Prompt — repository review

The prompt used to assess a repository against its Foundation.

Rendered from intransic-skills/skills/intransic/prompts/repository-review.mdat v0.2.2 · 50e2c2f38f69

Prompt — Repository Review (full reassessment / new-agent onboarding)

Two uses: (a) a full periodic reassessment producing a new Foundation, (b) onboarding a new agent or engineer through the Foundation.


Part A — Full reassessment

Operating rules

  1. Evidence before assertion. Documentation is a claim to verify, never evidence. Compare implementation, executed tests, direct observation, configuration, schema, infrastructure — then documentation last. Where sources conflict, record the conflict rather than silently choosing.
  2. Do not invent maturity. A capability is implemented only when the operational path exists. An interface, mock, design doc, commented-out block, roadmap entry or dormant module is not sufficient.
  3. Separate public from private. Internal assets keep full technical detail; public assets expose no secrets, internal paths, exploitable detail or operational weaknesses.
  4. Do not modify production behaviour. Add documentation, schemas and validation assets only. No refactoring, dependency upgrades, renames or infrastructure changes.
  5. Trace every material claim to a file path, symbol, test name, configuration key, schema, route or commit.

Sequence

1 — Inspect only. Cover: root documentation, manifests, dependencies, source directories, API routes, schemas and migrations, domain models, authentication and authorisation, data access, validation, error handling, logging and observability, background jobs, infrastructure-as-code, deployment, CI/CD, test structure, environment configuration, secrets handling, privacy and security controls, external integrations, ADRs, feature flags, and TODO/FIXME/deprecated paths. Inspect enough git history to find major decisions, replaced approaches, abandoned components and decisions documentation no longer reflects.

Produce a repository summary, components discovered, likely-authoritative files, obvious contradictions, areas requiring caution, and analysis confidence. Create no files yet.

2 — Present the generation plan. Files to create, files to amend, assumptions, information excluded from public assets, likely Foundation sections and Trust Wall claims, validation approach. Wait for approval.

3 — Generate. Foundation, internal assets, Trust Wall, validation tooling.

4 — Validate. Schemas parse; YAML/JSON valid; evidence paths resolve; public claims map to evidence; no secrets in public files; internal and public maturity agree; scripts execute.

5 — Report honestly. Files created and modified, validation results, claims safe to publish, claims withheld, unresolved contradictions, highest-risk gaps, recommended next action, exact rerun commands.

Standing cautions

These hold in every repository:

Cautions specific to this repository

Not yet established — populate from the step 1 assessment.


Part B — New-agent onboarding

Explain the system through the Foundation

Read foundation/FOUNDATION.md first, then AGENTS.md. Understand, in order: what the system is and its lifecycle stage; its trust boundaries; its invariants and which are currently violated; the conventions; and what requires human review.

Correct extension points

To do this Go here
Add an endpoint …
Add business logic …
Change identity handling … — review required
Change schema …
Add background work …

Common architectural mistakes

Universal, and worth stating before the repository-specific list:

  1. Trusting the documentation. Verify against code first; check the contradiction register.
  2. Assuming tests protect you. Check whether an assertion exists before relying on one. If it does not, add it yourself.
  3. Assuming a capability is live because code exists. An interface, a mock, or an extensive but unregistered subsystem is not a capability.
  4. Removing a guard to make something work. If an authorisation attribute blocks you, the fix is a correct credential — not removing the attribute.

Repository-specific mistakes: populate from the contradiction and risk registers.

Evidence-backed guidance

When proposing a change, cite the file and line, the invariant it satisfies, the test asserting it, and whether the Foundation or ledger needs updating. Do not describe a change as safe because it "follows the existing pattern" — verify the pattern is the Foundation's, not merely a habit in a drifted area of the codebase.


Output (Part A)

Follow the five-step sequence, presenting results at each step. Do not create files before step 2 is approved. Prefer under-claiming to over-claiming at every step, and record what you could not determine rather than filling gaps with plausible inference.