Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Stage 0 of the intervention lifecycle (venue-FREE). Answers 'why might this intervention work?' Documents the opportunity, expected impact, audience, channel hunch, mechanism hypothesis. Output: 0-lifecycle/0-seed/0-seed.md + _LOG_0-seed.md (+ 1-probes/ feasibility questions). Markdown only. Modeled on haipipe-paper-seed. Trigger: seed, opportunity, why this intervention, /haipipe-application seed.
argument-hint
[intervention-path] [intent...]
allowed-tools
Bash, Read, Write, Edit, Grep, Glob, Skill
metadata
{"version":"0.6.1","last_updated":"2026-07-19","summary":"Seed stage (stage 0, venue-FREE ROOT of the DIKW ladder) — states why this intervention might work before evidence is mature: opportunity, impact, audience, channel hunch, mechanism. DRAFT may WebSearch to orient; PROBE is FEASIBILITY-light (novelty + external-data obtainable); internal-data needs FORWARD to the ladder as [FORWARD -> CLAIMS] pointers. History: ./CHANGELOG.md."}
Skill: haipipe-application-seed
Stage 0 of the intervention lifecycle: the venue-FREE ROOT the whole DIKW ladder grows from.
It answers one question — why might this intervention work? — and keeps that possibility alive before the evidence is mature.
0-seed why this might work (the opportunity) <- THIS STAGE
1a-descriptions what the data looks like
1b-themes what patterns/topics emerge
1c-claims what generalizes (the ledger)
1d-advice what the evidence advises (the deliverable)
The user invokes this skill; it drives DRAFT → PROBE → REVISE → CHECK internally via the 2-phase/ workers.
Read first: ../../../PHILOSOPHY.md.
What's special: three things make a seed a seed
1. Venue-FREE, and the root every rung above depends on.
The channel hunch (sms / push / in-app / dashboard / email) is context, not a commitment — the venue is pinned AFTER the ladder via /haipipe-application venue, and the seed survives that retargeting untouched.
Seed writes [FORWARD -> CLAIMS] pointer lines in _LOG_0-seed.md for any internal-data need it surfaces (the token stays CLAIMS for grep-stability); rung 1a CONSUMES them at its DRAFT open, and an unconsumed pointer fails the 1a gate.
2. Its probes are FEASIBILITY only — "can this intervention exist at all?"
Two shapes, both light and both discovery-side: is the angle NOVEL (landscape / prior interventions / 查新), and is the EXTERNAL data OBTAINABLE (benchmarks, field norms, outside labeled data)?
Profiling OUR OWN cohort/engagement data is task work that belongs to rung 1a — never a seed probe; it leaves as a FORWARD pointer (need + why, no card).
This bounds the seed's cost to the feasibility question and stops it doing ladder evidence work early.
3. DRAFT may search; PROBE must dispatch.
Inline WebSearch is legitimate DRAFT fuel — orientation that becomes prose AND buffered state: planned question skeletons — but it is NEVER evidence.
The PROBE phase must ALWAYS run the real worker; an inline result has no project-side ledger, so the PROBE phase did not happen.
The invariant is section STATE: planned (DRAFT) vs read with a resolving target: (PROBE), mechanically enforced by the probe checker.
The four phases, in seed
DRAFT settle the five content sections with the user (haipipe-application-draft); MAY WebSearch to
ORIENT (crowded space? prior interventions? benchmark rates?) — weave the result into prose AND
AUTHOR the probe plan: ① ORGANIZE each feasibility question into a `## QX<n>` ENTRY in
1-probes/, ② MATCH it against the bank (a read-only grep is legal), leaving it `state: planned`;
register internal-data needs as [FORWARD -> CLAIMS] pointers in _LOG_0-seed.md
PROBE EXACTLY ONE worker call — `Skill("haipipe-application-probe", args="from-buffer <root>")` —
feasibility only (novelty + external-data obtainable). It RUNS THE DRAFT-AUTHORED PLAN FORWARD:
③ DISPATCH what the bank still owes, ④ POINT each `target`, ⑤ INTERPRET. ①② already happened at
DRAFT and are AUTHORITATIVE — PROBE does not re-raise and does not re-match.
Inline search is FORBIDDEN here.
