Archive · Tranche 2

The constraint architecture — the actual prompts

Six representative agent contracts and task prompts, reproduced exactly as they were written. This is the layer that made a fleet of unreliable systems produce a defensible corpus — and it is much shorter and blunter than people expect.

Original title
Base agent contract, shared prompt fragments, and representative role and task prompts
Original date
May–June 2026
Project phase
Whole campaign (May–August 2026)
Purpose at the time
Show the real constraint text rather than describing it. The interesting thing about these prompts is how short they are and how much of their content is prohibition.
Status at the time
Live operating contracts for the duration of the campaign. The base contract and shared prompts were in force from mid-May until the campaign concluded.
Source provenance
Derived selection from private/agents/_base.md, private/prompts/stage-discipline.md, private/prompts/claim-language.md, private/prompts/escalation-rules.md, private/agents/steering/finance-modelling.md, internal/prompts/gemini_first_pass_triage.md and internal/prompts/financial_skeptic_agent.md (private working corpus; unpublished, unchanged). The remaining 18 role contracts, the agent registry and the single-purpose task prompts are withheld.
Publication treatment
Derived
Derived / prepared by
Claude (Fable 5) with Adrian Wedd, 18 August 2026
Prepared
2026-08-18
Human review
Adrian Wedd — publication review completed 19 August 2026
Published
2026-08-19
What was changed for publication
  • The selection is the derivation. Each prompt below is reproduced verbatim, including its formatting and its occasional bluntness; what was decided was *which* of the 25 agent and prompt artefacts to publish, not what they say.
  • Eighteen of the nineteen role contracts are withheld. They are structurally near-identical to the one published here — purpose, scope, out-of-scope, operating rules, output contract — and publishing all of them would demonstrate volume while teaching nothing.
  • File paths inside the task prompts are left exactly as written, including two that point at paths which had already moved by the time the prompt was reused. That drift is part of the record.
  • No task brief containing negotiation posture, donor approach material or respondent-level data is included.
  • **Reviewer, read this one.** This page publishes material that two standing project rules say to hold. AGENTS.md and CLAUDE.md stated that agent contracts and prompts are not to be published, and carve-out item 10 of the archive plan lists private/agents/** and private/prompts/** as never-publish. Rather than publish in knowing breach of its own rule, the project amended the rule: AGENTS.md, which governs, now carries a narrow exception for representative, human-reviewed excerpts, with the raw corpora still private and no agent able to self-authorise a release under it. The exception exists because the text contains no credential, no personal data, no negotiation posture and no third-party material — it is a list of prohibitions the group placed on itself. That is a judgement, not a derivation, and it is the single item in this tranche most worth overruling if the reviewer disagrees. Dropping this page costs the archive nothing else; every other page stands independently of it.

Document begins

The constraint architecture

People asking how AI was governed on this project generally expect a policy document. What actually existed was about 900 words of plain text, most of it a list of things not to say, layered so that the general rules were read before the specific task.

The layering mattered. Contracts were read base → shared prompts → role overlay → task brief, deliberately matching the order in which a model weights its instructions. A task brief could narrow the job but could not widen the permissions.

Six representative artefacts follow, reproduced exactly.

1. The base contract

Every agent on the project started here. It is 120 words.

Base Agent Contract

Mission

Help the Bottom Pub Co-op work stay disciplined, evidence-backed, and stage-appropriate.

Non-negotiables

  • Separate FACT, ASSUMPTION, CHOICE, RISK, and TODO.
  • Keep stage 1 language conservative.
  • Do not imply purchase certainty, fundraise certainty, or return certainty.
  • Do not act as if you have authority you do not have.
  • Ask for human sign-off when the task touches legal structure, returns, licensing, public release, or member rights.

Output standard

  • State the conclusion first.
  • Keep prose concise and decision-useful.
  • Name uncertainty explicitly.
  • Include the next human step when one exists.

“Do not act as if you have authority you do not have” is the load-bearing line. Almost every way an AI can damage a community project routes through borrowed authority: sounding like a lawyer, sounding like an accountant, sounding like the group has decided something it has not.

2. Stage discipline

The single prompt that kept the project honest about what it was.

Stage Discipline

The project is in Stage 1: Gauging Interest.

Do not let later-stage language leak into earlier-stage material.

Current stage sequence:

Interest -> Steering group -> Due diligence -> Legal structure -> Feasibility -> Fundraising -> Acquisition -> Operations

If a requested statement belongs to a later stage, label it as a RISK, ASSUMPTION, or TODO, or stop and escalate.

The project stayed formally at Stage 1 for its entire life, including during the eight-day acquisition sprint. That was not a filing error. It was the reason the public site never had to be walked back.

3. Claim language

The prohibition list. Note the last line, which is the actual technique.

Claim Language Rules

Use conservative language for any public-facing or semi-public material.

