Skip to main content
Corpus
workflow / six steps

Corpus

Structured agent work, checked at every step.

Define the work, run agents, verify outputs, keep the evidence.

Start the loop Read the docs

  • Plain markdown

    Readable in any repo.

  • Any agent

    Bring the tool you already use.

  • Human review

    Checks provide evidence, not verdicts.

loop preview

One run, six records.

ready
  1. 01Pull
  2. 02Spec
  3. 03Task
  4. 04Run
  5. 05Review
  6. 06Close
pwrcheckevidence

# current run

$ corpus review TASK-auth-refresh

PASS AC-001 — output pasted

UNVERIFIED AC-002 — manual resize pending

HUMAN ATTENTION — retry.ts changed outside scope

REVIEW-auth-refresh

review packet / example

review

AC-001 pass

Evidence: `npm test auth-refresh` pasted, exit 0.

AC-002 unverified

Evidence missing. Human attention required before merge.

No evidence, no Pass.

common failure modes

Make review visible.

A useful run leaves a trail: request, scope, checks, evidence, and findings.

  1. INTAKE

    Vague tickets

    Keep the request. Turn it into checkable requirements.

  2. SCOPE

    Agent drift

    Name the files, limits, and checks before the run starts.

  3. EVIDENCE

    Unbacked completion

    A Pass needs output, a CI link, or a named observation.

  4. LEDGER

    Lost findings

    Save useful lessons so later tasks can reuse them.

workflow / six steps

The loop is the product.

Pull, Spec, Task, Run, Review, Close. Each step writes or checks a record.

010203040506

workflow / six steps

The loop at a glance.

Six steps around the hexagon; the inscribed triangle is the spine — Spec, Run, Close. Each pass leaves a file the next step can use.

  1. 01
    Pulloptional

    Point the spec's sources at the origin, or capture the raw request verbatim in an intake file first.

  2. 02
    Spec

    State the requirements one per id, each with the check that verifies it. The spec is the unit.

  3. 03
    Taskoptional

    When one spec splits into parallel slices, hand each agent a bounded packet. Most work needs none.

  4. 04
    Run

    The worker implements the spec, runs every check, and pastes the real output beside each requirement.

  5. 05
    Reviewoptional

    A non-implementer checks the evidence per requirement; the full review packet is reserved for substantial changes.

  6. 06
    Close

    Merge, record the decision, save any durable finding, and update the board. Close feeds the next pull.

Close feeds the next Pull — the loop closes

review packet / example

Put evidence next to the claim.

Specs and reviews are working records. Keep commands, notes, and results in the same place.

SPEC-auth-refresh

spec example / acceptance criterion

spec

AC-003 — Expired refresh token redirects to login

The client must clear local session state and route the user to `/login`.

Verify with: auth-refresh-expired.test

what you get

The pieces.

spec

Spec-first

Write the contract. Hand the agent a bounded task.

review

Review by exception

Show evidence per requirement. Escalate the gaps.

branch

Worktree discipline

One task, one branch, one diff to inspect.

limits

Honesty framework

Mark what is convention, checklist, toolable, or enforced.

plain answers

Plain answers.

Is Corpus an agent?

No. Your coding tool writes code. Corpus structures the work around it.

Does Corpus decide whether code ships?

No. Checks produce evidence. Review decides what it means.

Do I need the CLI?

No. Corpus works as markdown. The CLI handles setup and checks.

Why the seal?

The six points are the six steps: Pull, Spec, Task, Run, Review, Close.

start / first pass

Start with one spec.

Copy the kit, write one contract, run one bounded task, review the evidence.

Start the loop Copy the kit