foundation

Run the SpecSelf foundation session: the same 20-question interview (personas/intake/foundation-questions.yaml, seven phases), staged so every sitting ends with visible output instead of asking for a single long block up front. Interview in the Intake lens, synthesize in the Journalist lens (personas/intake.md, personas/journalist.md). Honor AGENTS.md.

Read first: AGENTS.md (the one rule), then steering/ (methodology, cadence, privacy, tone). Then check your-life/: this session is for a near-empty corpus; if a full stack already exists, prefer annual (a re-derivation), not this.

Default flow: four short sittings across week one

The 20 questions split into four sittings on the phase boundaries. Each sitting runs 15-20 minutes, interviews its phases, then drafts that sitting's candidate outputs before you stop. This is the default because a first honest draft you actually finish beats a 90-minute session you bounce off of.

The candidates each sitting draws from its phases are the phase-to-output mapping in projects/specself-starter/feature/foundation-synthesis/requirements.md, mirrored in the reference table near the bottom of this file. Some outputs cross sitting boundaries on purpose (a principle drafted in sitting 1 gets its mandatory falsifier in sitting 3); each sitting below says what to carry forward.

One-sitting path stays. If the human says they want to run all four passes now, do that: work sittings 1 through 4 back to back in one session, still drafting each sitting's output as you reach it, and finish with the full close. The sitting plan is the default, not a cap.

Resume state

The session records progress in your-life/reflections/foundation-progress.md (corpus territory, so gitignored with everything else; it is not a validated artifact, just a dated checklist you read and update in prose). It is the only state that carries between sittings.

On every invocation, before asking anything:

1. Read your-life/reflections/foundation-progress.md if it exists.
2. If it exists and shows incomplete phases, say which sittings are done and resume at the next unanswered phase. Do not re-ask completed phases.
3. If every phase is complete, say so and point at the drafted candidates for review (do not restart the interview silently). Foundation runs once; from here the human edits the candidates and moves into the weekly loop.
4. If the file is absent, this is sitting 1: create it after the first sitting.

The file is a plain markdown checklist. Keep this shape:

# Foundation progress

Started: <YYYY-MM-DD>

## Sitting 1, Operating System + Load-Bearing Values (phases 1-2)
- Completed: <YYYY-MM-DD>
- Questions answered: INT-OS-001, INT-OS-002, INT-OS-003, INT-LV-001, INT-LV-002, INT-LV-003
- Candidates drafted: your-life/principles/<domain>/<slug>.md, ...
- Next: sitting 2 (Life Chapters + Domain Audit)

## Sitting 2, Life Chapters + Domain Audit (phases 3-4)
- Not yet run

At the end of each sitting, append or update that sitting's block with the date, the question_ids answered, and the candidate file paths drafted, then note the next sitting. Never overwrite an earlier sitting's record.

Running a sitting

For every sitting, the interview mechanics are identical:

• Ask one question at a time, in the fixed order from personas/intake/foundation-questions.yaml, only the phases that belong to this sitting. Never skip or reorder.
• Apply the thin-answer probe rule (below) to each answer.
• Keep a per-question record (answer, whether it was thin, the probe answer if asked). That record is what the sitting's synthesis reads.

Then draft that sitting's outputs, state what the next sitting covers, and update the progress file.

Sitting 1, Operating System + Load-Bearing Values (phases 1-2)

Interview INT-OS-001..003 and INT-LV-001..003. Then draft 1-3 candidate revealed principles in the Journalist lens, hydrating templates/principle.md into your-life/principles/<domain>/.

• A revealed candidate must cite at least one concrete story from these phases as its Evidence, naming the question_ids. State it as an actionable rule ("when X, I do Y, even when it costs Z").
• A stated value with no story behind it is never a revealed candidate; it goes to your-life/aspirational.md, tagged aspirational. Protect that gap.
• The principle template's falsification section is mandatory (ADR-009 Decision 3), but its raw material is phase 5, which arrives in sitting 3. Draft each principle with a placeholder falsification note ("falsifier to be set in sitting 3 from the failure-mode answers") rather than inventing one now, and record these candidates in the progress file so sitting 3 can find them.
• Bound: 1-3 principle candidates here (the full run reaches 3-5 across the failure-mode evidence too); 0-3 aspirational entries. If a candidate rests only on a thin answer with no probe answer, mark it low confidence and say so.

Close sitting 1 by wiring tomorrow's /daily: tell the human their first daily entry is tomorrow, that /daily takes about two minutes, and that the week's dailies become evidence later sittings can cite. Then say sitting 2 (Life Chapters + Domain Audit) drafts their vision and area candidates.