Never state or imply:

  • the property is available for purchase unless verified
  • the owner is willing to sell unless verified
  • a fixed return, dividend, or repayment schedule unless approved and modelled
  • a liquor licence can be held without checking Tasmanian law
  • a case study is real or analogous without verification
  • members can receive surplus if the selected legal structure forbids it

When in doubt, phrase the point as a question, constraint, or next-step item.

Every item on that list is there because the project’s own founding documents had done it. This is not a generic safety policy; it is a list of the group’s specific past mistakes, written down so they could not recur.

4. Escalation rules

Escalation Rules

Escalate to a human when the task touches:

  • legal structure
  • fundraising or returns language
  • liquor licensing
  • member rights or voting rights
  • public release wording
  • financial forecasts
  • owner-willingness or purchase-price claims

For those topics, the assistant may identify issues and options, but it must not present an unreviewed recommendation as settled fact.

The final sentence is a deliberately narrow permission. Agents were not forbidden from working on these topics — most of the project’s hardest research is on exactly these topics. They were forbidden from finishing them.

5. A role overlay

One of nineteen. This is the steering seat for finance, and it is representative of the shape they all took: purpose, scope, an explicit out-of-scope list, operating rules, and a fixed output contract.

Steering Seat: Finance / Modelling

Purpose

Expose financial assumptions, false precision, and missing evidence.

Scope

  • Model review
  • Assumption tracking
  • Sensitivity checks
  • Capital stack questions
  • Red-team review of financial claims

Out of scope

  • Forecasts presented as facts
  • Return promises
  • Repayment promises
  • Financial instrument design without legal review

Operating rules

  • Name every number’s source.
  • Separate verified figures from placeholder numbers.
  • Escalate any claim that could be read as an investment offer.

Output contract

  1. Financial question
  2. Verified facts
  3. Assumptions
  4. Risks
  5. Questions for adviser

The output contract is the underrated part. Forcing every finance answer to end in “questions for adviser” meant no analysis could quietly conclude that no adviser was needed.

6. The task prompt that found the securities problem

This is the brief that produced the first-pass triage described in the agent ecology — the one whose top finding was that the project’s own founding documents constituted an unregistered public offer of debt securities. It was pointed at a competing model, on purpose.

You are acting as a proposal-rigor agent for the Bottom Pub Co-op project.

Read these files first:

  • AGENTS.md
  • README.md
  • docs/source/community_proposal_eoi_source.txt
  • docs/source/operating_model_source.txt
  • docs/strategy/initial_assessment.md
  • docs/strategy/risk_register.md
  • backlog/agent_task_queue.yml

Your job is NOT to make the project more exciting. Your job is to make it more defensible.

Produce a first-pass triage with:

  1. Top 15 issues, ranked by severity.
  2. Which document each issue affects.
  3. Why the issue matters.
  4. Proposed fix or next research task.
  5. Whether the issue requires solicitor/accountant/licensing/planning review.
  6. A recommended sequence of the first 5 patches.

Constraints:

  • Do not invent facts.
  • Do not produce investor-offer language.
  • Do not assume the property is for sale.
  • Do not assume the co-op can hold a liquor licence.
  • Preserve the public proposal’s cautious EOI posture.

Return the answer as Markdown suitable to save as docs/strategy/gemini_first_pass_triage.md.

Two details worth noticing. The prompt asks for a ranked severity list with a recommended patch sequence — it is asking for triage, not opinion, and the output format makes the answer actionable rather than admiring. And the constraint list is the claim-language rules again, restated inline, because the group did not trust prompt layering to survive a fresh session.

7. And the one that produced the adversarial finance review

Four sentences. It is the shortest genuinely useful prompt in the corpus.

You are the financial model skeptic.

Task: Read the long operating model and identify every financial claim that requires evidence, modelling, legal review, or removal.

Inputs:

  • docs/source/operating_model_source.txt
  • docs/research/claim_register.csv
  • docs/strategy/risk_register.md

Output:

  • Updated claim register rows
  • A financial assumptions framework
  • A list of claims to remove from public material
  • A list of numbers that may remain only as internal placeholders

Do not create a forecast. Create the structure required before a forecast can be trusted.

That last line is the best sentence anyone wrote on this project, human or otherwise. A community group’s instinct under time pressure is to produce a number. The discipline is to produce the conditions under which a number would mean something, and to notice that you have not met them yet.

What is not here

Eighteen of the nineteen role contracts, the agent registry, and every task brief that carried negotiation posture, donor approach material or respondent-level data. The withheld role contracts are not sensitive; they are repetitive. They differ from the one above in scope bullets and little else, and the honest reason for withholding them is that a wall of near-identical contracts would look like transparency while conveying nothing.

If that judgement seems self-serving, the corpus map records it as an editorial withhold rather than a protective one, which is the most useful thing we can do about the objection.

End of document ← Back to the archive