Paid Chronicle · Public Preview
Reading Coding-Agent Prompts Without Mistaking Them for Architecture
Evaluate six entries from a historical coding-agent account, using a harness rubric that keeps missing source evidence separate from architectural inference.
Paid edition · $9 one-time · lifetime read access with email recovery · article only, no software license.
Read the sample, then see what is included and unlock- Public prompt snapshots do not expose every hidden vendor runtime or policy layer.
- The source corpus is unpinned; its provenance and a complete sampling manifest are not established here.
- Architectural implications are Greyforge inferences, not measured reliability results.
Prompt snapshots reveal harness shape, not the whole product
A public system prompt can reveal the identity assigned to a model, the tools the identified artifact names or describes, visible behavioral constraints, output rules, and some task-management structure. That is enough to compare parts of a harness. It is not enough to infer every hidden runtime service, evaluation loop, policy layer, or product behavior.
A bounded comparison would score only visible structure. Does the surface separate planning from execution? Can review be independent from the action loop? Is durable state named? Are tool permissions explicit? Is recovery a first-class phase or merely another instruction to the same context window? Answers would describe an identified snapshot; this archive lacks the pinned source manifest needed to reproduce those scores.
The original account inferred a recurring single-loop pattern from public prompt material. Its missing source manifest prevents a reproducible sample-level conclusion. Greyforge's interest in separate authority contracts remains a design position, not proof of complete vendor architectures or superior multi-role reliability.
Harness comparison rubric
- Identity
- What role and operating objective does the visible prompt declare?
- Tools
- Which read, write, terminal, browser, and external actions are exposed?
- Authority
- Are planning, implementation, review, and publication separated?
- State
- How does the visible harness carry task and durable context?
- Verification
- Can evidence block completion, or does the same loop judge itself?
- Recovery
- What forces a new diagnosis after repeated failure?
Included in this edition
Everything below is part of the article itself unless it says otherwise. No software, repository access, or product license is included.
Public source ledger
Checklist in the article
Retained publisher accounts, the missing source manifest, and dated official correction context.
Harness comparison rubric
Table in the article
Identity, tools, authority, state, review, and recovery fields.
Six-entry historical evidence ledger
Table in the article
Reported themes and unresolved capability boundaries, without vendor scores.
Authority-boundary map
Checklist in the article
A proposed division of planning, implementation, review and release duties.
- Edition
- Comparative edition 1.0
- Published
- Updated
- Reading time
- 8 minutes
- Full edition
- 1,925 words
- Article status
- Historical record
Who this is for
- Agent-tool builders comparing planning, execution, and review boundaries.
- Technical evaluators who need a reusable comparison rubric and explicit source gaps.
- Readers studying the difference between one-loop and separated-role harnesses.
Not for
- Readers seeking undisclosed vendor internals, a current product ranking, or a claim that prompt snapshots reveal every runtime behavior.
Detailed contents
- 01
Public corpus and selection boundary
- 02
A common harness rubric
- 03
Six attributed historical entries
- 04
The single-loop hypothesis
- 05
Separated roles and review
- 06
Six-entry historical evidence ledger
- 07
Reported themes and unresolved capabilities
- 08
Limitations and dated conclusions
Evidence and method
literature synthesis: Editorial review of six entries in the February publisher account. A reusable rubric is separated from unverified source descriptions and unmeasured architectural hypotheses.
Limitations
- Public prompt snapshots do not expose every hidden vendor runtime or policy layer.
- The source corpus is unpinned; its provenance and a complete sampling manifest are not established here.
- Architectural implications are Greyforge inferences, not measured reliability results.
Full research disclosure
Literature Synthesis
- Research question or engineering problem
- What can incomplete coding-agent prompt material establish about a harness?
- Principal finding
- The historical account describes six entries but lacks a pinned sampling manifest. Its architecture hypothesis is not a verified product comparison.
- Evidence type
- Historical publisher summaries, source-note limitations and separately dated official documentation.
- Method summary
- Separate reported themes, unresolved capabilities and proposed design criteria; do not infer product-wide absence from incomplete excerpts.
- Scope
- February 2026 publisher account, corrected in September; six attributed entries and later design aids, not a current ranking.
- Limitations
- Source provenance, exact files and the originally claimed seventh sample are unresolved.
- No controlled comparison establishes Greyforge superiority or reliability gains.
- Public source or reproduction note
- Editorial sources and limits
- Published
- 2026-02-22
- Last verified
- 2026-09-05
- Status
- Historical record
Access and updates
Purchase includes lifetime read access to this edition, email-based recovery, and revisions published to the same edition.
Public companion: Read the public source limits and correction