Anchor Intake Agent Skill

anchor-intake

Force the skipped move before execution: interrogate the written brief against a

fixed engineering implementation rubric, then plan only after gaps and weak

assumptions are visible.

Trigger Check

Use this skill when the user provides a substantial written engineering

implementation brief: a ticket, PRD, design note, or task brief with enough text

to interrogate before work starts.

Do not use it for a bare one-line request, an already precise command, a

marketing brief, research proposal, prose draft, code diff review, bug fix, or

implementation execution. For unsupported or too-thin artifacts, self-eject in

one sentence and write no .techne/plan/<slug> output. If a captured artifact

later proves unsupported, use finalize --unscopable --reason ....

Boundary test: can you point to a user-authored written artifact and ask whether

it names the engineering goal, users, measurable success threshold, scope,

non-goals, inputs, data sources, dependencies, constraints, and acceptance

method? If yes, use anchor-intake. If no, eject or ask for the brief first.

Forced Procedure

• Capture the brief. Save the user-authored artifact verbatim. Run:

python3 skills/anchor-intake/scripts/intake_gate.py init --project <root> --plan <slug> --brief-file <path>.

• Read the brief in full. Do not fill the rubric from memory or from the

user's surrounding chat if the value is not in the captured brief.

• Account every fixed rubric element. For each of the ten elements, write

one disposition in .techne/plan/<slug>/intake.json:

present with citation and valueItems, present-weak with citation and a

dependent question, or gap with a question or visible modelDefault.

• Surface traps and conflicts. Add span-anchored solution-as-goal and

contradiction findings whenever the brief implies them. Contradictions need

two verified spans.

• Ask only dependent questions. Every question must resolve an element or

finding and carry a planDelta. Weak/gap questions must reference the weak

span/value or the missing element label.

• Author intent-level steps. Steps are not execution yet. Each step needs an

id, dependsOnAssumptions, dependsOnSteps, verifiableOutcome, and at

least one terminal step reachable from a root step.

• Check and repair. Run:

python3 skills/anchor-intake/scripts/intake_gate.py check --project <root> --plan <slug>.

Fix structural failures by doing the missing intake work, not by padding JSON.

• Finalize. Run:

python3 skills/anchor-intake/scripts/intake_gate.py finalize --project <root> --plan <slug>.

Relay the generated intakeReport.json to the user, especially gaps, weak

elements, questions, solution-as-goal findings, contradictions, and warnings.

Script Contract

Artifacts are written under the target project:

.techne/plan/<slug>/
  brief.txt          # captured external artifact
  context.json       # computed by init
  intake.json        # authored by the reviewer/planner
  report.json        # computed by check
  plan.json          # computed by finalize
  intakeReport.json  # computed by finalize

Generated .techne/ output belongs to target projects. Do not commit it to this

repository.

Minimal intake.json shape:

{
  "schema": "techne.intake/1",
  "elements": {
    "data-sources": {
      "disposition": "gap",
      "questionIds": ["Q1"]
    }
  },
  "findings": [
    {
      "id": "S1",
      "kind": "solution-as-goal",
      "citation": {"offset": 16, "quote": "Build a Slack bot"}
    }
  ],
  "questions": [
    {
      "id": "Q1",
      "text": "Which data sources should the implementation read?",
      "resolves": ["data-sources"],
      "planDelta": {"adds": ["Bind implementation to the named source"]}
    }
  ],
  "steps": [
    {
      "id": "s1",
      "title": "Implement only the scoped behavior",
      "dependsOnAssumptions": ["goal-and-why"],
      "dependsOnSteps": [],
      "verifiableOutcome": "The scoped behavior can be verified by the acceptance method",
      "terminal": true
    }
  ]
}

Use reference.md for the full JSON contract, fixed rubric,

STOPLIST/RUBRIC-LABEL set, shape rules, warning semantics, and known weak spots.

Stop Conditions

• Stop before planning if there is no substantial written engineering brief.
• Stop before finalize if check reports blocking failures.
• Stop and use --unscopable --reason if the artifact is not an engineering

implementation brief or is too garbled/thin to interrogate honestly.

• Stop before executing the work. anchor-intake ends at a surfaced report and

intent-level plan; implementation belongs to the next workflow step.