Super Agent Presentation

An agent-native reporting framework: scenario-adaptive report protocols, bounded context routing, checkpointed long-task memory, and mechanical structural audits — so the same task handed to an agent stops producing a different, unreadable report every time.

中文版 README | English

💡 Use it as an installed Skill in Claude Code / Codex / Cursor / GitHub Copilot, or hand any agent one Markdown index in link-only mode — no framework, no daemon, no lock-in.

🤖 AI agents: read AGENT_START.md instead — the smallest bootstrap, structured for LLM consumption, not human browsing.

🪶 Radically lightweight. The entire skill layer is plain Markdown plus one stdlib-only Python CLI. No database to maintain, no Docker to configure, no dependencies to install — fork it, rewrite it, adapt it to your stack.

🧾 Hand your agent the same task twice — get the same readable report shape both times. Route → bounded bundle → checkpoint → mechanical audit: the reporting contract survives hours of work and context compaction, and a bad delivery can be blocked by exit code instead of by hope.

Contents

1. Why — more than a style prompt
2. What's New · dated changelog
3. What you get · modes / modules / profiles / templates / CLI
4. Quick start · no-install trial + install + agent workflow
5. Real examples · 16 receipt-backed reports + live deck
6. How it survives long tasks · checkpointed anti-forgetting
7. Catalog
8. Deployment and constraint ladder
9. Evidence status · what is and is not claimed
10. Repository structure
11. Verification
12. Design docs
13. Contributing

---

1. 🎯 Why — more than a style prompt

Give an agent the same substantive task twice and you get two differently shaped reports; some are unreadable for humans. Style guides alone do not survive long tasks: after hours of work and context compaction, the agent has forgotten the reporting contract, and a giant "always follow this template" prompt taxes every turn and flattens short answers and deep experiment reports into the same shape.

This framework treats reporting as a routed, checkpointed, auditable protocol instead of a style preference. Whatever the wording, a reader should reliably find:

answer or status → evidence → explanation → boundaries or uncertainty → useful next step

Short answers may stay one sentence. Experiment reports must carry protocol, metrics, run counts, uncertainty, tables, and conclusion boundaries. An explicit user-requested format (JSON, three sentences, a paper section) always wins.

💭 Why routing instead of one big prompt? A resident mega-template taxes every turn and dies at the first compaction. Here a tiny micro-contract only decides when to activate; the full protocol is loaded once, at the reporting boundary, and the checkpoint file — not the context window — carries the contract across compaction.

💭 Why mechanical audits instead of "please follow the style"? Prompts reduce forgetting; they cannot block a bad delivery. audit --strict returns a real exit code you can wire into CI, so structure is enforceable — while the repository stays honest that no audit can verify your facts. Every finding the audit can emit, with its trigger and fix, is listed in docs/AUDIT-CODES.md.

