Long prompt, direct output
A one-pass weekly brief and its raw run record under the same source pack and report requirements.
Run packet in preparationThe policy was updated on June 31, 2026.
Switch the three typical synthetic cases below and compare the two ways to deliver the same material.
…
Reads smooth and confident — the error is hidden inside, with no warning.
…Synthetic cases to explain the workflow — not an automated fact-checking benchmark.
Prompts and skills tell the agent what it should do; BriefLoop records what it actually did, and decides which version is allowed to ship.
The real difference isn't one more reviewer agent — it's that the writing agent has no authority to declare its own draft “ready to deliver”.
Writes well, but the rules only live in this one conversation.
Can call tools and run multi-step, but the writer and the judge are still the same one.
Puts the key facts, checks, and approval outside the agent.
| ChatGPT + long prompt | Agent runtime + Skill.md | BriefLoop | |
|---|---|---|---|
| Main role | Generate content | Automate tasks | Manage a report from draft to delivery |
| How rules take effect | Agent interprets the prompt | Agent runs skill + runtime | Key boundaries by external state, program checks, and permissions |
| Who decides it's done | Usually the model or user | Usually the agent or runtime flow | The writing agent can't declare it deliverable alone |
| How key numbers are kept | In the material and the chat | Maybe in intermediate files | Key facts registered separately; changes must leave a trace |
| After an error is found | Keep chatting, regenerate | Agent retries or reruns steps | Issue logged, delivery blocked, rechecked after repair |
| Which version is final | The last reply or generated file | Whatever the agent marks done | The version that passed checks and was approved |
| After an interruption | Depends on the chat log and context | Depends on whether the runtime kept state | Recovered from artifacts and run records |
| Best for | One-off, low-risk writing | Multi-step automation | Recurring, high-stakes, multi-reviewer reports |
ChatGPT thinks and writes, the skill teaches it how, the runtime makes it run — BriefLoop keeps the ledger, blocks, and releases.
An agent runtime can implement all of this itself — persistent state, gates, human approval, recovery. But once it has, it's essentially rebuilt BriefLoop. The point isn't that others can't; it's that you don't have to reassemble it from prompts and scripts on every team.
A skill says “please do this,” BriefLoop says “you don't get past this until it's done.”
Local files, reports, spreadsheets, or registered web material.
Currently supported in Codex.
Whether numbers have a source, material is stale, or a big new conclusion suddenly appears.
Unresolved issues don't count as final. A human always releases, and it's recorded.
New runs in v0.15.3 use only the Experimental Codex SQLite path. The SQLite ControlStore (briefloop.db), its receipts, and ledger relations are the sole runtime authority. Agents write invocation-scratch proposals; only deterministic services may accept effects. The legacy JSON control plane has been deleted: no import, migration, or dual-write. Legacy JSON control files and report, status, Quality Panel, Markdown/JSON/JSONL/HTML exports are non-authoritative projections. Strict action, envelope, and human-request JSON payloads are only write-boundary carriers: they are revalidated against ControlStore and are not authority by themselves. See How it works.
See what each is for first; open the filename to see more.
The report that becomes final only after passing every check and human approval.
Which material and page each key number maps to, at a glance.
Where a source is still missing, a conflict remains, or a human still needs to judge.
Who released it, when, and on which version — traced throughout.
v0.15.3 can export one self-contained, read-only HTML view: Brief, Quality, Market Data, LAJ advice, and improvement status. It is a Store/LAJ display projection, not runtime authority, a quality score, or a delivery verdict. LAJ is Experimental; utility is NOT MEASURED.
Synthetic data · read-only display. The third page honestly reports Improvement Ledger unavailable; there is no disposition, write, or “accept suggestion to improve the next run” capability.
The same time window, source pack, model, and delivery requirements. Once the runs are complete and verified, this section will publish all three outputs, findings, and run packets — not cherry-pick the best-looking result.
A one-pass weekly brief and its raw run record under the same source pack and report requirements.
Run packet in preparationA weekly brief and run record produced by an agent using a public skill under the same inputs.
Run packet in preparationThe same writing agent, with fact registration, findings, repair state, and human release status attached.
Run packet in preparationThree of the most common recurring deliverables, each with its own easiest mistake.
Deliver: weekly industry moves and competitor tracking.
Often blocks: unverified data and opinions that creep in after many edits; freeze the core numbers before polishing.
Deliver: a monthly operating and decision brief for leadership.
Often blocks: “planned” written as “done”; overly absolute conclusions without attribution.
Deliver: source and compliance review conclusions on an AI draft.
Often blocks: stale or unsupported statements flagged and blocked; review only the exceptions.
BriefLoop won't automatically prove a sentence is absolutely correct, and it won't replace legal, compliance, or management judgment. What it does is put sources, conflicts, stale material, and open issues in front of you — and keep them from quietly becoming final.
For example: it won't prove “demand definitely hasn't recovered”; it only points out that the material provided isn't enough to support that conclusion.
These pages explain the v0.15.3 product boundary. They use synthetic data and do not establish measured capability.
The real briefloop init --web command is an Experimental one-shot local wizard. This site page is a read-only interface preview: it creates no workspace and submits no transaction.
Post-final advisory review. It never changes Gates, delivery, approval, next action, or factual-support conclusions. Utility is NOT MEASURED.
Open design preview →The HTML Improvement tab is a read-only unavailable state. There is no automatic learning. v0.15.3 Human observations and approved guidance live as Store Receipts; a successor run carries them only with explicit --include-approved-guidance.
You can see it catch errors without installing anything; to really run it, pick one below.
In Codex, say:
“I want to use briefloop.ai for weekly reports”
Before taking any action, read:
https://briefloop.ai/.well-known/briefloop-agent.json
Then read the canonical Skill specified by the manifest. Do not infer repository, installation, runtime, or capabilities from this homepage.
It should read the installation reference, confirm directory and permissions, and verify each setup step for your platform. A conversational “done” is not proof of installation.
git clone https://github.com/Stahl-G/briefloop.git
cd briefloop && bash scripts/setup.sh
Requires Python 3.12. After setup, source .venv/bin/activate, then you can run bash scripts/demo.sh. Create a workspace with briefloop new industry-weekly ./my-weekly, or run briefloop onboard then init --from-onboarding. On Windows, follow the PowerShell installation path; do not translate these Bash commands into Git Bash. pipx install briefloop is not the launch path yet.
Want to try online first? See the three catch demos. Full installation and governance details are in the agent installation guide and How it works. Current capability boundaries are in architecture-status. The on-site architecture reference v0.6.1 is a v0.12.1 code snapshot, not a v0.15.3 capability claim.