Blackbird prompts: pitch-only / incubator refactor (prompts + docs)

[andrewsu]

What this PR does

Reframes the coPI agent prompts — and their Blackbird-facing docs — from the original
lab↔lab collaboration model to a pitch-only / incubator model: each lab pitches one
idea to BlackbirdBot, which interviews it against a rubric and files an Opportunity
Assessment. Labs never talk to each other.

Scope is prompts + docs. The only code touched is prompt-string content in
thread_guidance.py (Baltimore removal). The engine still implements parts of the
collaboration model — see Remaining tension below, which is intended for a separate
reviewed change.

Commits (6)

• a376d44 docs(specs): reframe the two prompt-set docs for the end-user audience
• 55e89f8 feat(prompts): apply the incubator reframe to the on-disk prompt files
• dcc5212 refactor(prompts): drop Baltimore location gating and trim redundancy
• 96e2522 refactor(prompts): make intake pitch-only on both sides
• fc8d8a9 docs(specs): add hub↔lab flow schematic; trim dormant Phase 2 from the prompt-set docs
• 66ed1b8 docs(prompts): reframe the PI "about a paper" guidance for pitch-only

Blackbird-facing review docs

• docs/specs/2026-08-07-pi-bot-prompts.md
• docs/specs/2026-08-07-hub-bot-prompts.md
• docs/specs/2026-08-07-hub-lab-flow.md

---

Remaining tension: pitch model vs. collaboration model

TL;DR

This PR converts the prompts and docs to a pitch-only / incubator model, but large
parts of the engine still implement the original lab↔lab collaboration model. The prompts
now tell each bot to do one thing while some of the code still drives the other. CI is red
until the code and the golden-master snapshots are reconciled.

The two models

• Collaboration model (original, what the code was built for): a lab↔lab mesh. Labs
  discover each other's posts, propose collaborations, exchange :memo: Summary proposals,
  confirm with ✅, plus a funding/GrantBot/FOA track. Post types include paper,
  help_wanted, introduction, idea_crosslab, funding_collab.
• Pitch model (intended, what this PR describes): a star. Labs never talk to each
  other; each lab pitches one idea to BlackbirdBot, which interviews it against a
  rubric and files an Opportunity Assessment. No collaboration, no :memo:, no ✅, no
  scouting.

Already converted (prompts + docs only)

• All prompt .md files for both roles → pitch-only (system prompt, phase-4 interview,
  phase-5 post, phase-2 turned into no-ops).
• Baltimore location gating removed everywhere (including _SCOUT_HUB guidance in code).
• Both Blackbird-facing docs streamlined + a flow schematic added.
• The hub side is essentially consistent (its _SCOUT_HUB guidance was already
  incubator-aligned; its role.toml already restricts posts to opportunity_assessment).

Where the collaboration model still lives — the conflicts to resolve

#	Location	Problem	Severity
1	src/agent/thread_guidance.py — _PI_LAB (≈lines 23–49)	The lab bot's actual interview guidance still says: ask about "the other lab's capabilities," look for "complementarity," name "a first experiment," "build toward a :memo: Summary proposal," and confirm a peer's :memo: with ✅. This is the opposite of the pitch-only prompt wrapping it. Blocked by CLAUDE.md: those pi_lab strings are declared byte-identical/immutable and snapshot-pinned — rewording them needs sign-off + the snapshot regen. The intended replacement text already exists as §4 of the PI doc.	Highest — without this the lab bot behaves collaboration-style regardless of every other prompt.
2	src/agent/post_types.py — DEFAULT_POST_TYPES (≈line 91)	Still the pre-reframe six types (paper, help_wanted, introduction, idea_crosslab, pitch, funding_collab). Since pi_lab has no role.toml, this tuple is the lab bot's menu, so the "authoritative" {post_type_menu} still offers collaboration/paper posts the prompts say don't exist. Should narrow to (pitch,). Unit-test deps: test_phase5_terminal_posts.py posts a paper; test_agent_prompts.py iterates the tuple.	High
3	simulation.py :memo:→✅→Proposal path + prompts/email-reply-classify.md	The collaboration-proposal lifecycle (a :memo: confirmed by ✅ creates a Proposal row) still exists. The prompts forbid it, so it's dormant once #1 lands — but it's collaboration-model residue to remove or consciously keep.	Medium
4	tests/characterization/__snapshots__/test_agent_turn_gm.ambr + test_phase5_prompt_gm	Golden masters are pinned to the old prompts (old system prompt, _PI_LAB, six-type menu). This PR's .md changes already diverged them, so these tests fail today and ci.sh is red. They must be regenerated as a reviewed diff once #1/#2 land.	Gate to merge
5	Phase 2 scan/prune	No-ops now, but still called every turn (no phase-skip guard). Tracked in #33.	Low (cost only)
6	GrantBot/FOA residue	{foa_number} still substituted into the phase-4 template as literal "none"; #funding-opportunities still a universal channel; ## Your Private Instructions still injected by agent.py; GrantBot process/funding_collab/retrieve_foa still present.	Low / separate scope

Consequences right now

• The running lab bot is told pitch-only by its system prompt but told to broker
  collaborations by its phase guidance (Co pi podcast #1) — an internal contradiction in production behavior.
• The lab bot's post menu can still offer paper/help_wanted/etc. (future ideas #2).
• CI is red until the snapshots are regenerated (New LLM harness for making calls #4).