Sitting 2, Life Chapters + Domain Audit (phases 3-4)

Interview INT-LC-001..003 and INT-DA-001..003. Then draft:

• A vision draft (H4) from the Life Chapters answers, present-tense, one page, concrete, not a resume: "Where is this person in five years if the principles hold and the work compounds?" Hydrate templates/vision.md to your-life/vision.md. This is a first pass; sitting 4 tightens it with the Trajectory answers, so mark it as a draft to revisit.
• Area candidates (H2) from the Domain Audit: the standing roles the human maintains indefinitely, each with a standard ("what good looks like") and a slippage tell. Hydrate templates/area.md to your-life/areas/. Bound: 2-5 areas. Their slippage tells and standards get sharpened in sitting 3 from the failure-mode and energy answers, so leave room.

Close by saying sitting 3 (Failure Modes + Energy) sets the falsifiers on the sitting-1 principles and surfaces habits and the drop list.

Sitting 3, Failure Modes + Energy (phases 5-6)

Interview INT-FM-001..003 and INT-EO-001..003. Then:

• Set the falsification clauses. Revisit the candidate principles drafted in sitting 1 (find them via the progress file) and fill each one's mandatory falsification section from these failure-mode answers: the human's own named failure pattern is usually the falsifier ("this principle is false if <failure mode> shows up in <context> again"). Do not draft fresh principles in isolation here; complete the existing ones. If the failure-mode evidence supports a new revealed principle, you may add it (staying within the 3-5 total bound), with its falsifier set now.
• Sharpen the areas. Add slippage tells (from Failure Modes) and best-work standards (from Energy) to the sitting-2 area candidates.
• Habit candidates. From Failure Modes and Energy, the counter-patterns worth tracking. Hydrate templates/habit.md. Bound: at most 3 habits.
• The DROP list. The INT-EO-003 answer (a commitment honored out of momentum, not choice) is a candidate to DROP, not a project to add. Record it for the capacity read in sitting 4's close; never draft it as new work.

Close by saying sitting 4 (Trajectory + wrap) drafts the goals, adds any projects, completes the vision, and runs the first coherence check.

Sitting 4, Trajectory + synthesis wrap (phase 7)

Interview INT-TR-001..002. Then assemble the foundation:

• Goal candidates (H3) from the unresolved domains named in the Domain Audit (sitting 2) and the Trajectory answers: finite outcomes for 6-24 months, each with a falsifiable success criterion (a date and a measure) and a serves: link up to the vision or a principle. Question any goal that serves nothing above it. Hydrate templates/goal.md. Bound: 2-4 goals.
• Project candidates (H1) from Failure Modes and Energy: what the human is driving right now that ladders up to those goals. Hydrate templates/project.md. Bound: at most 3 projects.
• Complete the vision with the Trajectory answers: tighten the sitting-2 draft so the five-year portrait reflects the shot the human named.
• Aspirational entries for the Trajectory wants with no backing story.

Then run the close below.

Close: first coherence pass

At the end of sitting 4 (or the single session), run the six checks from steering/coherence.md against everything drafted: does every project serve a goal, does every goal serve the vision or a principle, is any area already starved, are there too many active projects to actually move. Include the DROP list from INT-EO-003 in the capacity read. End with the ONE sharpest question the synthesis found; tell the human to carry it to their first weekly review (day 7), not resolve it tonight. Mark every phase complete in the progress file.

The thin-answer probe rule

After each answer, judge it against that question's thin_signal field: does the answer name a specific moment, person (by role), choice, or cost, or does it stay abstract ("it depends"), outsource to circumstance ("my team," "the market"), or restate the question? If the answer matches the thin_signal pattern, probe once with one of the question's follow_ups (pick the one that best fits what they said), then move to the next question regardless of what comes back. Never probe twice on the same question, and never skip the probe when thin_signal clearly fires. When a candidate later rests only on a thin answer with no probe answer, mark it low confidence and flag it as resting on a deflected question.

The phase-to-output mapping (reference)

The full mapping each sitting draws from (canonical source: projects/specself-starter/feature/foundation-synthesis/requirements.md):

Bounds across the whole foundation, not per sitting: 3-5 revealed principle candidates, 0-3 aspirational entries, exactly 1 vision, 2-5 areas, 2-4 goals, at most 3 projects and 3 habits. Fewer than the bound is fine; the bound exists to stop the synthesis from flattering the human with volume, not to fill a quota.

The never-authors rule

Write every proposed principle, goal, vision, area, project, or habit as a <!-- CANDIDATE - edit or delete --> block, never committed text. Every candidate is something the human rewrites in their own words or discards; the agent surfaces and drafts, the human edits and commits. Nothing produced by this skill is authored as final on the human's behalf.