case-bootstrap

case-bootstrap

Overview

Scaffolds a new case folder, collects intake one missing fact at a time, initializes internal/state.json per the project convention, and routes to the first downstream skill. Designed to be the first thing a user (or another agent) invokes when starting a new case study.

When to Use This Skill

TRIGGER when:

• Starting a new business case study from scratch.
• The user says "I want to write a case about X" or "new case".
• Re-opening an existing case whose internal/state.json is missing or stale.

SKIP when:

• The case folder already has current intake and a populated internal/state.json (use the downstream skill in next_skill directly).
• The user wants to revise an existing case's metadata without re-scaffolding (do it manually).
• The task is research, drafting, or testing on a case that is already bootstrapped.

Workflow

At the start, read an existing internal/state.json and stop with a blocked history entry if an applicable blockers item prevents this run.

1. Intake, one fact: read what the author already said and any existing internal/intake.yaml. Store every unambiguous value. Follow references/intake-questions.md. When a blocking fact is still open, ask that one fact and stop the turn. Use the harness question tool for a real choice, and ordinary text for an open fact such as the company or the protagonist. Ask in the author's language. Do not list the other fields. When nothing blocks, continue in this same turn. Store teaching_tension and contrast_set when the author already states them, leave each null otherwise, and pass any stored value to research; they are not blocking questions (references/intake-questions.md, references/state-schema.md).
2. Slug: once the subject is known, derive a kebab-case case-<slug>/ from it. ASCII, no accents.
3. Scaffold: create case-<slug>/internal/ and case-<slug>/research/sources/public/ plus research/sources/private/. Do this only after the blocking questions are answered or explicitly deferred. Do not create placeholder documents. The full layout is in references/case-folder-convention.md.
4. State: check applicable blockers in an existing state, then write internal/state.json per references/state-schema.md, with stage: "bootstrap", work_target: null for a new case, and a bootstrap history entry. On rerun, preserve the existing successful stage, work_target, and history. Choose next_skill from the routing table below and actual remaining work. Leave uncollected audience, location, and journal as null.
5. Routing:
   • No clear idea yet → case-research (it will elicit the decision point)
   • Idea + interviews/transcripts ready → case-research
   • Everything written, heading to publication → case-peer-review (ask the journal first when it is unset)
   • Existing draft or case needing adaptation or anonymization, including before its first pilot → case-disguise
6. Idempotency: re-running on the same slug must not duplicate files. Update internal/state.json and internal/intake.yaml only when needed; leave case artifacts alone.
7. Report and continue: after scaffolding, print the case folder path and a one-line summary of the intake, say in one line that the next step starts, and run next_skill in this same turn, following references/pipeline-flow.md. Do not report a folder you have not created, and do not end the turn with this report.
8. Disclosure: append this run's inputs, outputs, AI contribution, human contribution, checks, and pending checks to internal/disclosure.md. Do not render that internal audit file.

Limits

• Does not collect case content. Does not run case-research. Does not analyze primary material.
• Does not edit AGENTS.md, INDICE.md, governanca/decisoes-sobre-skills.md, or any other global file.
• Does not push, merge, or publish. Pushes happen via the project's standard commit/push workflow.
• Does not create placeholder documents and does not render internal/. A later skill renders a document when it writes the Markdown. See references/reader-formats.md.

Continuing the pipeline

When this skill completes, follow references/pipeline-flow.md: tell the author in one line what comes next, and run next_skill in the same turn. Stop only for a reason listed there. End every turn with the harness question tool and options the author can pick, such as "Seguir com a pesquisa".

Inputs

• Whatever intake facts the author has already given. Missing facts are collected one at a time, per references/intake-questions.md.
• The existing case-<slug>/ if rerunning for idempotency.

Outputs

• case-<slug>/internal/state.json initialized with schema_version: 2, stage: "bootstrap", work_target: null, and one history entry.
• case-<slug>/internal/intake.yaml with the intake config.
• case-<slug>/internal/disclosure.md with the run's audit entry.
• case-<slug>/research/sources/public/ and research/sources/private/, ready for case-research.

References

• references/taxonomy.md — the terms the reader sees, per language: headings, captions, cross references, and cover labels ("Exhibit" is "Anexo" in Portuguese). Use it for every reader-facing document.
• references/pipeline-flow.md — continue into the next skill, when to stop, and the question that ends every turn.
• references/case-folder-convention.md — project-wide folder + state.json spec.
• references/intake-questions.md — canonical question wording and validation rules.
• references/state-schema.md — full internal/state.json schema reference.
• references/reader-formats.md — canonical Markdown, HTML and PDF copies, the paths to list, and the one main deliverable to offer to open. Downstream skills render; bootstrap does not render scaffold files.

Common Pitfalls

1. Re-running on a non-empty case-<slug>/ without checking existing state — overwrite risk.
2. Generating the slug with accents or uppercase letters — breaks downstream scripts and tooling.
3. Setting next_skill to null after bootstrap — leaves the pipeline stalled.
4. Forgetting to write the history entry for bootstrap — breaks the audit trail.
5. Sending the whole intake as one numbered list. Ask the first open blocking fact and stop the turn.
6. Ending the turn after scaffolding with a report and no question. Continue into next_skill; when something does stop the run, end with the harness question tool and a "continue" option.