2. 📢 What's New

• 2026-09-12 — ✍️ Human-form contract and cross-model editing (reportctl edit-prompt). Measured eight machine-written forms in the same-facts A/B and found half of them induced by the protocol itself: a gpt-6-astra draft narrated eleven inferences it declined to make (bare prompt: three); a gpt-5.5 draft wrapped 77 numbers in code spans because the protocol's own Markdown does. The core contract is rewritten around the reader's next decision (omit what the reader knows, state what is known and stop, give one position and the condition that flips it), bundle strips the protocol's word-level code spans, and three audits catch what survives (non-inference-statement, code-span-number, process-leakage). New edit-prompt hands the draft to a second model with a fidelity contract and the eight-item cut list. Re-run: narrated non-inferences 11 to 1, code spans 77 to 0, report 25% shorter; the edited draft kept every number and audited clean. Default bundle budget 16,000 to 20,000 after a mode was found silently over the old limit. Same day: finished exemplars (reportctl exemplar MODE) for status-update, experiment-report, decision-brief, and research-idea — a passage in the target register, EN and CN, read before the first sentence; rules describe the form, the exemplar shows it. Each passage is audited under its own mode in the test suite.
• 2026-09-07 — 🔍 Cross-model semantic review (reportctl review-prompt). The audit checks form; this step checks meaning: claim-versus-visual consistency, derived-number premises, reasoning validity, unsurfaced source contradictions, fidelity to the fact sheet. Fed three pre-fix showcase drafts that had passed a zero-warning audit, a second model flagged all three human-found reasoning errors as blockers and found two more the human had missed. Also fixed two regressions the framework caused on chat surfaces, found by a same-facts same-model A/B against a bare prompt: no section markers and 1.7x hedging (under-sectioned, hedge-saturation warnings; delta columns in baseline tables). Honest result of that A/B: for a strong model on a one-shot short report, bare and framework are close to a draw; the framework helped gpt-5.5 clearly more.
• 2026-09-06 — 🧭 Natural-language routing, audit-codes reference, audit boundary fixes. A probe matrix of natural requests that avoid the literal signal words routed only 12 of 25 correctly (a request containing the word postmortem lost a tie to investigation-report; incident, papers, risks scored nothing because only compound phrases were registered; , , , , were unknown). Nine modes gained plain-word and colloquial signals; the matrix now routes 28/28 with all harness and activation routes unchanged, and it is a unit test. New lists every one of the 42 audit findings with trigger and fix, pinned to the source by a drift test (); the number-presentation checks stop reading inline-code tokens as context, exempt explicit denials of best-run reporting, and no longer misread as a threshold ().

• v0.7.0 (2026-09-02) — two blind controlled studies published (second passes the primary quality gate; efficiency gates still fail), 30+ reporting standards absorbed, natural-tone / ablation / benchmarking modules, sbar-handoff / executive-onepager / rebuttal-response / release-card templates, seven readability audit warnings, report-in-response and proportional-ceremony contract fixes, ARIS-style README restructure.
• v0.6.0 (2026-08-25) — reportctl --version, GitHub Pages demo, citation metadata, cloud-sync test preflight.
• v0.5.0 (2026-08-25) — examples/showcase-20260825/ (16 finished reports + deck + receipts), EN/CN README split, CJK print-clipping fix with real-render regression tests.
• v0.4.0 (2026-08-24) — research-idea mode, 4 bounded research profiles (RL / embodied / world models / VLA), 8 exact template assets, slide surface.
• v0.3.x (2026-08-24) — preregistered study controller with immutable receipts, typed host adapter, blind A/B packets, claim gates; checkpoint artifact receipts; one-case Codex pilot published as insufficient_evidence.
• v0.2.0 (2026-08-24) — schema-v2 full-intent checkpoints, same-checkpoint final audit, shared bounded scanner.
• v0.1.0 (2026-08-24) — resident micro-contract, routed Agent Skill, structural audit, 11 modes, 5 modules, 4 host adapters.

Full detail in CHANGELOG.md.

3. ✨ What you get

• 12 primary report modes — concise answer, implementation handoff, status update, investigation, experiment report, research idea, decision brief, risk report, incident update, postmortem, academic synthesis, review report.
• 8 display modules — visuals, tables, conclusions, evidence, academic-paper presentation, ablation design, performance benchmarking, natural tone (a fact-preserving de-AI pass) — loaded only when the report needs them.
• 4 bounded research profiles — reinforcement learning, embodied AI, world models, VLA — domain protocol cards without hardcoding any benchmark's habits.
• 12 exact template assets — detailed experiment reports per domain, paper-idea brief, dependency-free HTML/PPT-style academic deck, Quarto Reveal.js source, SBAR handoff, executive one-pager, reviewer response, and model/dataset release card — retrieved one at a time, never bundled into context.
• A stdlib-only CLI (reportctl) that routes, bundles bounded context, checkpoints long tasks, scaffolds, audits structure mechanically, emits a cross-model semantic review prompt, and renders a strict JSON report IR.
• Anti-forgetting for long tasks — a tiny persistent micro-contract, a small on-disk checkpoint saved at task start, and a final audit driven by the same checkpoint file, so the reporting contract survives context compaction.
• Evidence-first culture — audits check structure mechanically; the repository never claims your facts are correct, and it does not claim measured quality gains it has not measured (see Evidence status).
• Absorbed reporting standards — 30+ named sources (ASA p-value statement, CONSORT/PRISMA accounting, benchmarking-crimes literature, error-bar and uncertainty-visualization research, model cards and datasheets, peer-review response norms) encoded into modes, modules, and templates, with every mapping auditable in docs/REPORTING-STANDARDS.md.

