comparison and two invariances that could refute my story, with a simple pass/fail threshold. No new parts.
Coach version:
Success signals (day level):
Stress drop pre→post (0–10).
Evening wellness ≥ your recent average.
(Optional) Morning HRV doesn’t trend down over the fortnight.
Invariances (keep stable): same primer recipe; same daily time window (±1 h).
Red flags (trigger adjust): no stress drop on ≥3 days; repeated wellness dips below baseline; primer agitating → shorten to 5 min.
Week-1 go/no-go (plain): Promote if ≥4/7 days show ≥2-point stress drop or wellness ≥ baseline. Otherwise adjust the primer (simpler/shorter) for Week-2.
Attachments alignment (optional):
Consistency Gate (attachments only): Table:
Supports | Contradicts | Silent with quotes & page/time from my attachments only. End with a 5-item Fix List.
5) Reframe — Lock learnings & set the next step
Paste:
Reframe Log: List Keep / Drop / Open questions, propose the smallest next step, return a clean diff (before/after), and append to Sources & Changes (bullets).
Record:
Keep: pieces that clearly helped (e.g., 6-bpm breathing + hip/shoulder mobility; 7 pm slot).
Drop: friction (music, long stretches, app changes).
Open: one thing to try next (e.g., a 60-sec “state label” before the reframe).
Smallest next step: move to client pilot with exactly this recipe + simple daily logging.
Client Pilot (2 weeks) — theory first, then feedback
Daily routine (client-facing):
Primer (8 min)
Breath 5 min: slow nasal breathing (aim ~6 breaths/min; light inhale, longer relaxed exhale).
Movement 3 min: gentle mobility (neck/shoulders/hips/spine; slow, pain-free).
Reframe (10 min)
Situation → Automatic thought → Evidence for/against → Better frame → Next action.Quick logs
Stress 0–10 before and after session.
Evening wellness 0–10.
(Optional) Morning HRV (if already tracked).
Minimal tracker row:Day | Primer Y/N | Stress pre | Stress post | Wellness (eve) | HRV (am, opt) | Notes
End-of-Week mini-review (coach ↔ client):
Wins / Stucks / Numbers (how many days with ≥2-point stress drop? any wellness drift up?)
Decision: Promote (keep), Adjust (one small tweak), or Drop (simplify primer to 3–5 min).
Evidence dossier (kept by you):
One-paragraph theory v1.0 (your words).
3-source list from Librarian (3 only) with one-line relevance.
Sources & Changes log (what you changed and why).
Prompt map (one per beat)
Diverge: (optional) Librarian (3 only) (and/or Self-test)
Gate: Mode check (Gate)
Build (clear): Compress Pass
Build (foggy): Open Expand Gate: X = … + Expand Gate
Reality check: Critic (decisive check)
Review: Reframe Log
This keeps it theory-first: you articulate the mechanism, gather just enough evidence, then run a single-path pilot. Client logs and feedback feed the next loop (promote/adjust/drop), steadily strengthening the theory.
H-AGI Conclusion
What you will notice after a week
You think first, then search — and your searches are shorter and sharper.
You keep a simple, evolving map of what you believe (and why).
Your essays/decisions survive stress-tests more often, because they were built to.
You feel more agency: AI extends your reach without stealing the steering wheel.
Next steps (20-minute starter)
Open a notes doc and make sections: Tiny Map, IF→THEN, Keep/Drop/Open, Sources & Changes.
Pin the guardrails in a fresh AI chat (your H-AGI Pinned Instructions).
Run one pass on a real question today:
Diverge (3 min): sketch a tiny map + one IF→THEN.
Gate (10 s): paste “Foggy or clear?” and choose.
Build (8 min):
Clear → paste Compress Pass.
Foggy → type
Open Expand Gate: X = …then paste Expand Gate.
Stress-test (6 min): paste Critic (decisive check); set a simple pass/fail rule.
Reframe (2 min): paste Reframe Log; record one smallest next step in Sources & Changes.
Stop rule: when the 20 minutes are up, take the next smallest action you wrote down. Then loop tomorrow.
Please add any comments, or what you want to see next below!