socratic-code-theory-recovery

Socratic Code-Theory Recovery

Reverse-engineer a bounded context into documentation without hallucinating the parts the code cannot tell you.

On invocation

When this skill is invoked:

1. Check whether the user named a bounded context. Look at the same message that invoked the skill and at the immediately preceding messages. A valid bounded-context pointer is a path (relative or absolute) to a directory, plus a short human-readable name for what the context is (e.g. src/auth, "Authentication"). If both are present, proceed to Phase 1.

2. If no bounded context is named, ask for it before doing anything else. Do not start Phase 1 against the current working directory by default — Phase 1 produces files (QUESTION_TREE.adoc, OPEN_QUESTIONS.adoc) and running it on the wrong directory wastes work. Ask exactly:

   Which bounded context should I apply Socratic Code-Theory Recovery to? Give me a directory path (the bounded context's code root) and a short human-readable name. If you want the whole current repo treated as one bounded context, say so explicitly.

3. Once you have the pointer, run Phase 1. Use prompts/phase-1-question-tree.md — substitute [bounded context path] with the user's path. Do not change the leaf classification, Q-ID scheme, or output files.

4. Stop after Phase 1. Phase 2 must wait for the team to answer the [OPEN] leaves in OPEN_QUESTIONS.adoc. Tell the user that Phase 1 is complete, where the two output files are, and what the next manual step is — do not proceed to Phase 2 in the same session unless the user explicitly asks.

When to use this skill

Use this skill on a brownfield codebase when:

• Documentation is missing, outdated, or untrusted.
• A change is about to be made and you need a spec before you can change safely.
• You want documentation that distinguishes code-derived facts from team-supplied context — auditable, not generated prose.
• You want to surface the open questions in the system, not just synthesize an answer the team has not seen.

Do not use this skill when:

• You are doing greenfield development — use the spec-driven workflow instead.
• The whole system needs to be documented at once — work one bounded context at a time.
• The code is not runnable — fix that first.

The contract

This skill implements the Socratic Code-Theory Recovery contract from the Semantic Anchors project. The methodology rests on Peter Naur's 1985 paper Programming as Theory Building: a program's theory lives in the heads of its developers and cannot be fully captured in code alone. A documentation-recovery process that ignores this produces confident-looking prose that fills in the gaps with invention.

The fix: model the gaps explicitly. Every question about the system is either [ANSWERED] from code (with file:line evidence) or [OPEN] (with a category and the role that must answer it). The OPEN leaves are the handoff to humans.

Two-phase workflow

                  ┌────────────────────────────────┐
   Phase 1        │  CODE  ──►  Question Tree       │
                  │             ├─ [ANSWERED] leaves│
                  │             └─ [OPEN] leaves    │
                  └────────────────┬───────────────┘
                                   ▼
                  ┌────────────────────────────────┐
   Between        │  OPEN_QUESTIONS.adoc           │
                  │  ──► team (routed by role)     │
                  │  ──► answers fill in OPENs     │
                  └────────────────┬───────────────┘
                                   ▼
                  ┌────────────────────────────────┐
   Phase 2        │  Answered tree ──► Docs         │
                  │  PRD · Cockburn UCs · arc42 ·   │
                  │  Nygard ADRs (claims cite code) │
                  └────────────────────────────────┘

Phase 1: Build the Question Tree

Use prompts/phase-1-question-tree.md. Adapt the bounded-context path and the Q1-Q5 wording; do not change the fixed second level, the leaf classification, the Q-ID scheme, or the output files.

Outputs:

• QUESTION_TREE.adoc — the full hierarchical reasoning trace
• OPEN_QUESTIONS.adoc — only the [OPEN] leaves, grouped by Ask role

The five root questions decompose into a fixed second level — the same enumerated node set on every run, so Q-IDs are stable and trees from different runs can be diffed node-by-node. Free, code-driven decomposition applies only below the fixed level. The fixed nodes:

• Q1.1–Q1.6 — product identity, primary users, channels, why-built, success metrics, segment priority.
• Q2.1–Q2.6 — actors, use-case catalog, per-interface system specs, data/entity model, acceptance criteria, cross-cutting business rules. See references/cockburn-use-cases.md.
• Q3.1–Q3.12 — the twelve arc42 chapters, in arc42 order. See references/arc42.md. Design rationale lives in the Q3.9 chapter — see references/nygard-adrs.md.
• Q4.1–Q4.8 — the eight ISO/IEC 25010 characteristics; Q4.9 — which characteristic has priority. See references/iso-25010.md.
• Q5.1–Q5.5 — technical debt, security risks, operational risks, dependency/supply-chain risks, scaling/performance risks.

Every fixed node is emitted even when its only leaf is [OPEN] or [ANSWERED: not applicable].

Leaf classification rules and Q-ID scheme: references/output-schema.md.

Worked examples — one [ANSWERED] and one [OPEN] leaf for each major branch: references/examples.md.

Between Phases: Team answers the OPEN leaves

Route OPEN_QUESTIONS.adoc to the people whose role appears in each section: Product Owner, Architect, Developer, Domain Expert, Operations. In one controlled experiment with a 13,000-line Go codebase, 11 targeted OPEN questions were enough to close the gap to the original documentation.

Team answers are written directly into OPEN_QUESTIONS.adoc under each question, marked clearly. Do not call Phase 2 until every OPEN leaf has either an answer or an explicit (deferred) marker.

Phase 2: Synthesize documentation

Use prompts/phase-2-synthesize.md. The Phase 2 LLM reads the answered tree and produces:

• PRD from the Q1 branch (problem, users, goals, success criteria)
• Specification from the Q2 branch (Cockburn use cases at User Goal level, system use cases for each technical interface, supplementary specifications)
• arc42 with all 12 chapters from the Q3 branch
• Nygard ADRs with Pugh Matrix from the Q3.9 branch

Code-derived claims cite the file:line evidence from their [ANSWERED] leaf — a reference to the code, the only durable, canonical artifact. Team-supplied information is marked (team answer). The Question Tree is temporary scaffolding, so its Q-IDs are not written into the final documents; during synthesis every claim is still traced back to a leaf as a build-time check. This dual traceability — code evidence plus team input — is the difference from a simple reverse-engineering prompt that fills in gaps silently.

What the LLM can and cannot recover

A controlled experiment (deleting documentation from a greenfield project and regenerating it from code) showed:

Derivable from code: functional requirements, acceptance criteria, building-block views, glossary, security mechanisms, crosscutting concepts.

NOT derivable from code: business context, design rationale (the ADR "why"), quality-goal priorities, stakeholder concerns, aspirational features, performance budgets, tutorials, review results.

If your synthesized documentation contains a claim from the second list without a (team answer) marker, the LLM hallucinated it. Mark it [OPEN] and ask the team.

Spec drift and reconciliation

After this skill produces documentation, the implementation LLM will add security hardening, validation rules, and edge cases that are not in the spec. This is structural, not a discipline problem. Re-run Phase 1 against the current code periodically — before a release, after a security review, before onboarding — and diff against the existing spec. The diff reveals NEW (in code, not in spec), CHANGED (diverged), and DEAD (in spec, not in code).

Further reading

• Peter Naur, Programming as Theory Building (1985). https://pages.cs.wisc.edu/~remzi/Naur.pdf
• Brownfield Workflow (Semantic Anchors). https://llm-coding.github.io/Semantic-Anchors/brownfield
• Brownfield Experiment Report. https://llm-coding.github.io/Semantic-Anchors/brownfield-experiment-report
• Fair Comparison Report (three recovery approaches). https://llm-coding.github.io/Semantic-Anchors/brownfield-fair-comparison