4. 🚀 Quick start

Fastest trial (no install)

Paste this together with the repository link into your agent:

请先读取该仓库的 AGENT_START.md，并用其中的最小路由完成本次最终汇报；
不要读取全部协议。仓库：https://github.com/asimfish/super_agent_presentation

Or use the English bootstrap prompt in AGENT_START.md. Link-only mode is best-effort; installation is the persistent path.

Install into a project

git clone https://github.com/asimfish/super_agent_presentation.git
cd super_agent_presentation
python3 scripts/install.py plan  --target /path/to/your/project --host codex   # preview
python3 scripts/install.py apply --target /path/to/your/project --host codex   # apply

--host claude, --host cursor, and --host copilot are also supported. The installer never overwrites existing Skills or instruction files; see INSTALL.md.

The agent workflow

python3 skills/agentic-reporting/scripts/reportctl.py list
python3 skills/agentic-reporting/scripts/reportctl.py bundle \
  --task "Report a five-seed RL ablation" \
  --mode experiment-report --profile reinforcement-learning --module tables
python3 skills/agentic-reporting/scripts/reportctl.py audit \
  --file report.md --mode experiment-report --strict

For a long task, save a checkpoint near the start, work freely, then let the final audit read the same checkpoint:

python3 skills/agentic-reporting/scripts/reportctl.py checkpoint \
  --task "Report implementation results, verification, remaining risk" \
  --mode implementation-handoff --surface chat \
  --must-show "Verification evidence" \
  --output <private-scratch>/agent-report.json

python3 skills/agentic-reporting/scripts/reportctl.py audit \
  --file report.md --checkpoint <private-scratch>/agent-report.json --strict

5. 🖼️ Real examples

examples/showcase-20260825/ holds one finished sample for every mode and profile — generated through the real CLI workflow with every receipt preserved, final verdict ALL PASS (16/16 strict audit, 16/16 semantic oracle, 7/7 manual page review).

See	Where
12 core-mode reports (Chinese samples)	examples/showcase-20260825/modes/
4 research-profile reports (RL / embodied / world models / VLA)	examples/showcase-20260825/profiles/
7-page academic HTML deck + Chrome-printed PDF	html/deck.html · render/academic-talk.pdf
Every first-attempt failure and its fix, kept honestly	first-failures.md
Machine-readable manifest binding reports to receipts	manifest.json

All example facts are synthetic fixtures; rendering was verified on macOS Chrome 151 only. Details and boundaries in examples/README.md.

6. 🔄 How it survives long tasks

flowchart LR
    accTitle: Three-layer execution mechanism of agentic reporting
    accDescr: A tiny resident contract saves a small checkpoint at long-task start and releases the detailed protocol; at delivery it reloads one primary protocol, at most two modules, and audits with the same checkpoint.
    A[Resident micro-contract] --> B{Likely long task?}
    B -- yes --> C[Route briefly, save tiny checkpoint]
    C --> D[Release detailed protocol]
    B -- no --> E[Do the task]
    D --> E
    E --> F[Reporting boundary]
    F --> G[Reload checkpoint or pick one primary protocol]
    G --> H[Research task: at most one domain profile]
    H --> K[Prefer one display module, max two]
    K --> L[Surface guide only when needed; template assets stay out of bundles]
    L --> I[Audit with the same checkpoint, verify facts]
    I --> J[Final report]

• The resident layer only decides when to activate; it never keeps the manual in context.
• Progressive disclosure loads one primary protocol, at most one research profile, and bounded modules — measured context budgets in docs/CONTEXT-BUDGET.md.
• Prompt layers reduce forgetting; they cannot mechanically block a bad delivery. For that, wrap checkpoint + final audit --checkpoint exit codes in CI or a wrapper.