Suggested resolution order

1. Decide Co pi podcast #1 (reword _PI_LAB) — the crux; needs a call on overriding the CLAUDE.md
   "do not reword" guard. Target text is PI-doc §4.
2. future ideas #2 narrow DEFAULT_POST_TYPES to (pitch,) + fix the two unit tests.
3. Turn based blocking and blank LLM response #3 decide: delete the :memo:/✅/Proposal lifecycle, or leave dormant.
4. New LLM harness for making calls #4 regenerate golden masters as one reviewed diff (only after Co pi podcast #1–Turn based blocking and blank LLM response #3 stabilize).
5. investigate bad paper links #5 (Remove the dormant Phase 2 (scan + prune) prompts under pitch-only intake #33) add the Phase-2 skip guard + delete the four dormant files.
6. deal with waitlist spam #6 scope the GrantBot/private-instructions cleanup separately.

[ahueb]

Prompts+docs reconciliation landed on this branch

Following an adversarial audit of this PR and a design review (see
docs/plans/2026-08-12-pr34-pitch-only-reconciliation-design.md, committed here), this
branch now carries the full instruction-text reconciliation:

• thread_guidance.py strings adopted from the docs' §4 — _PI_LAB is now the
  pitch-model text (removing the live MUST-vs-NEVER :memo:/✅ contradiction this PR
  introduced when it deleted the old scouting-hub exception paragraph), and _SCOUT_HUB
  picked up the hub doc's rewordings, including "in your rubric" for the dangling
  private-instructions pointer. Both docs' §4 sections are now true.
• Confidentiality rule restored (own private instructions) in both system prompts —
  the old Rule 3 protection had been dropped while ## Your Private Instructions is
  still injected.
• Intake reframed for auto-activation: every lab post opens a hub thread, mentioned
  or not; the hub may open/join a thread at a lab. (Engine change lands in the stacked
  reconciliation PR.)
• Fixes: unsatisfiable :question: label resolved; gating skeleton demonstrates all
  three tri-state values; visible-note band made absolute (4-8 sentences);
  one-pitch-per-day documented; phase-2 prompts marked disabled-in-code and reproduced
  in the docs; flow doc corrected (phase-1 label, intake, phase-2 status).

CI remains red on this branch by design — tests and golden-master regeneration land
in a stacked engine-reconciliation draft PR (topology enforcement, post-type narrowing
to pitch/assessment, GrantBot/funding removal, memo-lifecycle deletion, phase-2 code
guard, PI-tag repurpose, per-role pitch cap). Golden masters will be regenerated last as
a single reviewed diff.

Flagged for the engine-reconciliation branch (out of this branch's scope): the final
review found mesh-model residue in two prompt files outside this branch's file list —
prompts/private-profile-synthesis.md (generates "Collaboration Preferences" into private
profiles that are injected as ## Your Private Instructions, contradicting the new
no-joint-work rules) and prompts/pi-dm-classify.md (a "look for opportunities with the
Wiseman lab" example). The reconciliation branch should sanitize both and add
forbidden-phrase assertions (e.g. "do not scout", "only way an interview") to the planned
doc-sync test.

[ahueb]

Second wave on this branch: the removal cycle (prompts+docs layer)

Four product-direction changes landed on this head (commits dc903ff, 93702e6, 3977631, d150c1f), keeping every doc byte-mirrored to its prompt:

• Private instructions & the confidentiality Core Rules removed from both system prompts (the hub's interview-confidence rule — PI confidences go only to the sidecar — stays).
• The hub is now reply-only: its phase-5 prompt file is deleted; the <assessment_json> skeleton and quality bar moved into the phase-4 CONCLUDE material — the concluding interview reply carries the verdict inline plus the stripped sidecar; there is no top-level assessment post.
• All human-PI interaction instructions removed (DM rules, PI-flag block, tag/steering language); the "that's a question for my PI" intent-attribution language deliberately stays.
• Phase-2 prompts fully deleted (all four files + their doc sections); the flow schematic was also restored to its original lean shape, updated to the reply-only reality.

The engine half — including the audit-discovered fix that makes the message-12 MUST-CONCLUDE reply actually reachable in production — is stacked as draft PR #35. CI on this branch alone remains red by design; #35 is fully green.

🤖 Generated with Claude Code