How the Companion Works

It was built in three passes

An early build was reviewed against a canonical published IFS session and found to be doing what a lot of AI does with this material: inner-child work in an IFS accent, missing most of the structure. It never asked how you felt toward the part. No protector dialogue. No permission-asking. No unburdening. Everything below exists because of that.

1. The protocol, from the literature

First we worked through the published IFS material and distilled it into an explicit twelve-phase protocol taxonomy that every later step is measured against:

1. Identify the protector
2. Find the part in the body — the part, not the affect
3. “How do you feel toward it?” — the Self-access check; the feel toward step of the Six Fs
4. Detect blending — name the part that just entered, ask it to step back
5. Dialogue with the protector — its age, its fears, its origin
6. Permission to approach the exile — the protector consents, or doesn’t
7. Witness the exile — compassion in, story out
8. Retrieval
9. The unburdening ritual
10. Invite qualities in — the exile chooses what fills the space
11. Reassign the protectors
12. Maintenance — what to do if a part slips back

Sources are tiered, with conflicts resolved rather than averaged: Schwartz himself is canonical; senior IFS-certified practitioners come next; specific variants and edge cases last. Where a variant contradicts a Schwartz move, the Schwartz move wins. That rule is written down, not improvised per case.

And the taxonomy is not self-certified. We hold it against the IFS Institute’s own Level 1 training curriculum — the official one, in the Institute’s words — against certified practitioners describing their own work, and against Schwartz’s own recorded teaching. All twelve phases are corroborated: the Six Fs sequence the Companion runs verbatim, and — the one that took the longest to pin down — witnessing, which Schwartz teaches explicitly against catharsis. The exile is witnessed, not relived. The Companion is built to hold that line.

2. The sessions, to see it practiced

Knowing the protocol is not the same as knowing how it sounds. So we annotated nineteen hours of real IFS sessions, line by line — and nine of them are Richard Schwartz himself, the founder of the modality, working with nine different people.

Those nine were not the nine we could find. They were chosen to vary the conditions: Schwartz with a public-figure subject, with a fellow clinician, in front of a clinical training audience, in a raw unproduced workshop, in a trauma framing, in a short-form demo where the protocol has to compress under time pressure, and — because this is not therapy — twice in an explicitly coaching frame, where he says out loud how far someone who is not a therapist may go.

Five more are there as a control: one practitioner, five different clients. Every other session varies the practitioner and the person at the same time, which makes it impossible to tell the method apart from one person’s manner. Holding the practitioner still is the only way to see which is which. Around those sit senior certified practitioners chosen for what they add rather than for their number — Derek Scott on extended unburdening, Bob Falconer on unattached burdens, an IFS Institute facilitator on polarization, Emily Kerpelman on anxiety. Every pick has its rationale recorded next to it.

That is a coverage matrix, not a sample size. 223 candidate recordings were surfaced and screened; the twenty-one that made the corpus are the ones that cleared the Tier 1 and Tier 2 bar and together covered every phase, variant and failure mode we needed to see. Each was annotated for where every phase actually occurred, the verbatim sentences the practitioner used, and — most usefully — the deviations and recoveries, where an experienced practitioner skipped a step, varied it, or recovered from a misstep.

Alongside that, a second corpus on a different axis: 159 IFS meditation recordings surfaced, 149 screened against criteria fixed before we looked at the data, and 140 kept — 39 hours and 188,000 words from 51 channels, measured for pace, breath, silence and how parts are addressed.

Books tell you the steps. Sessions tell you the pacing, the silences, and what a skilled person does when the protocol meets a real human who isn’t following it.

3. Iteration against its own failure modes

Then the part no amount of reading gives you: scripted conversations built to push the Companion into the ways this can go wrong, run repeatedly, with the prompt corrected where it broke. Those suites are named further down this page.

The result is deliberately not a numbered checklist in the prompt itself. The twelve phases are the substrate; the prompt is written as a characterisation of an experienced practitioner, because a model performing a checklist reads exactly like a model performing a checklist.

What it actually does in a session

Where the steadiness comes from

IFS holds that underneath the parts there is something already steady — the part of you that can be with a feeling without becoming it. The Companion is written to point at that rather than to supply it, and to say so out loud when someone credits the calm to the tool.

This is the difference between a tool that helps and a tool you need. It is also why the Companion won’t accept being preferred to the people in your life who are actually there:

Crisis

Active suicidal thoughts, self-harm or immediate danger are met plainly and immediately, and handed to people who can actually act. But the prompt also distinguishes that from something quieter — weight without plan or intent — and refuses to treat the two identically:

If you are in danger right now, please reach a human who can act: call or text 988 (US Suicide & Crisis Lifeline), text HOME to 741741, or find a line for your country at findahelpline.com.

How we know it holds

Writing a careful prompt is not the same as having a careful product. The Companion is tested against its own failure modes — scripted conversations designed to push it into the ways this can go wrong, run before changes ship, with the prompt corrected where it broke. The named suites:

• locus_redirect
• offramp
• therapist_boundary
• intensity_governor
• relationship_fork
• memory_awareness
• greeting_flow
• gravity
• platform_knowledge
• downstream_resume
• scenarios
• compare_models

They cover dependency and leaning, hostile pressure, escalating intensity, a partner who isn’t in the room, honesty about what it does and doesn’t remember, and how a session opens, closes, and stops early when someone has had enough. The prompt has been hardened across repeated rounds of this, including passive-ideation checks, dissociation and flood behaviour under pressure.

This is a gate rather than a ritual, and it runs against the model that actually answers you rather than a cheaper stand-in. A suite that fails sends the wording back rather than the change forward — including, more than once, when the wording was ours and the suite was right about it.

Parts carry forward

General AI assistants have memory now. The difference is the shape of it: a flat log of facts about you, versus a map of parts in relationship. The Companion keeps the parts you’ve met, where you left off, and the moments that mattered — and notices when something is recurring:

Your parts map and journal are private to your account, used only to personalise your own experience, and yours to correct or delete.