7. 🗂️ Catalog

12 modes · 8 display modules · 4 research profiles · 5 surfaces · 12 exact templates — the full inventory with per-item summaries, bounded route files, and finished-example links lives in docs/CATALOG.md.

Research/display need	Asset or profile	Enforced emphasis
General experiment	experiment-report-detailed	claim map, protocol, metric direction, run counts, uncertainty, reproducibility
Paper idea	research-idea mode + template	limitation, hypothesized mechanism, recent work, decisive experiment, falsifiers, kill criterion
Reinforcement learning	reinforcement-learning / rl-experiment-report	env steps, runs/seeds, tuning parity, learning curves, interval estimates, failed tasks
Embodied AI	embodied-ai / embodied-experiment-report	sim vs real, robot and sensors, success rule, resets/interventions, generalization axes, failure types
World models	world-models / world-model-experiment-report	model/data cards, open-loop prediction, closed-loop control, scaling, exploitation boundary
VLA	vla / vla-experiment-report	data mixture, embodiment and action interface, rollouts, generalization, latency, safety
HTML/PPT academic talk	academic-talk-html	dependency-free, responsive, printable, keyboard nav, assertion-evidence pages
Quarto slides	academic-talk-revealjs	Reveal.js, citations, speaker notes, self-contained HTML, appendix
Ablation study	ablation module	variant-versus-full-system comparison, interaction checks, tuning-policy parity, honest contribution tables
Performance comparison	benchmarking module	full-suite protocol, geometric/harmonic means, speedup and tail-latency discipline, platform disclosure
De-AI tone pass	natural-tone module	fidelity contract (facts, relations, scope survive rewriting), signal-to-action table for CN/EN boilerplate, misfire protection for technical vocabulary
Reviewer response	rebuttal-response	point-by-point quoting, outcome-first replies, precise revision locations, no unverifiable promises
Model/dataset release

8. 🪜 Deployment and constraint ladder

Usage	Fits	Constraint strength
Send the repo URL to the agent	quick trial	best-effort; a URL is not an install and does not raise instruction priority
Explicit $agentic-reporting invocation	one-off task with the Skill installed	medium
Installed Skill + host micro-contract	daily cross-task use	recommended; the host keeps reminding the finalization flow
Same-checkpoint audit + external wrapper/CI	long tasks, batch delivery	mechanically blockable by exit code; still does not verify facts
JSON IR + validator/renderer	APIs, durable formal reports	strongest structural constraint; does not replace evidence checking

For formal reports, start from skills/agentic-reporting/assets/templates/report-spec.json, then validate-spec → render → audit --strict. The JSON IR separates verified / inference / recommendation claims and requires verified conclusions to cite evidence IDs. Details in README_CN.md and docs/ARCHITECTURE.md.

9. 🔬 Evidence status

This repository practices what it preaches: it does not claim measured effectiveness it does not have. Two private blind controlled studies have been run through the preregistered pipeline and their controller-generated aggregates are published under evals/runs/controlled/:

Run	Design	Primary composite (95% CI)	Primary gate	Token overhead (median / p90)	Claim status
codex-20260825	28 held-out cases × 2 conditions × 3 seeds = 168 Codex runs; 84 blind pairs; 2 independent model raters	+0.214 [0.015, 0.478]	fail (< +0.3)	8.1x / 27.8x	insufficient_evidence
codex-20260826	identical design, treatment pinned after the report-in-response and proportional-ceremony fixes	+0.315 [0.110, 0.579]	pass	7.9x / 26.7x	insufficient_evidence