Routing mechanics + seed specifics: ../../../2-phase/1-probe/haipipe-application-probe/SKILL.md
REVISE tighten wording; weave probe takeaways into Opportunity/Mechanism; the Q-consumer section holds the questions (haipipe-application-revise)
CHECK exit criteria below → Gate Ledger row; the gate RUNS the probe checker (haipipe-application-check)
Seed RECEIVES its feasibility evidence, never PRODUCES it inline: it raises questions; haipipe-application-probe binds them via the stake-free collector Agent(haipipe-probe-q-executor-agent), never an orchestrator directly.
If the intervention folder does not exist, route to /haipipe-application enter <path> (get-or-create owns scaffolding).
Migrate a legacy per-stage _PROBE/ card into 1-probes/ in the new shape on first touch only.
The artifact
0-lifecycle/0-seed/0-seed.md — full skeleton in ref/seed-template.md:
Opportunity 2-3 sentences: what gap exists, what behavior we want to change
Expected impact directional estimate ("increase refill adherence by 5-15pp")
Audience who receives it — a specific subset, not "everyone"
Channel hunch sms | push | in-app | dashboard | email — a HUNCH, not the venue pin
Mechanism hypothesis one sentence: why this audience + this content might respond
Q-consumer feasibility questions raised (novelty + external-data), one `## Q-Seed-<n>` block each
Each Q-consumer question is one ## Q-Seed-<n> block. Its fields and their FILL RULES live inline in ref/seed-template.md as <!-- RULE --> comments — that template is the single home, and this file does not restate them (JL ruling 2026-07-19, D3(b): a second copy is what drifts). The id carries the stage name (Seed-) so the flat probe pool shows which stage raised the question. State is deliberately minimal — a question is OPEN while Answer is unfilled, ANSWERED once it is not; no state: field lives in the seed doc (that machinery belongs to the 1-probes/ ENTRY).
Sidecar: _LOG_0-seed.md (phase journal + the [FORWARD -> CLAIMS] pointers).
Formatting: ===== title / ----- section rules; content sections use no #, Q-consumer questions use ## Q-Seed-<n>; one sentence per line.
Venue-FREE: the seed survives retargeting; the channel hunch is context, not a commitment.
Done: all five content sections carry real content (not placeholders); Audience and channel hunch are specific; the Q-consumer section raises at least the novelty/landscape question as a ## Q-Seed-<n> block (Description/Reason/Probe/Answer, per ref/seed-template.md), with internal-data needs appearing only as [FORWARD -> CLAIMS] pointers; the probe checker exits clean at the gate — the APPLICATION family's copy, named explicitly:
TWO files named check-probe-cards.sh exist on disk (paper + application) and their invariants differ, so a bare -name find resolves to whichever the filesystem returns first and silently asserts this intervention against PAPER rules.
Questions this stage typically raises
DRAFT's RAISE+PLAN step raises what the draft cannot answer. These are the kinds this stage is prone to — read this list, then walk the draft against it.
👣 occupied ground Has this intervention already been tried on this population?
Name the closest prior programme, or name the ground as open.
🧪 mechanism Is there evidence the hypothesised mechanism works AT ALL —
anywhere, for anyone? Not "is it plausible", but "who showed it".
📡 channel reach Does this channel actually reach this audience? Name the
coverage, not the intention.
This list bounds DISPATCH, never RAISING. A question of any other shape — profiling OUR OWN cohort/engagement data, or any other prerequisite — is still RAISED: it keeps its ## Q-Seed-<n> block, takes Answer: deferred -> CLAIMS, and carries a [FORWARD -> CLAIMS] pointer in _LOG_0-seed.md. What it does not get is an ENTRY in 1-probes/.
⚠️ The pointer token is [FORWARD -> CLAIMS], NOT DESCRIPTIONS, even though rung 1a-descriptions is what consumes it. The token is frozen for grep-stability (see "Venue-FREE" above, and 1a's consume-grep). Writing DESCRIPTIONS produces a pointer that 1a's grep never finds — the need silently disappears.
Exits
promote -> /haipipe-application ladder run the 1a-1d sweep (or `descriptions` to start rung-by-rung)
End every reply with the closing block (stage line via ../../../haipipe-application/stage-strip.sh).