What the studies support: under blind rating the framework condition scores higher on 6 of 7 quality dimensions, concision reached parity in the second run, visual selection precision and recall are 1.0, and zero framework responses leaked a local path (45 of 46 critical-error records in the second run sit on the baseline side). What they do not support: the efficiency gates fail structurally — the framework writes full reports where the baseline often answers in a line, so output tokens run ~8x and semantic density per 1,000 tokens collapses — and the 65% win-rate gate is unmet (45.2%). Cases were synthetic, raters were models rather than humans, and the provider revision was unpinned. The controller therefore keeps effectiveness_claim_eligible: false, and this README claims no readability or efficiency gain. The earlier one-case pilot remains published as insufficient_evidence for the same reason. Gates, blinding, and receipts are specified in BENCHMARK.md and evals/.

What is verified: deterministic routing, bounded bundles, checkpoint/audit mechanics, template rendering (329 unit tests + harness smoke + real-render regression on macOS Chrome 151), and the end-to-end showcase run with receipts in examples/.

10. 📁 Repository structure

AGENT_START.md                 # smallest bootstrap for link-only agents
AGENTS.md                      # resident micro-contract for this repo
adapters/                      # Claude, Codex, Cursor, Copilot host adapters
skills/agentic-reporting/
  SKILL.md                     # routing and finalization workflow
  references/                  # core, primary modes, display modules
  assets/templates/            # exact Markdown/JSON structure guards
  assets/presentations/        # HTML/Quarto academic deck templates
  scripts/reportctl.py         # route/bundle/checkpoint/audit/render CLI
dist/                          # prebuilt link-only route bundles
examples/showcase-20260825/    # 16 finished reports + deck + audit receipts
evals/                         # study schemas, templates, sanitized run records
scripts/                       # installer, benchmark harness, study controller
docs/                          # architecture, research, catalog, ADRs
tests/                         # 329 unit tests

11. ✅ Verification

python3 scripts/check_test_env.py               # preflight: cloud-sync/eviction hazards
python3 -m pip install -r requirements-dev.txt
python3 -m unittest discover -s tests -v
python3 scripts/presentation_benchmark.py smoke

The preflight warns when the clone sits in a cloud-synced folder (iCloud Desktop/Documents included) and fails when file content has been evicted — both make the subprocess-heavy tests hang or flake.

smoke validates known-good/bad fixtures and routing and prints host_activation_observed: false — it calls no real host or model.

12. 📚 Design docs

Architecture · Catalog · Audit codes · Research sources · Template provenance · Reporting standards · Context budgets · Performance · Security review · Benchmark protocol · Security policy · ADRs

13. 🤝 Contributing

See CONTRIBUTING.md. The short version: run the full test suite, keep dist/ reproducible, and never add an effectiveness claim without the evidence gates in BENCHMARK.md.

14. 📖 Citation

If this framework is useful in your work, cite it via CITATION.cff (GitHub's "Cite this repository" button) or:

@software{super_agent_presentation,
  author  = {asimfish},
  title   = {Super Agent Presentation: an agent-native reporting framework},
  year    = {2026},
  version = {0.7.0},
  url     = {https://github.com/asimfish/super_agent_presentation}
}

15. ⭐ Star History

16. 🙏 Acknowledgements

Absorbed standards — the framework encodes 30+ named reporting sources (the ASA statement on p-values, CONSORT/PRISMA accounting, IBCS, NN/g scanning research, Google SRE postmortem culture, plain-language guidelines, benchmarking-crimes literature, model cards and datasheets, peer-review response norms, …). Every mapping is auditable in docs/REPORTING-STANDARDS.md; template provenance is tracked source-by-source in docs/TEMPLATE-SOURCES.md.

Adapted with license — the natural-tone module distills the MIT-licensed shuorenhua de-AI rewriting skill by @MrGeDiao into a fidelity-contracted display module plus a deterministic audit rule.

README conventions — the structure of this README (dated What's New, numbered contents, honest boundary callouts) follows the example set by ARIS.

Host platforms — Claude Code · Codex · Cursor · GitHub Copilot.

17. License

MIT — see LICENSE.